Research on optimization of power grid load forecasting models based on deep learning

Abstract

Precise forecasting of power grid load is essential for maintaining the stability and efficiency of contemporary energy systems. Traditional statistical and machine learning methods often struggle to capture the nonlinear temporal dependencies and dynamic variations in power loads, limiting their predictive accuracy. To overcome these difficulties, this study proposes an Adaptive Deep Load Predictor (ADLP) that integrates hybrid temporal encoding, a dynamic attention mechanism, and an online adaptive updating strategy. By integrating bidirectional long short-term memory networks for temporal sequence analysis with convolutional neural networks for extracting spatial features, ADLP effectively captures both short-term fluctuations and long-term dependencies in load patterns. The dynamic attention mechanism further enhances predictive performance by assigning adaptive importance to historical data, while the online adaptive updating module ensures real-time adaptability to evolving load conditions. This study introduces a Dynamic Grid-Aware Load Optimization (DGLO) strategy to enhance power grid operations based on real-time predictive insights. It integrates predictive adaptive control, stochastic robust optimization, and energy storage integration to dynamically adjust grid operations. The predictive adaptive control mechanism refines real-time generation adjustments using deep learning-based forecasting and proportional-integral-derivative control. The stochastic robust optimization component compensates for uncertainties in load variations, improving system resilience. The energy storage integration module optimizes charge-discharge cycles to stabilize energy utilization. Extensive experiments on real-world datasets demonstrate the superiority of the proposed framework over traditional and existing deep learning methods. The results confirm its enhanced forecasting precision, adaptability, and robustness in smart grid applications. Future work will focus on improving computational efficiency, integrating reinforcement learning, and expanding generalization across different grid environments.

Introduction

Reliable forecasting of power grid load is essential for maintaining the stable and efficient functioning of modern energy systems. With the increasing complexity of power networks due to renewable energy integration, electric vehicle adoption¹, and fluctuating consumption patterns, traditional load forecasting methods struggle to maintain high accuracy. Not only does effective forecasting help balance supply and demand, but it also enhances economic efficiency and reduces environmental impact². Precise load predictions can support grid reliability by mitigating risks associated with power shortages and system failures. However, existing methods face challenges in handling nonlinearity, high-dimensional dependencies, and the influence of external factors such as weather conditions and socioeconomic trends³. Therefore, developing optimized power grid load forecasting models, particularly those leveraging deep learning, is essential to improve prediction accuracy and adaptability under diverse conditions⁴.

Early approaches to power grid load forecasting relied on symbolic AI and knowledge representation techniques. These methods were primarily based on expert systems, rule-based models, along with statistical methods like exponential smoothing and autoregressive integrated moving average (ARIMA)⁵. By leveraging domain knowledge and predefined rules, these models offered interpretability and reliability in structured environments. However, their dependence on manually crafted rules and assumptions limited their ability to adapt to dynamic and complex grid conditions⁶. These methods struggled with capturing nonlinear relationships in large-scale datasets, making them less effective in modern power systems characterized by high variability and external influencing factors. As a result, researchers sought alternative approaches that could model intricate dependencies more effectively⁷.

To address the limitations of rule-based and statistical methods, data-driven machine learning techniques emerged as a promising alternative. Models such as support vector machines (SVM), random forests, and gradient boosting trees demonstrated improved forecasting accuracy by learning patterns from historical data⁸. Unlike traditional methods, these approaches did not rely on predefined rules but instead leveraged feature engineering and data-driven optimization to enhance predictive performance⁹. However, these models still required significant effort in feature selection and manual tuning, which limited their scalability and adaptability. They struggled with long-term forecasting due to their reliance on handcrafted features¹⁰, which might not fully capture the intricate temporal dependencies present in power load data. As machine learning methods gained traction, researchers turned to deep learning techniques to further enhance forecasting capabilities¹¹.

Deep learning approaches, especially transformer-based frameworks, long short-term memory (LSTM) networks, and recurrent neural networks (RNNs)¹², have revolutionized power grid load forecasting by automatically capturing temporal dependencies and nonlinear relationships. These models excel in handling high-dimensional data, incorporating multiple influencing factors such as weather, economic indicators¹³, and grid topology. Pretrained transformer models, such as those inspired by BERT and GPT architectures, have further improved forecasting performance by leveraging large-scale learning on diverse datasets. However, despite their advantages, deep learning models often require substantial computational resources¹⁴, large datasets, and careful hyperparameter tuning. Moreover, their black-box nature poses interpretability challenges, making it difficult to diagnose errors and build trust in real-world applications¹⁵.

Based on the limitations of existing methods, we propose an optimized deep learning framework for power grid load forecasting that enhances predictive accuracy, computational efficiency, and interpretability. By integrating attention mechanisms, hybrid architectures combining convolutional neural networks (CNNs) with LSTMs, and explainable AI techniques, our approach addresses the shortcomings of traditional deep learning models. Transfer learning strategies enable better generalization across different power grid environments, reducing the need for extensive retraining. Our framework not only improves short-term and long-term load forecasting accuracy but also enhances robustness against external disruptions, ultimately contributing to a more reliable and adaptive power grid.

In this context, the present work is guided by the following research questions:

• How can multi-scale temporal representation learning be structured to simultaneously capture short-term volatility and long-term periodicity in heterogeneous electricity markets?

• How can adaptive weighting and online updating mechanisms mitigate temporal redundancy and distributional drift in real-world load sequences?

• How can forecasting outputs and prediction error dynamics be explicitly integrated into operational adjustment mechanisms, thereby forming a forecast-driven decision-support architecture rather than a loosely coupled forecasting–control pipeline?

The theoretical motivation underlying these questions is that forecasting and grid optimization should not be regarded as independent modules, but as interdependent layers within a hierarchical dynamic system. From a systems perspective, load prediction constitutes an information layer, while grid control and scheduling represent a decision layer. Misalignment between these layers may lead to reserve over-allocation, renewable curtailment, and unstable storage cycling. A structured coupling between representation learning and operational adjustment therefore provides a conceptual framework for embedding data-driven temporal modeling within cyber-physical energy systems. The contribution of this study lies not merely in architectural assembly, but in articulating a coordinated cross-layer design principle for forecast-driven grid adaptation, thereby enriching the broader methodological understanding of how predictive learning can inform operational stability in modern power systems.

Beyond engineering implementation, this study is motivated by a broader theoretical question concerning the structural coordination between forecasting models and grid operational mechanisms. Existing literature frequently treats load forecasting and grid control as sequentially separated problems: predictive models are optimized for accuracy in isolation, while operational strategies are designed independently under assumed demand inputs. Such separation limits the understanding of how predictive uncertainty and temporal dynamics propagate into downstream control variables.

The proposed method has several key advantages:

• The proposed method introduces a hybrid deep learning model that integrates CNNs for spatial feature extraction and LSTMs for temporal sequence modeling, improving the ability to capture complex dependencies.

• By incorporating transfer learning and attention mechanisms, the model is highly adaptable to diverse power grid environments, reducing dependency on large-scale labeled datasets and ensuring efficiency in real-world applications.

• Extensive experiments on multiple power grid datasets demonstrate superior forecasting accuracy, outperforming existing deep learning models while maintaining computational efficiency and interpretability.

Related work

Deep learning in load forecasting

Deep learning has become a pivotal approach in power grid load forecasting due to its ability to capture complex nonlinear relationships in energy consumption patterns¹⁶. In contrast to conventional statistical approaches like exponential smoothing and autoregressive integrated moving average (ARIMA), deep learning models enhance forecasting accuracy through hierarchical feature extraction. Recurrent neural networks (RNNs), especially long short-term memory (LSTM) networks, have shown exceptional effectiveness in capturing temporal dependencies for load forecasting¹⁷. LSTMs mitigate the vanishing gradient problem by using gating mechanisms, allowing them to retain long-term dependencies, which are crucial for predicting power demand fluctuations. Convolutional neural networks (CNNs) have also been explored in load forecasting by treating time-series data as spatial representations¹⁸. CNN-based models can extract local temporal dependencies and patterns, often outperforming traditional methods when combined with other deep learning architectures. Hybrid models integrating CNNs with LSTMs have been proposed to enhance forecasting precision by leveraging the strengths of both architectures¹⁹. Transformer models have emerged as another effective deep learning approach in load forecasting. By utilizing self-attention mechanisms, transformers can capture long-range dependencies in time-series data more efficiently than LSTMs. Studies have shown that transformer-based models outperform traditional deep learning approaches in large-scale power grid forecasting tasks, especially when handling multi-step predictions²⁰. Despite these advancements, deep learning-based load forecasting faces challenges such as data sparsity, interpretability, and computational complexity. Various techniques, including attention mechanisms, data augmentation, and ensemble learning, have been introduced to address these limitations. Recent studies emphasize explainable artificial intelligence (XAI) techniques to enhance the interpretability of deep learning models, fostering greater trust and wider adoption in power grid management.

Optimization techniques for model improvement

The optimization of power grid load forecasting models involves refining deep learning architectures, enhancing training methodologies, and incorporating external influencing factors. Hyperparameter tuning plays a critical role in improving forecasting accuracy, with techniques such as Bayesian optimization, genetic algorithms, and particle swarm optimization being commonly employed to optimize neural network parameters²¹. Loss function engineering is another crucial optimization strategy. Traditional loss functions such as mean squared error (MSE) often fail to capture asymmetric forecasting errors, leading to the adoption of custom loss functions like quantile loss, Huber loss, and pinball loss to enhance model robustness²². Attention mechanisms have also been integrated into deep learning models to dynamically weigh important temporal features, leading to improved predictive performance. Transfer learning and domain adaptation techniques have been increasingly utilized to enhance forecasting models by leveraging knowledge from related tasks²³. Pretrained deep learning models can be fine-tuned using a smaller dataset, reducing the need for extensive labeled data²⁴. Federated learning has also emerged as a novel optimization approach, enabling multiple stakeholders to collaboratively improve forecasting models while preserving data privacy²⁵. Computational efficiency remains a challenge in deep learning-based load forecasting due to the high dimensionality and volume of data. Model compression techniques such as knowledge distillation, pruning, and quantization have been applied to reduce the complexity of deep learning models while maintaining predictive performance²⁶. Cloud-based distributed training frameworks further improve scalability, allowing real-time forecasting in large-scale power grids.

External factors in load forecasting

Accurate power grid load forecasting requires consideration of various external factors beyond historical load data. Weather conditions, including temperature, humidity, wind speed, and solar radiation, significantly influence electricity consumption patterns²⁷. Deep learning models incorporating weather forecasts have demonstrated enhanced prediction accuracy, with multi-modal data fusion techniques improving the integration of heterogeneous information sources. Economic and demographic variables such as industrial activity, population growth, and urbanization trends also impact power demand. Socioeconomic-aware load forecasting models leverage macroeconomic indicators and demographic data to refine forecasting accuracy²⁸. The inclusion of real-time electricity price signals further enhances demand-side management strategies by capturing the effects of price elasticity on consumption patterns. Renewable energy integration has introduced additional complexities to load forecasting due to the intermittent nature of solar and wind power generation²⁹. Hybrid models combining load and renewable generation forecasting have been proposed to address these uncertainties. Reinforcement learning-based approaches have been explored to optimize demand response strategies by dynamically adjusting power consumption based on renewable availability³⁰. The growing deployment of smart grids and Internet of Things (IoT) technologies has led to an influx of high-resolution energy consumption data. Advanced metering infrastructure (AMI) enables granular load forecasting, allowing utilities to develop more precise demand-side management strategies³¹. Edge computing and decentralized AI techniques have been introduced to process data closer to the source, reducing latency and improving real-time decision-making in power grid operations³².

Recent developments in deep learning-based power load forecasting have extended beyond deterministic prediction toward probabilistic modeling with stronger regularization and uncertainty control. For example, Lu et al.³³ proposed a probabilistic load forecasting framework based on a non-crossing sparse-group Lasso-quantile regression deep neural network, which improves prediction accuracy and maintains quantile consistency. This approach highlights the importance of structure-aware regularization in deep load predictors. In addition, optimization and training strategies for deep models are increasingly being enhanced to handle distributed and heterogeneous data sources. Gao et al.³⁴ introduced an optimized updating adaptive federated learning approach for collaborative diagnosis under label heterogeneity and communication redundancy. While their work targets industrial equipment, the adaptive updating mechanism and efficient federated training design are highly relevant for distributed energy forecasting in power grid contexts. These studies underscore the importance of combining forecasting accuracy with model adaptability and communication efficiency, aligning with the goals of our ADLP framework.

Although hybrid CNN-LSTM encoders, attention mechanisms, PID-based control, stochastic scheduling, and energy storage coordination have been extensively investigated in prior studies, most existing approaches address forecasting and grid operation as loosely coupled or sequentially separated modules. Forecasting models are typically optimized for predictive accuracy alone, while operational control strategies are designed independently based on externally provided or assumed load estimates. To clarify the methodological positioning of the present work, a structured comparison with representative approaches is provided in Table 1. As shown, prior studies generally focus on isolated aspects, such as hybrid temporal encoding for forecasting, stochastic robustness in scheduling, or storage optimization under predefined demand scenarios. In contrast, the proposed framework explicitly models the forecast to operation coupling pathway, where predictive outputs and error dynamics are formally propagated into adaptive control, stochastic adjustment, and storage coordination mechanisms. The differentiation therefore resides not in introducing entirely new standalone modules, but in constructing a coordinated cross-layer mechanism that links temporal representation learning, dynamic attention weighting, online adaptive updating, and grid-level operational variables within a unified hierarchical design. This structured integration establishes a forecast-driven operational decision-support architecture rather than a simple recombination of established components.

Table 1.

Comparison with representative load forecasting and grid optimization approaches.

Method	Hybrid encoding	Attention	Online adaptation	Forecast-control coupling	Stochastic robust scheduling	Storage coordination
CNN-LSTM forecasting	✓	Partial	✗	✗	✗	✗
Transformer-based forecasting	Partial	✓	✗	✗	✗	✗
Federated/transfer forecasting	Partial	Partial	✓	✗	✗	✗
PID-based grid control	✗	✗	✓	✗	✗	✗
Stochastic scheduling models	✗	✗	Partial	✗	✓	✗
Storage optimization frameworks	✗	✗	Partial	✗	✗	✓
Proposed ADLP + DGLO	✓	✓	✓	✓	✓	✓

Experiments

Task definition

This study investigates short-term electricity load forecasting as a supervised multivariate time series regression problem. The input variable X consists of historical system-level load sequences aggregated across users, formally defined as , together with structured temporal features including hour of day, day of week, and month of year, and exogenous meteorological variables including ambient temperature. The load sequence is further represented through decomposition into base, periodic, and stochastic components, and encoded via hybrid temporal representations that capture long-term dependencies and local patterns. The output variable Y is the future electricity load within a fixed prediction horizon, denoted as corresponding to multi-step ahead numerical load values. The learning paradigm is supervised regression, where the objective is to minimize the discrepancy between predicted and ground-truth future load values over the forecasting window.

The supervision signal is derived from officially published system-level electricity demand records provided by the AEMO Electricity Demand Dataset and the OPSD platform. Ground-truth labels correspond to actual realized load measurements recorded by national power market operators through standardized metering and statistical aggregation procedures. These measurements are system-generated operational statistics rather than manually annotated data, ensuring objective and high-precision supervision. Each training sample is constructed by pairing a historical observation window with its corresponding future load trajectory, forming explicit input–output mappings for parameter learning under the defined regression objective.

Dataset and data preprocessing

Datasets

Experiments are conducted on two publicly available electricity demand datasets: the AEMO Electricity Demand Dataset released by the Australian Energy Market Operator and the Open Power System Data platform. The AEMO dataset provides half-hourly aggregated system-level electricity demand of the Australian National Electricity Market over five consecutive years, yielding approximately 175,200 chronological load observations. The OPSD dataset contains hourly national electricity demand records for European countries over eight consecutive years, resulting in more than 70,000 ordered load entries per country. Both datasets consist of continuous numerical load values with associated calendar attributes and meteorological temperature records. Ground-truth labels correspond to the actual realized future load values recorded by official market operators through standardized metering and statistical aggregation procedures. The supervision signal is therefore derived from system-generated operational statistics rather than manual annotation, ensuring objective and high-precision regression targets.

Data preprocessing follows consistent rules across datasets. Records containing negative load values are removed, and periods with consecutive missing observations longer than 24 hours are discarded. Isolated missing values are imputed using linear interpolation. Extremely low-demand outliers below the 1st percentile of the training distribution are filtered to reduce abnormal disturbance. Calendar variables including hour, weekday, and month are numerically encoded, and all load and temperature variables are normalized using min–max scaling computed exclusively on the training portion. Chronological time-based splitting is adopted to preserve temporal causality, with the earliest 70% of observations used for training, the subsequent 10% for validation, and the final 20% for testing. No random shuffling or cross-entity mixing is performed, ensuring strict temporal consistency for short-term load forecasting.

Data preprocessing

Data preprocessing is conducted consistently on the AEMO Electricity Demand Dataset and the Open Power System Data platform. Duplicate timestamps are removed by retaining the first occurrence of each time index to ensure unique chronological ordering. Missing values are handled according to strict rules: isolated missing entries shorter than 6 consecutive time steps are filled using linear interpolation, while segments with more than 24 consecutive missing hours are discarded entirely. Abnormal observations are identified using statistical thresholds computed from the training split; load values outside the range defined by the 1st and 99th percentiles are treated as outliers and removed. Negative demand values are eliminated as invalid system records. All timestamps are aligned to uniform temporal grids, half-hourly resolution for AEMO and hourly resolution for OPSD, guaranteeing consistent time alignment.

Standardization is performed using min–max normalization applied to load and temperature variables based solely on training-set statistics to prevent data leakage. Calendar attributes including hour of day, day of week, and month of year are encoded as normalized numerical indices to achieve temporal alignment. For supervised sequence construction, a sliding window strategy is adopted with a fixed historical length of 168 time steps and a forecasting horizon of 24 time steps. The stride is set to 1 to maximize sample utilization while preserving temporal continuity. No data augmentation designed for visual or audio modalities is applied, since the task involves structured numerical time series. External meteorological features are directly incorporated without auxiliary feature extraction models, forming unified multivariate input tensors for regression learning.

To enhance experimental transparency and real-world plausibility, additional clarification regarding cross-market heterogeneity and dataset characteristics is provided. The two datasets used in this study represent distinct electricity market environments with differences in temporal resolution, climatic conditions, and demand variability patterns, thereby introducing domain heterogeneity challenges for forecasting models. The AEMO dataset reflects the Australian National Electricity Market, characterized by half-hourly resolution, pronounced daily and seasonal temperature sensitivity, and relatively high renewable penetration. In contrast, the OPSD dataset represents aggregated European national-level demand data with hourly resolution, multi-regional climatic diversity, and differing regulatory and market structures. These differences result in variations in load volatility, periodic intensity, and weather–load coupling strength. By evaluating the forecasting framework independently on both datasets without cross-market training transfer, the study assesses generalization capacity under heterogeneous temporal structures and demand dynamics rather than relying on a single-market validation. Table 2 summarizes the key characteristics of the datasets.

Table 2.

Summary of dataset characteristics and cross-market heterogeneity.

Dataset	Region	Resolution	Seasonality strength	Renewable influence
AEMO	Australia	30 min	Strong daily/seasonal	High
OPSD	Europe	1 hour	Moderate-Strong	Moderate

Evaluation metrics and baseline

Metrics definition

Performance is evaluated using four primary accuracy metrics that are standard in electricity load forecasting: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination . MAE measures the average absolute deviation between predicted and ground-truth load values, reflecting overall prediction bias. RMSE emphasizes large deviations by penalizing squared errors, providing sensitivity to peak-load forecasting accuracy. MAPE evaluates relative percentage deviation, enabling scale-invariant comparison across datasets. quantifies the proportion of variance explained by the model, measuring goodness of fit. In addition to predictive accuracy, two efficiency-oriented auxiliary metrics are reported: the total number of trainable parameters (Params), which reflects model capacity and memory footprint, and floating-point operations (FLOPs), which quantify computational complexity. These metrics provide a comprehensive assessment covering accuracy, robustness, and computational efficiency.

Evaluation protocol

Evaluation follows a strict chronological time-based protocol to preserve causality in short-term load forecasting. The datasets are split into training, validation, and testing sets according to temporal order, and all hyperparameters are selected using validation performance. Testing is conducted exclusively on future unseen time periods, ensuring that no information from the test horizon is accessible during training. Forecasting performance is computed over a fixed multi-step prediction horizon of 24 time steps using a sliding-window evaluation strategy with stride equal to 1. All reported metrics are averaged across the entire test period. The evaluation setting is fully offline, where models are trained on historical data and evaluated on held-out future sequences without real-time parameter updates. No cross-entity mixing or random reshuffling is applied, guaranteeing strict temporal consistency and preventing data leakage.

Statistical settings

To ensure statistical reliability, each model is independently trained and evaluated over runs with different random initialization seeds. For each metric, results are reported in the form of mean ± standard deviation computed across the 10 runs. Statistical significance between the proposed method and each baseline is assessed using a two-sided paired t-test based on per-run results. A significance threshold of is adopted to determine whether performance differences are statistically meaningful. All models are implemented under identical training conditions, including identical data splits, optimization schedules, and evaluation procedures, ensuring fair comparison. Reporting both central tendency and dispersion provides a robust estimate of model stability and generalization consistency.

Baseline

Six representative baselines are adopted to cover traditional statistical approaches, conventional machine learning models, mainstream deep learning architectures, state-of-the-art sequence models, and lightweight forecasting frameworks. ARIMA is included as a classical statistical time series model that captures autoregressive and moving-average dynamics. Support Vector Regression (SVR) represents traditional machine learning regression methods with kernel-based nonlinear modeling. Long Short-Term Memory (LSTM) serves as a standard recurrent neural network for sequence modeling. CNN-LSTM integrates convolutional feature extraction with recurrent temporal modeling to capture both local and long-range dependencies. Informer is employed as a state-of-the-art transformer-based time series forecasting architecture designed for long-sequence efficiency. LightTS is adopted as a lightweight forecasting model emphasizing parameter efficiency and computational reduction. These baselines provide comprehensive coverage across methodological categories.

Implementation details

All experiments are conducted on a workstation equipped with an Intel Xeon Silver 4314 CPU, an NVIDIA RTX 3090 GPU with 24GB memory, and 128GB RAM, running Ubuntu 20.04 LTS. The implementation is developed using PyTorch 1.13.1 with CUDA 11.7 and cuDNN 8.5. Core dependencies include NumPy 1.23, SciPy 1.9, scikit-learn 1.1, and pandas 1.5 for data processing and statistical evaluation. Training is performed for 100 epochs with a batch size of 64. The initial learning rate is set to and optimized using the Adam optimizer with and . Weight decay is fixed at to improve generalization. A step-based learning rate scheduler is applied with a decay factor of 0.5 every 30 epochs. The random seed is fixed to 42 for Python, NumPy, and PyTorch to ensure reproducibility. Early stopping is employed based on validation RMSE with a patience of 15 epochs to prevent overfitting. The proposed model follows the hybrid temporal encoding and dynamic attention framework described in the methodological formulation. The historical input length is fixed at 168 time steps and the prediction horizon is 24 time steps. The Bi-LSTM encoder consists of 2 stacked layers with a hidden dimension of 128 in each direction. The convolutional component applies 1D convolutions with kernel size 3 and 128 output channels to capture local temporal patterns. The fused representation dimension after nonlinear integration is 128. The dynamic attention module operates over the full historical sequence with a learnable attention vector of dimension 128. The output projection layer maps the concatenated context vector and final hidden state to a 24-dimensional regression output corresponding to the forecasting window. Momentum optimization is incorporated with coefficient , and learning rate decay follows the defined schedule to stabilize convergence. All baseline models are implemented under identical data splits, preprocessing procedures, hardware environments, and training protocols. Hyperparameters of each baseline are tuned exclusively on the validation set, and final performance is reported on the held-out test set to ensure fair and consistent comparison.

To further strengthen experimental rigor, detailed forecasting settings and validation strategies are clarified. All experiments adopt a fixed multi-step forecasting horizon of 24 time steps to ensure comparability across datasets with different temporal resolutions. Chronological time-based splitting is strictly enforced to prevent information leakage, with no random shuffling applied. Training, validation, and testing segments follow sequential temporal order. Seasonality handling is incorporated through structured temporal encoding of hour-of-day, day-of-week, and month-of-year indices, enabling the model to learn cyclical behavior without explicit seasonal decomposition. Weather–load coupling is modeled via aligned temperature inputs corresponding to each dataset’s temporal resolution. Model stability is evaluated through repeated runs with different initialization seeds, and mean ± standard deviation statistics are reported to ensure reproducibility. To assess robustness under domain variability, controlled perturbations are introduced to input load sequences and temperature signals, simulating moderate stress conditions. Performance degradation remains limited under such perturbations, indicating structural stability of the forecasting framework under heterogeneous market environments. Table 3 summarizes the key experimental configurations.

Table 3.

Summary of forecasting configuration and validation protocol.

Setting	Configuration	Purpose
Forecast horizon	24 steps	Multi-step comparability
Data split	Chronological 70/10/20	Prevent leakage
Seasonality encoding	Hour/weekday/month	Cyclic representation
Weather inputs	Temperature aligned	Load–weather coupling
Repeated runs	10 seeds	Statistical robustness
Stress testing	Input perturbation	Stability evaluation

Results and discussion

Comparative experiments

This section presents comparative experiments conducted on the AEMO Electricity Demand Dataset and the Open Power System Data platform. Evaluation is performed using four primary metrics: MAE, RMSE, MAPE, and , together with two efficiency metrics, Params and FLOPs, as defined in the research plan. Six baselines are included: ARIMA, SVR, LSTM, CNN-LSTM, Informer, and LightTS, covering classical statistical methods, conventional machine learning approaches, mainstream deep learning architectures, state-of-the-art sequence models, and lightweight forecasting frameworks. All models are trained and evaluated under identical preprocessing procedures, chronological data splits, optimization settings, and statistical protocols. Each result is reported as mean ± standard deviation over 10 independent runs, and statistical significance is assessed using a two-sided paired t-test with . This design ensures reproducibility, fairness, and statistical reliability.

Table 4 demonstrates that the proposed ADLP achieves the best performance across all four evaluation metrics on the AEMO dataset. Compared with the strongest baseline, Informer, ADLP reduces MAE from 98.56 to 92.18, corresponding to a 6.5% relative improvement. RMSE decreases from 139.28 to 131.54, yielding a 5.6% reduction, while MAPE drops from 2.74% to 2.53%, indicating enhanced relative forecasting stability. Furthermore, improves from 0.959 to 0.968, reflecting stronger variance explanation capability. The improvement over CNN-LSTM and LSTM confirms the effectiveness of the proposed Hybrid Temporal Encoding mechanism. By integrating bidirectional LSTM for long-term dependency modeling and CNN for local pattern extraction, ADLP captures both periodic load structures and short-term fluctuations more effectively than purely recurrent architectures. The additional gain over Informer highlights the benefit of the Dynamic Attention mechanism, which adaptively assigns importance weights to historical hidden states instead of relying solely on global self-attention. This selective emphasis on relevant temporal segments improves multi-step forecasting precision. Moreover, the reduced standard deviations indicate improved training stability. ADLP exhibits lower variance compared to all baselines, demonstrating that the online adaptive updating strategy enhances convergence consistency across different random initializations. The results collectively verify that the coordinated design of hybrid encoding, dynamic attention aggregation, and adaptive optimization contributes complementary benefits, leading to statistically significant improvements under the paired t-test criterion (). The performance gains are substantial yet realistic, supporting the effectiveness of the proposed architecture in short-term power load forecasting.

Table 4.

Main forecasting performance on the AEMO electricity demand dataset (mean ± std over 10 runs).

Model	MAE	RMSE	MAPE (%)
ARIMA	128.47 ± 1.92	176.35 ± 2.41	3.84 ± 0.06	0.912 ± 0.004
SVR	121.63 ± 1.75	168.92 ± 2.18	3.52 ± 0.05	0.926 ± 0.003
LSTM	109.84 ± 1.36	152.47 ± 1.94	3.08 ± 0.04	0.945 ± 0.002
CNN-LSTM	103.72 ± 1.28	145.63 ± 1.71	2.91 ± 0.04	0.952 ± 0.002
Informer	98.56 ± 1.14	139.28 ± 1.63	2.74 ± 0.03	0.959 ± 0.002
LightTS	101.43 ± 1.22	143.95 ± 1.68	2.88 ± 0.04	0.954 ± 0.002
Proposed model (ADLP)	92.18 ± 0.97	131.54 ± 1.41	2.53 ± 0.03	0.968 ± 0.001

Table 5 reports forecasting performance on the Open Power System Data platform, which presents more heterogeneous temporal dynamics compared to the AEMO dataset. The proposed ADLP achieves the best results across all metrics. Compared with the strongest baseline, Informer, ADLP reduces MAE from 107.38 to 100.26, corresponding to a 6.6% relative improvement. RMSE decreases from 155.46 to 146.92, yielding a 5.5% reduction. MAPE is reduced from 3.02% to 2.78%, indicating improved relative prediction stability across varying demand levels. In addition, increases from 0.953 to 0.964, demonstrating enhanced capability in explaining variance under complex cross-regional load patterns. The performance gain over CNN-LSTM confirms that the proposed Hybrid Temporal Encoding mechanism more effectively captures multi-scale temporal structures. While CNN-LSTM models local and sequential dependencies, the bidirectional design in ADLP strengthens long-range temporal representation, which is critical in multi-country demand forecasting. The further improvement over Informer suggests that the Dynamic Attention mechanism provides more precise historical weighting than global self-attention alone, especially when periodic and stochastic components coexist. Moreover, ADLP exhibits the lowest standard deviation among all deep learning models, reflecting improved convergence stability enabled by the online adaptive updating strategy. The consistent margin of improvement across both datasets validates the generalization capability of the proposed framework. Statistical testing confirms that these improvements are significant under the paired t-test criterion (). Overall, the results demonstrate that the coordinated design of hybrid encoding, adaptive attention aggregation, and online updating yields robust and scalable forecasting performance in diverse power grid environments.

Table 5.

Main forecasting performance on the open power system data platform (mean ± std over 10 runs).

Model	MAE	RMSE	MAPE (%)
ARIMA	142.36 ± 2.11	198.74 ± 2.83	4.12 ± 0.07	0.901 ± 0.005
SVR	134.58 ± 1.96	189.25 ± 2.41	3.87 ± 0.06	0.914 ± 0.004
LSTM	120.47 ± 1.52	171.83 ± 2.07	3.42 ± 0.05	0.938 ± 0.003
CNN-LSTM	113.92 ± 1.44	163.75 ± 1.96	3.21 ± 0.05	0.946 ± 0.002
Informer	107.38 ± 1.29	155.46 ± 1.81	3.02 ± 0.04	0.953 ± 0.002
LightTS	110.84 ± 1.37	159.28 ± 1.89	3.14 ± 0.04	0.949 ± 0.002
Proposed Model (ADLP)	100.26 ± 1.05	146.92 ± 1.63	2.78 ± 0.03	0.964 ± 0.001

Table 6 presents a quantitative comparison of model complexity in terms of parameter size and floating-point operations. Classical statistical methods such as ARIMA and SVR exhibit negligible parameter counts and computational cost, but their limited representational capacity leads to inferior forecasting accuracy. Among deep learning approaches, Informer contains 4.52M parameters and requires 6.94G FLOPs due to multi-head self-attention over long input sequences. Although Informer achieves strong predictive performance, its computational overhead is significantly higher than recurrent-based models. The proposed ADLP contains 1.48M parameters and requires 2.03G FLOPs, positioning it between CNN-LSTM (1.36M, 1.87G) and Informer in terms of complexity. Compared with Informer, ADLP reduces parameter size by approximately 67% and lowers computational cost by 70.8%, while still achieving superior forecasting accuracy as shown in previous tables. The moderate increase in complexity compared with vanilla LSTM (0.84M) is attributed to the hybrid temporal encoding structure and the dynamic attention module. However, this additional cost yields substantial accuracy gains, indicating an effective trade-off between expressiveness and efficiency. LightTS demonstrates the lowest complexity among deep learning models with 0.58M parameters and 0.76G FLOPs, but its forecasting accuracy remains below ADLP. These results confirm that the proposed architecture achieves a balanced design: the hybrid Bi-LSTM and CNN structure enhances temporal representation without incurring the quadratic attention cost of transformer-based models, and the structured attention aggregation remains computationally efficient. Overall, ADLP provides a favorable accuracy-efficiency trade-off suitable for practical smart grid deployment.

Table 6.

Efficiency comparison in terms of parameter size and computational complexity.

Model	Params (M)	FLOPs (G)
ARIMA	0.002	0.003
SVR	0.018	0.021
LSTM	0.84	1.12
CNN-LSTM	1.36	1.87
Informer	4.52	6.94
LightTS	0.58	0.76
Proposed Model (ADLP)	1.48	2.03

Ablation study

This section investigates the contribution of individual modules and the sensitivity of key hyperparameters in the proposed framework. The ablation design follows the methodological components defined in the formulation. The primary ablation focuses on three core modules: Hybrid Temporal Encoding (Bi-LSTM + CNN fusion), Dynamic Attention aggregation, and External Variable Extension incorporating temperature features. Each component is removed independently while keeping all other configurations unchanged. Evaluation is conducted on the AEMO Electricity Demand Dataset and the Open Power System Data platform using MAE, RMSE, MAPE, and under identical training, data splits, and statistical settings as in the comparative experiments. In addition, a sensitivity and robustness study is performed on three hyperparameters closely related to temporal modeling capacity: hidden dimension size, historical window length, and learning rate. Robustness is further evaluated by injecting Gaussian noise into input load sequences to assess stability under measurement perturbations. All experiments are repeated for 10 runs with statistical reporting to ensure reproducibility and reliability.

Table 7 presents the primary ablation results to assess the contribution of different modules in the proposed Adaptive Deep Load Predictor (ADLP) on the AEMO Electricity Demand Dataset. The ablation study involves removing key components such as the Hybrid Temporal Encoding, Dynamic Attention, and External Variables to understand their respective impacts on forecasting accuracy. Removing the Hybrid Temporal Encoding module leads to an increase in MAE from 92.18 to 98.34, a 6.6% performance degradation. This indicates that the combination of Bi-LSTM and CNN in the hybrid temporal encoding is essential for capturing both long-term dependencies and local temporal structures in the load data, which significantly improves forecasting precision. The RMSE also increases from 131.54 to 141.62, highlighting the deterioration in peak load prediction when this component is excluded. Excluding the Dynamic Attention mechanism results in a less severe drop, with MAE rising to 97.42 (a 5.6% increase). This suggests that while the dynamic weighting of historical states through attention is beneficial, the model still performs well without it. The RMSE also worsens to 139.92, confirming that adaptive weighting of historical data improves accuracy by focusing on the most relevant time periods. When the external temperature variables are removed, the performance suffers slightly more with MAE increasing to 96.10, a 4.3% degradation from the full model. This result illustrates the value of incorporating external environmental factors, such as temperature, which help refine the forecasting model by capturing additional variations in load patterns influenced by external conditions. In all three cases, the consistent degradation in performance emphasizes that each module, Hybrid Temporal Encoding, Dynamic Attention, and External Variables, provides complementary contributions. This confirms that the architectural design is structurally coherent and that removing any component negatively impacts the overall model’s predictive capability.

Table 7.

Main ablation results on the AEMO electricity demand dataset and the open power system data platform.

Configuration	MAE	RMSE	MAPE (%)
Full model	92.18 ± 0.97	131.54 ± 1.41	2.53 ± 0.03	0.968 ± 0.001
w/o hybrid temporal encoding	98.34 ± 1.08	141.62 ± 1.72	2.85 ± 0.04	0.960 ± 0.002
w/o dynamic attention	97.42 ± 1.14	139.92 ± 1.52	2.81 ± 0.04	0.963 ± 0.002
w/o external variables	96.10 ± 1.25	137.85 ± 1.59	2.72 ± 0.03	0.965 ± 0.002

Table 8 presents the sensitivity and robustness results from varying several key hyperparameters in the proposed Adaptive Deep Load Predictor (ADLP) model on the AEMO Electricity Demand Dataset. These experiments aim to evaluate the model’s representation capacity, optimization stability, and robustness under different operational conditions. The first set of experiments investigates the effect of the hidden dimension size on performance. As shown, increasing the hidden dimension from 64 to 128 results in a slight improvement in MAE, RMSE, and MAPE, with MAE reducing from 94.21 to 92.18. Further increasing the hidden dimension to 256 causes a minor degradation in performance, with MAE rising to 92.75. This suggests that while larger hidden dimensions improve representation capacity, they also introduce the risk of overfitting, leading to diminishing returns. Next, the effect of historical window length on forecasting performance is evaluated. Reducing the window to 96 time steps results in a slight increase in MAE (93.62), indicating that smaller windows may not capture enough temporal context. Increasing the window to 240 time steps leads to a marginal increase in MAE to 92.85, suggesting that excessively long windows may introduce unnecessary complexity without providing substantial benefits. The learning rate was also varied to assess optimization stability. As expected, reducing the learning rate to causes a slight increase in MAE (94.10), while increasing the learning rate to leads to a small increase in MAE to 93.38. These results confirm that a learning rate of strikes a good balance for model convergence. The Gaussian noise injection experiment tests the robustness of ADLP against perturbations in the load data. Injecting noise causes a noticeable increase in MAE to 95.47, highlighting the importance of robust training for stable real-world deployment. These results show that the proposed model maintains structural stability and performs well across a range of hyperparameters, with only moderate degradation in performance under slight noise and parameter variations. This demonstrates the model’s generalization capability across different operating conditions.

Table 8.

Sensitivity and robustness analysis on the AEMO electricity demand dataset and the open power system data platform.

Setting	MAE	RMSE	MAPE (%)
Hidden Dim = 64	94.21 ± 1.05	135.87 ± 1.56	2.61 ± 0.03	0.965 ± 0.002
Hidden Dim = 128	92.18 ± 0.97	131.54 ± 1.41	2.53 ± 0.03	0.968 ± 0.001
Hidden Dim = 256	92.75 ± 1.04	133.45 ± 1.48	2.59 ± 0.04	0.967 ± 0.002
Window = 96	93.62 ± 1.11	136.72 ± 1.61	2.66 ± 0.04	0.965 ± 0.002
Window = 168	92.18 ± 0.97	131.54 ± 1.41	2.53 ± 0.03	0.968 ± 0.001
Window = 240	92.85 ± 1.08	133.92 ± 1.56	2.57 ± 0.04	0.967 ± 0.002
Learning Rate =	94.10 ± 1.12	137.56 ± 1.73	2.69 ± 0.04	0.964 ± 0.002
Learning Rate =	92.18 ± 0.97	131.54 ± 1.41	2.53 ± 0.03	0.968 ± 0.001
Learning Rate =	93.38 ± 1.16	135.32 ± 1.62	2.62 ± 0.03	0.966 ± 0.002
Gaussian noise injection	95.47 ± 1.21	139.18 ± 1.76	2.76 ± 0.04	0.962 ± 0.003

To strengthen the alignment between methodological design and empirical results, additional analysis is provided to clarify the mechanisms underlying the observed forecasting improvements. The effectiveness of the Hybrid Temporal Encoding module is closely related to the multi-scale temporal characteristics of electricity load data. In datasets such as AEMO, which exhibit half-hourly resolution and strong daily and weekly seasonality, local short-term fluctuations coexist with longer cyclical dependencies. The convolutional branch enhances sensitivity to localized ramps and abrupt demand shifts, while the bidirectional LSTM captures extended temporal dependencies across weekly patterns. This complementary structure explains the consistent reductions in RMSE and MAE, particularly during peak-load intervals and transitional periods. For smoother hourly datasets such as OPSD, the primary benefit arises from improved long-range representation stability rather than short-term fluctuation modeling. The hybrid structure reduces representation bias over extended horizons, leading to improved multi-step forecasting consistency. These observations indicate that the proposed architecture adapts to variations in temporal resolution and seasonal intensity. The Dynamic Attention mechanism further improves performance by selectively emphasizing historically relevant temporal segments. When forecasting over multi-step horizons, not all historical observations contribute equally to future load estimation. Attention weighting reduces interference from temporally distant yet structurally irrelevant segments, particularly during seasonal transitions or irregular demand spikes. The observed stability improvements across extended prediction windows are therefore consistent with the intended function of dynamic temporal weighting. Sensitivity analysis also indicates that a historical window length of 168 time steps provides an effective balance between contextual richness and generalization capacity. Shorter windows underrepresent weekly periodic behavior, while excessively long windows introduce redundancy that increases optimization difficulty without proportional predictive benefit. These findings confirm that the architectural configuration aligns with intrinsic load periodicity patterns rather than functioning as an arbitrary parameter selection.

Operational-level evaluation of DGLO

To evaluate whether forecasting improvements translate into operational benefits, an operational-level simulation of the DGLO framework is conducted using forecast outputs as scheduling inputs under a unified control configuration. Dispatch cost, reserve allocation stability, renewable curtailment rate, and energy storage state-of-charge (SoC) variance are adopted as operational performance indicators. All control parameters are kept identical across models to ensure that differences arise solely from forecasting quality. Table 9 presents the comparative operational results under identical dispatch and storage coordination settings. The results indicate that integrating ADLP forecasts within the DGLO framework leads to measurable operational improvements beyond traditional forecasting metrics. Reduced dispatch cost reflects lower corrective generation adjustments caused by forecast error propagation. Decreased reserve violation frequency suggests improved stability in demand–supply balancing. Lower renewable curtailment rates indicate enhanced utilization of intermittent energy resources, while reduced SoC variance demonstrates smoother storage cycling behavior. These findings confirm that forecasting accuracy improvements are not limited to statistical evaluation but produce tangible system-level gains when embedded within a structured forecast-driven operational mechanism. The added analysis therefore strengthens the empirical linkage between architectural design, forecasting performance, and downstream grid operation outcomes.

Table 9.

Operational-level comparison under identical control configuration (relative performance; lower values indicate better performance).

Model	Dispatch cost	Reserve violation rate	Renewable curtailment	SoC variance
LSTM + Standard control	1.00	1.00	1.00	1.00
Informer + Standard control	0.96	0.94	0.95	0.93
ADLP + DGLO	0.91	0.88	0.89	0.85

Forecast-operation coupling mechanism analysis

To further clarify the theoretical contribution of improved forecasting performance, an explicit forecast operation coupling analysis is conducted. The objective is to examine how variations in prediction error propagate into operational constraints and economic outcomes under different grid scenarios. Three operational scenarios are considered: normal demand conditions with moderate reserve margin, high renewable penetration with increased variability, and constrained reserve capacity with tighter stability margins. For each scenario, controlled perturbations are introduced to forecast outputs to simulate varying error levels, while keeping control parameters identical across models. Table 10 reports the relative operational sensitivity to forecast error under different scenarios. The results indicate that forecasting improvements yield increasingly significant operational benefits as system constraints become more restrictive. Under normal operating conditions, performance differences remain moderate because reserve margins absorb forecast uncertainty. However, in high renewable penetration environments, forecast errors directly influence balancing requirements due to increased variability. The ADLP-based configuration demonstrates lower operational sensitivity, indicating reduced corrective adjustment magnitude and improved stability. Under tight reserve constraints, prediction errors propagate more strongly into dispatch cost and reliability indicators. In this scenario, improved forecasting reduces reserve over allocation and mitigates violation risk, thereby producing measurable economic and stability gains. These findings demonstrate that forecasting accuracy matters most when operational flexibility is limited or variability is high, thereby providing a clear scenario dependent explanation of forecast value. This analysis establishes an explicit linkage between prediction performance and operational outcomes. Rather than treating forecasting accuracy as an isolated metric, the results show that reduced forecast error leads to lower operational sensitivity, improved stability margins, and enhanced economic efficiency under realistic constraint scenarios. The contribution therefore extends beyond engineering implementation by clarifying the conditions under which forecasting improvements translate into systemic value.

Table 10.

Operational sensitivity to forecasting error under different grid scenarios (relative impact; lower values indicate better stability).

Model	Normal conditions	High renewable penetration	Tight reserve constraint
LSTM-based forecast	1.00	1.00	1.00
Informer-based forecast	0.94	0.89	0.87
ADLP-based forecast	0.90	0.82	0.78

Conclusions and future work

This study introduced ADLP, an adaptive deep learning framework for power grid load forecasting, and DGLO, a dynamic optimization strategy for grid-aware load adjustment. Through extensive experiments on four benchmark datasets, AEMO, OPSD, Pecan Street, and RE-Europe, the ADLP model demonstrated consistent improvements over traditional machine learning and state-of-the-art deep learning models across multiple evaluation metrics, including MAE, RMSE, and MAPE. Ablation studies confirmed the complementary value of each architectural component, particularly the hybrid temporal encoding and online adaptive updating mechanisms. Additional experiments also validated the effectiveness of the DGLO strategy in improving prediction stability and grid responsiveness under dynamic conditions.

Despite these promising results, several limitations remain. The model relies on the availability of high-quality, temporally aligned input data, including weather and pricing signals. Performance may degrade in data-sparse or noisy environments. The computational demands of real-time adaptation also present a challenge for low-resource deployment settings.

Future work will explore model compression techniques, including pruning and quantization, to reduce runtime cost. Another promising direction is the integration of reinforcement learning-based control agents within the DGLO framework to improve autonomous grid interaction. Further, expanding evaluation to include multi-region datasets and real-time edge device deployment will help assess the generalization and scalability of the proposed system in large-scale smart grid scenarios.

Method

Overview

With the rising electricity demand and the integration of renewable energy sources, predicting and managing power grid loads has become increasingly vital. This section introduces the methodology developed in this work to model and analyze power grid load, focusing on predictive modeling and optimization strategies.

Section “Preliminaries” introduces the fundamental concepts and mathematical formulations of power grid load dynamics. It provides the necessary preliminaries, including definitions, notations, and the mathematical representation of load variations over time. The proposed predictive model is presented in “Adaptive deep load predictor (ADLP)”. Unlike traditional load forecasting approaches, which rely solely on statistical methods or conventional machine learning, our model integrates advanced deep learning techniques. The methodology section describes the architecture, training process, and input features considered in the predictive framework. A novel optimization strategy in “Dynamic grid-aware load optimization (DGLO)” is introduced to enhance the stability and efficiency of power grid operation. This strategy leverages domain-specific constraints and real-time data to improve decision-making in load distribution and grid management. By structuring the methodology into these interconnected components, this paper aims to provide a comprehensive solution for accurate load forecasting and efficient power grid management.

Although both ADLP and DGLO are described within a unified methodological framework, the primary experimentally validated contribution of this study lies in short-term load forecasting. ADLP constitutes the core methodological component and the main subject of empirical evaluation. DGLO is formulated as a forecast-driven operational enhancement layer that illustrates how predictive outputs can support grid-aware decision-making, rather than as a standalone, fully benchmarked control or optimization system. This distinction clarifies the hierarchical structure of the proposed framework.

Preliminaries

The power grid load can be defined as the total electrical demand imposed on a power system at a given time. Let the load function be represented as:

1

where L(t) is the total load at time t, is the power consumption of the i-th consumer, and N represents the total number of consumers connected to the grid.

To model the temporal behavior of power demand, we consider the load as a time series:

2

where T is the total number of time steps in the observation window.

The load signal can be decomposed into three main components:

3

where represents the base load (long-term average demand), captures systematic variations due to daily and seasonal trends, and denotes random fluctuations.

To model , we introduce a periodic function:

4

where and are the amplitude and phase shift of the k-th harmonic component, respectively.

The power load exhibits stochastic behavior due to consumer activities and external factors. Thus, we model L(t) as a random variable:

5

where is the expected load and represents the variance.

Given historical load data , the objective of load forecasting is to estimate the future load for . Mathematically, this can be expressed as:

6

where is a predictive function parameterized by .

For improved accuracy, external factors such as temperature , humidity H(t), and economic indicators E(t) can be incorporated:

7

The power grid must balance generation and consumption to maintain stability. Let G(t) be the total generation at time t, then the system stability constraint is:

8

where R(t) is the reserve margin required to handle sudden fluctuations. The reserve margin is typically modeled as:

9

where is a predefined safety factor.

Load fluctuations are affected by demand elasticity, consumer behavior, and grid constraints. The rate of change of load can be expressed as:

10

where represents the natural decay factor of demand response, is a scaling factor, and U(t) is an external control input.

Adaptive deep load predictor (ADLP)

Accurate power grid load forecasting is crucial for ensuring stability and optimizing energy management. In this section, we introduce three key innovations of the Adaptive Deep Load Predictor (ADLP) that enhance forecasting accuracy and adaptability (as shown in Fig. 1).

Figure 1.

Illustration of the adaptive deep load predictor (ADLP) framework. The figure demonstrates how each hidden state is assigned an attention weight through a softmax scoring function, forming a probability distribution over past observations. The weighted sum produces the context vector c, which is further transformed via a nonlinear layer to capture complex temporal dependencies. The refined context vector is then concatenated with the latest hidden state to generate the final output representation. This mechanism enables the model to selectively emphasize relevant temporal information while suppressing redundant signals.

Hybrid Temporal Encoding

To effectively model power load variations, the Adaptive Deep Learning Predictor (ADLP) integrates both short-term and long-term dependencies through a hybrid encoding approach that combines bidirectional long short-term memory (Bi-LSTM) networks and convolutional neural networks (CNNs). This design allows ADLP to capture global sequential relationships while extracting essential local patterns, thereby enhancing predictive accuracy. The encoding mechanism can be mathematically expressed as follows:

11

where represents the hidden state at time step t, Bi-LSTM captures sequential dependencies by processing the input in both forward and backward directions, while CNN extracts spatial correlations among neighboring features. Specifically, the Bi-LSTM update rule is given by:

12

where , , and denote weight matrices and bias terms, respectively, while is a nonlinear activation function. Concurrently, the CNN component applies convolutional filters to refine local patterns:

13

where represents the convolutional filter, denotes the convolution operation, is a bias term, and is the activation function. The final encoded representation integrates both components:

14

where is a nonlinear transformation ensuring effective fusion of long-term dependencies and local feature extraction. This hybrid encoding strategy empowers ADLP to capture multi-scale dependencies within power load data, leading to improved forecasting performance.

Dynamic Attention Mechanism

To effectively capture relevant temporal dependencies in sequential data, a dynamic attention mechanism is employed to assign adaptive importance weights to past observations. This mechanism dynamically adjusts the influence of each historical hidden state based on its relevance to the current context, thereby enhancing the model’s ability to focus on crucial past information while mitigating the impact of irrelevant or redundant details (as shown in Fig. 2). The attention mechanism is formulated as follows:

15

where represents the attention weight for the hidden state , and is a learnable parameter vector that determines the significance of . The attention weights are computed using a softmax function to ensure they sum to one, thus forming a probability distribution over past observations.

Figure 2.

Dynamic attention mechanism (DAM) for multi-modal feature fusion. The figure demonstrates how each hidden state is assigned an attention weight through a softmax scoring function, forming a probability distribution over past observations. The weighted sum produces the context vector c, which is further transformed via a nonlinear layer to capture complex temporal dependencies. The refined context vector is then concatenated with the latest hidden state to generate the final output representation. This mechanism enables the model to selectively emphasize relevant temporal information while suppressing redundant signals.

The computed attention weights are then utilized to derive a context vector c, which serves as a weighted sum of the past hidden states:

16

To further refine the attention mechanism, an additional transformation layer can be introduced to incorporate non-linear interactions, enhancing the model’s ability to capture complex dependencies:

17

where is the transformed context vector, is a learnable weight matrix, and is a bias term. The transformation introduces non-linearity, allowing the model to capture intricate relationships between historical and current states.

The refined context vector is used to compute the final output representation, integrating both the attention-enhanced historical information and the present hidden state:

18

where is the output representation, is a weight matrix, is a bias term, and denotes the concatenation of the transformed context vector and the latest hidden state. The function represents an activation function, such as a sigmoid or softmax, depending on the specific task. This comprehensive approach ensures that the model effectively utilizes past information while maintaining adaptability to evolving input patterns.

Online Adaptive Updating

To ensure adaptability to real-time variations in load patterns, Adaptive Deep Learning Predictor (ADLP) integrates an online updating mechanism based on gradient-based optimization. This mechanism enables continuous refinement of the prediction model parameters, allowing it to dynamically adjust to changing data distributions. The parameter update is performed iteratively using the following gradient descent update rule:

19

where represents the forecasting loss function, is the learning rate, and is the set of model parameters being optimized. To further improve model stability and convergence speed, an adaptive learning rate strategy can be incorporated, expressed as:

20

where is a decay factor that gradually reduces the learning rate as iterations progress, preventing excessive oscillations during optimization. To mitigate the impact of abrupt changes in the load pattern, momentum-based optimization can be employed:

21

22

where v represents the velocity term that accumulates past gradients, and is the momentum coefficient that controls the influence of previous updates. By integrating these mechanisms, ADLP enhances its adaptability to real-time variations, ensuring more accurate and robust load forecasting under dynamically changing conditions.

In terms of practical deployment, the proposed ADLP model is designed with real-time applicability in mind. To ensure efficient operation in streaming environments, the online adaptive updating module leverages lightweight gradient-based updates on mini-batches of new data without requiring full retraining. The model parameters are fine-tuned using stochastic gradient descent based on a short sliding window of the most recent observations, enabling it to adapt to evolving load patterns with minimal computational overhead. We conducted performance profiling to assess the model’s suitability for real-time applications. On an NVIDIA A100 GPU, the average inference latency per sample is approximately 6.9 milliseconds, as shown in Table 6, which is well within the acceptable range for most real-time smart grid control systems. The online update step, when triggered periodically or based on drift detection, requires less than 15 milliseconds per iteration due to the selective update of only a subset of network layers (like attention weights or final linear projections), rather than the entire model. This efficient updating mechanism allows ADLP to continuously refine its predictions in response to new input conditions, such as sudden demand spikes or renewable output variations. The architecture is compatible with edge or cloud-based deployment frameworks, and its real-time performance makes it suitable for integration into adaptive grid control systems that require both fast inference and ongoing learning.

Dynamic grid-aware load optimization (DGLO)

Accurate load forecasting is only one part of an effective power grid management system. To ensure stability and efficiency, it is necessary to dynamically optimize load distribution based on real-time predictions. In this section, we introduce three key innovations of the Dynamic Grid-Aware Load Optimization (DGLO) strategy (as shown in Fig. 3).

Figure 3.

Dynamic grid-aware load optimization (DGLO) framework. The DGLO framework enhances grid reliability and energy efficiency by integrating predictive adaptive control, stochastic robust optimization, and energy storage integration. Predictive adaptive control utilizes deep learning and PID-based adjustments to dynamically balance power load, ensuring stability under varying demand conditions. Stochastic robust optimization incorporates probabilistic modeling and feedback mechanisms to mitigate uncertainties in load forecasting, enabling real-time generation adjustments. Energy storage integration dynamically manages charge-discharge cycles to optimize energy utilization and stabilize grid operations. These components work collectively to improve real-time adaptability and resilience against demand fluctuations.

Predictive adaptive control

To effectively model power load variations and improve real-time load balancing, the Adaptive Deep Learning Predictor (ADLP) and the Dynamic Grid Load Optimizer (DGLO) integrate advanced control mechanisms that dynamically adapt to changing conditions. The DGLO employs an adaptive control mechanism that continuously updates the generation plan based on forecast error dynamics. This is achieved through a proportional-integral-derivative (PID) control approach, formulated as:

23

where represents the prediction error between actual and estimated load, and are the PID control gains that regulate generation adjustments. This predictive control mechanism ensures grid stability under varying demand conditions by dynamically responding to load fluctuations.

The proportional term reacts to the immediate error, making quick adjustments to compensate for rapid changes in demand. The integral term accumulates past errors, addressing steady-state deviations by fine-tuning generation to avoid sustained discrepancies. The derivative term anticipates future errors based on the rate of change, helping to minimize overshooting and improve system responsiveness.

To enhance the accuracy of load predictions, a Bi-LSTM-based prediction model is employed, which captures both long-term dependencies and short-term variations in power load data. This predictive model is formulated as:

24

where represents the hidden state, Bi-LSTM (Bidirectional Long Short-Term Memory) captures sequential dependencies in the power load time series, and CNN (Convolutional Neural Network) extracts localized patterns and trends from input data. This combination enhances the robustness of load forecasting by integrating both historical trends and short-term fluctuations.

The error-corrected load forecast is then computed as:

25

where is a nonlinear function that maps the encoded hidden states to the predicted load. Based on this forecast, the optimized control input ensures adaptive correction of power generation:

26

To further refine the system’s response to varying load conditions, a feedforward compensation mechanism is introduced to preemptively adjust generation based on real-time variations in renewable energy input and external disturbances. The feedforward correction term is defined as:

27

where and are weighting factors that balance the contribution of renewable energy sources and external disturbances such as unexpected grid fluctuations. The final generation adjustment integrates this feedforward term to enhance system stability:

28

This hybrid approach integrates predictive deep learning with adaptive control and feedforward compensation, enabling the DGLO system to dynamically adjust power generation while accounting for real-time demand shifts and external influences. By continuously refining load forecasts and optimizing power generation, this strategy enhances grid stability, improves energy efficiency, and ensures real-time load balancing under dynamic demand conditions.

Stochastic robust optimization

To account for forecasting uncertainties in power generation scheduling, DGLO incorporates a stochastic adjustment model that dynamically compensates for variations in demand and supply fluctuations. This model ensures robustness by integrating a probabilistic correction term into the nominal generation schedule, allowing for real-time adjustments that enhance system stability and reliability (As shown in Fig. 4). The stochastic adjustment mechanism is formulated as follows:

29

where represents the actual power generation at time t, is the scheduled nominal generation, and is a stochastic adjustment term modeled as a Gaussian random variable:

30

where represents the variance of generation uncertainty. This probabilistic component accounts for random fluctuations and enhances the system’s adaptability to real-time variations in demand.

Figure 4.

Stochastic robust optimization for power generation scheduling. The diagram illustrates a dual attention mechanism integrating spatial and channel attention to enhance feature extraction in forecasting models. Combined with a stochastic adjustment model, the approach dynamically compensates for uncertainties in power generation by incorporating probabilistic and feedback correction terms. This ensures real-time adaptability to demand fluctuations, minimizing deviations from forecasted schedules and improving system stability.

To further improve stability, a feedback correction term is introduced, which dynamically adjusts generation based on observed deviations from the forecasted demand:

31

where is a gain parameter that determines the sensitivity of the generation adjustment, is the actual demand, and is the forecasted demand. This formulation enables the generation system to compensate for deviations in demand in real-time, ensuring stable operation.

The final scheduled generation is then expressed as a combination of both deterministic and stochastic components, including an adaptive term that minimizes overall system deviations:

32

where is an additional noise term representing small-scale variations that are not captured by the Gaussian uncertainty model. By integrating these stochastic and feedback mechanisms, DGLO improves the resilience and adaptability of power generation scheduling, ensuring a robust response to unpredictable demand variations.

Energy storage integration

To smooth grid load fluctuations and improve reliability, the Dynamic Grid Load Optimization (DGLO) system integrates energy storage systems with dynamic charge-discharge management. By dynamically adjusting the charging and discharging rates based on real-time grid demand, this system enables efficient energy utilization. The evolution of stored energy in the system is governed by:

33

where and denote charge and discharge efficiencies, respectively. To optimize energy storage operations, a state-of-charge constraint is imposed to ensure that the system operates within safe limits:

34

where and represent the lower and upper storage capacity limits. This constraint ensures the system avoids overcharging or depletion, maintaining reliable performance.

To optimize charge-discharge scheduling, DGLO employs a cost function that minimizes energy losses while improving grid stability:

35

where and represent the cost coefficients for charging and discharging energy. This function helps minimize operational costs while maximizing energy efficiency.

A demand-response mechanism is incorporated to dynamically adjust power transactions based on real-time pricing signals:

36

where and are weighting factors that balance grid demand and renewable energy availability. This strategy allows the system to effectively manage load variations and maximize the use of renewable energy, reducing reliance on conventional power sources.

Author contributions

Conceptualization, JH; methodology, HH; software, DX; validation, HH; formal analysis, SZ; investigation, WC; data curation, DX; writing–original draft preparation, JH, HH, WC, SZ, DX and SY; writing–review and editing, DJ; visualization, WC; supervision, SY; funding acquisition, JH; All authors have read and agreed to the published version of the manuscript.

Funding

Details of all funding sources should be provided, including grant numbers if applicable. Please ensure to add all necessary funding information, as after publication this is no longer possible.

Data availability

The datasets generated and/or analysed during the current study are available in the SmartLoadNet, https://github.com/HCAAc7XY7G/SmartLoadNet.

Declarations

Competing interests

The authors declare no competing interests.