Optimization based load forecasting and demand management in smart building microgrids with Greylag Goose and Bi level graph models

Abstract

The effective use of energy in Smart Building Microgrids (SBMGs) largely depends on accurate load prediction and synchronized demand response especially with battery degradation and unreliability of renewable power. In spite of progress, problems of inaccuracies in forecasting, mismatch in demand and supply, and non-optimal optimization methods still ruin the reliability of the system and the durability of battery resources. A domain-adapted forecasting and optimization framework is proposed in this paper intending to combine Greylag Goose Optimization (GGO) with a Relational Bi-Level Aggregation Graph Convoluted Network (RBAGCN) to be used in SBMGs. In this context, the RBAGCN has been reengineered to incorporate physical and operating interrelations among the energy variables, and the GGO has been utilized to stabilize network weight convergence when the load is non-stationary, as opposed to being an independent optimizer. Before the training of a model, Fast Resampled Iterative Filtering (FRIF) is used to clean up and normalize historical sequential data and Prairie Dog Optimization (PDO) is used to remove the least salient features, i.e. three phase discharge power, battery discharge power, time of day, solar voltage, and ambient temperature. The modified RBAGCN performs load forecasting and demand-based management forecasting, and the GGO automatically adjusts the model parameters to improve convergence strength. As the simulation experiments with the MATLAB R2022b show, the proposed framework achieves an average forecasting accuracy of 98.3%, which is better than the benchmark models, including RNN (82.6%), ABMO-ANN (85.4%), RNN-LSTM (91.7%), GA-DNN (88.5%), and RNN-GRU (90.3%). In addition to this, the framework has lower error measures, such as mean absolute error of 0.0164, mean absolute percentage error of 0.0128, and mean squared error of 0.0069 as compared to benchmark values of 0.042, 0.037 and 0.031 respectively. Moreover, prediction variance and convergence iterations are decreased to 26.5% and 17.8%, respectively, which suggests a higher level of statistical stability and demand management learning efficiency of SBMG load prediction and battery-conscious demand management.

Introduction

1. Background

A variety of components are involved in the development of load forecasting and demand management (DM) by the means of enhanced battery performance in shared building microgrids (SBMGs)¹. One, past data and forecasting models are used to obtain accurate energy demand forecasting that improves the process of planning and resource management². Then, smart energy systems, including sustainable plans that match the predicted demand and renewable energy production, minimize wasted energy³. It also includes the dynamically controlled demand which varies the consumption according to supply and system priorities to enhance load flexibility⁴. The limitations to the battery performance and lifecycle are tackled through the regulation of the charging and discharging cycles in order to achieve high efficacy and performance in the storage⁵. The pre-processing of data involves outlier detection, normalization, and time synchronization of data, thus having quality data to be utilized in forecasting models⁶. Adaptive control reacts allow the system to respond to the change in consumption, generation, and storage to ensure the energy balance⁷. Resource scheduling algorithms are efficient to handle the generation, consumption, and storage to achieve maximum efficiency⁸.

• (b)Challenges

Complex issues that are interdependent surround integrated load forecasting and DM with advanced battery performance in SBMGs. The nature of energy consumption is quite variable and complex, which makes it hard to achieve high forecasting accuracy⁹. The pre-processing and integration process of large amounts of heterogeneous data of different sources entails a robust mechanism. There can be a high level of computation requirements to synchronise load forecasts and DM decisions. Battery functionality also impairs with age, which is a barrier to uniform storage functionality and general system dependability. The periods of renewable resources introduce uncertainty in energy supply and demand events. The technical accuracy is required to design scalable and modular solutions that can fit any building design and not impact the performance of the building¹⁰. It is complex protocols that are needed to establish safe, privacy-sensitive communication in the microgrid. The challenge of optimising control strategies with conflicting objectives, e.g. cost savings versus energy savings, brings out more algorithmic issues. There is a lack of standardised performance evaluation measures and validation structures which makes it hard to ensure that there are coherent performance measures in the emerging systems.

• (c)Literature Survey

Several research studies have focused on Integrated Load Forecasting and DM in SBMGs, utilizing different techniques and approaches. A review of some of these works is provided below:

Previous research have presented a developed an RNN-based predictive control system for community Energy Management (EM) with the goal of using customers’ available energy flexibility to actively engage them in energy trading¹¹. Researchers presented ABMO-ANN for short-term electricity load demand prediction within Microgrid Energy Management Systems (MEMS)¹². This method improves both feature selection and parameters optimization, mitigating the issues of forecasting performance in complicated and temporal MG contexts. An RNN-LSTM method to anticipate Renewable Energy Generation (REG) and Electric Power Consumption (EPC) in a Science Block (SCB) at one- and five-minute intervals. The above was meant to help guide the development of the DT and DM programs¹³.

A GA-DNN for predicting the net electric load in a MG environment, where the load curve was influenced by multiple external factors. An improved forecasting method based on RNN-GRU for estimating electrical load in smart grid settings¹⁵. The process begins with the use of correlation analysis to select the input variables, having established a range of relevant historical power use attributes as source data. A day-ahead optimal scheduling model powered by artificial intelligence for a grid-connected AC MG¹⁶. This framework permits GA to optimize the dispatch scheduling of batteries and generate demand response. Subsequently, PV generation and building load consumption were predicted with Light Gradient Boosting Machine (Light GBM). A cascade-structured Artificial Neural Network (ANN) model integrated with Particle Swarm Optimization (PSO) to support a self-adaptable EM system aimed at enhancing MG performance. The measurement of load forecasting and DM to improve battery performance is shown in Table 1.

Table 1.

Literature Comparison of Integrated Load Forecasting and DM in Smart Building Solar MGs.

References	Technique	Objective	Limitations
11	RNN-based predictive control system	Community EM using customer flexibility and matrix-based control for energy trading	Scalability challenges and computational complexity in diverse community configurations
12	ABMO-ANN	Short-term load demand prediction in MEMS	High complexity in optimization and scalability issues in large systems
13	RNN-LSTM	Anticipate EPC and REG for DSM program construction	Challenges in forecasting highly erratic energy generation and consumption patterns
14	GA-DNN	Predict net electric load in a MG environment	Susceptibility to local optima in load pattern modeling
15	RNN-GRU	Electrical load forecasting in smart grids	Limitations in adapting to dynamic, unpredictable energy demand shifts
16	GA + LightGBM	Optimize battery dispatch scheduling for a grid-connected MG	Complexity in dealing with variations in energy demand and supply
17	ANN + PSO	Enhance MG performance through self-adaptable EM	Difficulty in optimizing for all MG parameters under diverse conditions

The Relational Bi-level Aggregation Graph Convolutional Network (RBAGCN) is one of these that proposes a two-step aggregation step, which considers not only the local dependencies between nodes but also the global relational hierarchies in the datasets including the time-varying information. The lower-level aggregation is learning the intra-node correlations within consecutive time periods and the upper-level aggregation is consolidating the inter-node associations. This allows the network to deal with spatially dispersed parameters- like temperature, irradiance and energy loads variations in Smart Building Microgrids (SBMGs). The stacked relation mechanism guarantees adaptive feature fusion and also reduces gradient degradation that is often witnessed in the traditional GCNs. Research using multi-level GCNs to predict spatiotemporal forecasts demonstrates better learning stability and decreased prediction Criterion which confirms RBAGCN is an effective framework to address the high-dimensional load forecasting tasks. The majority of deep-learning algorithms (RNN, LSTM, GRU) are able to model temporal patterns but do not model the spatial correlation of the interconnected microgrid variables and, on the other hand, hybrids based on optimization (ABMO-ANN, GA-DNN, PSO-ANN) are better converged but are still parameter-tuning sensitive and affected by overfitting. Furthermore, little has been done to combine adaptive feature-selection and graph-based representation learning into a common framework. The mentioned restrictions also point out the necessity of a collaborative architecture to merge relational feature aggregation with dynamic optimization, which drives the creation of the GGO-RBAGCN framework that is described in this paper.

• (d)Research Gap and Motivation

Although significant progress has been made on load forecasting and demand management within SBMGs, significant research gaps still exist in achieving accurate load prediction and optimally performing a battery. Prior use of RNN, LSTM, GRU, hybrid neural network models paired with metaheuristic optimizers suggest that using some combination of deep learning tools for modeling sequential data is useful. However, each provides its own challenges, such as poor generalization, over fitting, lack of robustness and computational inefficiency with more than two-objectives. Additionally, other methods used conventional pre-process methods and static feature selection techniques, which do not adequately account for the time and context of the problem for adaptive EM. In addition, the other research lacked the ability to update battery health indicators dynamically. The methods used to optimize the frameworks also do not offer reasonable robustness to deep learning parameter tuning issues associated with nonlinear and ever-evolving SBMG environments. Ultimately, there is a pressing need for an integrated intelligent framework combining advanced pre-process, dynamic interchangeable features, graph-based learning, and adaptive optimization—in order to face challenges associated with complex battery performance, load variability, and energy distribution within SBMGs. The most recent studies have also addressed attention-based and large-scale graph-based models such as STGCN, T-GCN, GAT, and Transformer models to be used in predicting power systems. However, the methodologies usually assume the denseness of spatial topology, the abundance of node-level measurement data, or grid-wide sensor placement, which, in smart building microgrids (SBMGs) with limited sensing capacity, a limited feature space, and high real-time computational needs are often impossible. Therefore, the current research has a higher priority in the formation of a feature-relational graph learning model particularly useful in the context of SBMG and thus avoiding the use of mass-scale grid-based attention models. To address these problems, this research investigates a hybrid framework which integrating Greylag Goose Optimization (GGO) algorithm and Relational Bi-level Aggregation Graph Convolutional Network (RBAGCN). while hybrid deep-learning optimization models approaches exist, most prior experiments focus classical optimizers like PSO, GA or Greek Wolf Optimizer in order to tune shallow or recurrent networks. This research, on the other hand, assesses how relational graph-based learning can be combined with bio-inspired optimization to solar-battery microgrid (SBMG) load forecasting, as well as the battery-sensitive demand management. As far as the authors know, there is no direct accounting of such an integration to be found in applications of SBMG, especially when it comes to the collective modelling of spatio-temporal energy interactions and convergence-stable learning.

• (e)Contribution

The following is a summary of main contributions of this manuscript:

• In this paper, a domain-adapted GGO -RBAGCN models that are useful in load forecasting and battery-oriented demand management in Smart Building Microgrids (SBMGs) are introduced.

• The new method incorporates FRIF for pre-processing raw data into cleaned, standardized, and formatted historical sequential data to ensure load forecasting.

• PDO is also used for feature selection to determine significant features while selecting only significant inputs back in the historical sequential model to improve accuracy of model output.

• GGO optimizes the weight parameters of RBAGCN, addressing optimization limitations and improving the adaptability and accuracy of demand prediction and EM.

• The proposed GGO-RBAGCN method significantly outperforms existing methods in terms of accuracy, demonstrating superior performance across all evaluation metrics, highlighting its effectiveness in load forecasting as well as EM.

• (f)Novelty

This research lies not in the presentation of new algorithms but in the domain specific adaptation and synthesis of the existing learning and optimisation strategies when applied to Smart Building Microgrid (SBMG) applications. Specifically, the Relational Bi-Level Aggregation Graph Convolutional Network (RBAGCN) which was initially developed in the context of relational data modelling has been redesigned to capture physical and functional interdependencies of the energy variables in SBMGs. Greylag Goose Optimization (GGO) is used to supplement the convergence stability and robustness of network weight optimisation in non-stationary load conditions, in place of being an independent novel optimiser. In addition, Fast Resampled Iterative Filtering (FRIF) and Prairie Dog Optimization (PDO) are used concurrently, which improves the quality of input signals and relevance of features before the learning step. By means of this systematic adaptation and integration, the suggested framework is capable of attaining a higher forecasting accuracy and battery-conscious demand-management on SBMG settings.

Proposed methodology

In this section, the proposed GGO-RBAGCN framework that will be used to predict loads and manage battery-focused demand in Smart Building Microgrids (SBMGs) is outlined. The methodology can be viewed as a processing pipeline with four interconnected steps that include Data acquisition; signal conditioning through Fast Resampled Iterative Filtering (FRIF); Adaptive feature selection based on Prairie Dog Optimization (PDO); and spatio-temporal load prediction via a Relational Bi-Level Aggregation Graph Convolutional Network (RBAGCN). All stages has the input of the results of its predecessor, thus guaranteeing the gradual improvement of the data fed in as well as the representations learned. The optimization is then used to stabilize training convergence of RBAGCN when in non-stationary load conditions, instead of being an independent optimization module, which is referred to as Greylag Goose Optimization (GGO). The general structure of the workflow of the proposed framework is shown in Fig. 1, and the separate parts are further discussed in the following subsections.

Fig. 1.

Flowchart of Proposed GGO-RBAGCN Methodology.

Data acquisition

This research gathered a variety of historic sequential data nearly one year long of buildings furnished with Smart Solar Microgrids (SSM) and Battery Management Systems (BMS). The dataset will consist of some internal parameters, including load power, battery status, solar energy, and grid information, and external environmental variables, including temperature, humidity, speed of wind, pressure, and solar radiation. Such an extensive data collection guarantees that both building and environmental variables required to make the correct prediction of loads are recorded. It has eleven features; it is possible to observe the list of features of the historical sequential data in Table 2. The Smart Solar Microgrid (SMG) testbed of Hindusthan College of Engineering and Technology in Coimbatore, India, was chosen as the source of the data collected. The ongoing logging lasted between January and December 2023 based on the calibrated IoT sensors that were part of the management system of the building. The system would record eleven operational and environmental variables after every five minutes, yielding over 105,000 pieces of information. Prior to analysis, filtered the records with missing or suspicious records due to sensor drifting or short communication failures. The dataset is likely to be biased because of the seasonal variation of the solar irradiance and building-specific load profiles that are a result of the users behavior. The way it dealt with these biases was to normalize data on daily and monthly levels, and to test the model on stratified time subsets. These measures ensured that the performance of the model remained the same over seasons and occupancy trends. Therefore, the detailed data provenance results in the fact that the GGO-RBAGCN model can be replicated and applied to different smart building microgrid configurations with high reliability.

Table 2.

List of features from Historical Sequential Data.

S.NO	FEATURE NAME
1	Three Phase Grid Current
2	Battery Discharged Current
3	Battery Discharged Power
4	Time of Day
5	Solar Voltage
6	Solar Current
7	Three Phase Grid Voltage
8	Power Factor
9	Temperature
10	Humidity
11	Wind Speed

Pre-processing using fast resampled iterative filtering

In this paper, FRIF is employed during pre-processing stage to prepare data for energy forecasting and DM within a SBMG ¹⁸. The raw data collected from integrated systems including solar generation units, battery storage, and environmental sensors undergoes a comprehensive cleaning and normalization process to enhance both model performance and battery life optimization¹⁹. To begin, missing data points will be reconstructed by employing the values of neighbouring time steps, while keeping the historical temporal order and continuity required for accurate predictive forecasting. Data will be checked for outliers and removed to avoid bias in model training and ensure that the behaviour of the system is accurately construed. The dataset will be normalized by use of min–max scaling. At this stage, all features will be normalised to the same range. Typically, normalisation removes discrepancies in scale between features such as power, temperature, and radiation, enabling the forecasting model to represent all input parameters equitably. This normalisation stage is of extreme importance in smart grids due to low balancing the importance of different input features, leading to suboptimal dispatching of energy or incorrect cycling of the battery. The FRIF method can reduce timeliness of this data transformation stage, and increase spatio-temporal congruently and the ability to reduce patterns within sequential data. General improvements in signal structure can also facilitate the models’ recognition of demand patterns, and by improving the battery use pattern for example, will result in increased battery life and better energy consumption or forecasting profiles. The traditional preprocessing schemes like the moving average or spline interpolation schemes assume non-stationary signals to be treated evenly. This may cleanse up and alter momentary load changes or battery load changes. The Fast Resampled Iterative Filtering (FRIF) algorithm is a method that will adapt to sequential data by resampling uneven time intervals and the procedure successively reduces the high-frequency noise retaining the major temporal features. FRIF maintains temporal continuity and aspect periodicity and eliminates disturbances of high frequency due to sensor drift or loss of data. The primary benefit of FRIF in SBMG applications is its capability to retain both temporal smoothness and amplitude integrity across the heterogeneous parameters, including power, voltage and temperature, and leads to cleaner input signals which enhance the stability and convergence of the RBAGCN model. FRIF can also be easily computed in real time with microgrids, as it is much faster than classical Empirical Mode Decomposition.

The empirical results of doing FRIF preprocessing are reflected in the experimental results, which show that the prediction variance decreases and the convergence in the forecasting model becomes more stable.

Prairie dog optimization of feature selection

In Smart Building Microgrids (SBMGs), feature selection is a crucial precondition on improving the strength of loading forecasting models and their overall generalisation in the analysis of demand-side control, as well as battery performance optimisation. The current study makes use of Prairie Dog Optimisation (PDO) as a feature selection mechanism on its own, hence specifying a small and highly effective subset of input variables of historical sequential data instead of implementing it as an optimisation algorithm in general. First, Pearson correlation is used to measure the statistical correlation between the individual candidate input features and the target load power. Attributes with a high level of correlation are kept and weakly correlated attributes are abandoned to reduce redundancy. In this respect, there are five characteristics, three-phase discharged power, battery discharged power, time of day, solar voltage and temperature that are discussed as most informative variables that affect load variation in SBMGs. PDO is then used to refine this set of correlated features to produce combinations that are most predictively relevant but least dimensional. The resulting small feature set leads to a better computational efficiency, reduced the risk of overfitting, and improved the model interpretability during further training of RBAGCN. The use of PDO is intentional to extract a small and high-impact feature set, and the necessity to use it is also supported by a sensitivity and variance inflation ratio (VIF) analysis, avoiding theoretical optimisation arguments. A sensitivity analysis was done to estimate loss of information by gradually adding excluded features of humidity, pressure, power factor, wind speed, grid voltage, and grid current in the RBAGCN model. This subsequent increase in forecasting error was less than 1.5 per cent, which implies that the variable not included added little in the way of predictive information. Furthermore, VIF analysis revealed that there is a strong multicollinearity between multiple attributes that were dropped and the VIFs of many of such attributes are high (78), hence, validating high redundancy in the original set of features. The resultant five feature subset (with constraints) also achieves an optimal balance between dimensional reduction and predictive accuracy. Table 3 summarises the features of the selected ones.

Table 3.

List of Selected features from Historical Sequential Data.

S. No	Feature Name
1	Three Phase Discharged Power
2	Battery Discharged Power
3	Time of Day
4	Solar Voltage
5	Temperature

After conducting the implementation of PDO, the top five characteristics, three-phase discharged power, battery discharged power, time of day, solar voltage, and temperature were chosen due to a consistently high-contribution score. In order to determine the possibility of information loss, that undertook a sensitivity analysis that introduced the omitted features (humidity, pressure, power factor, wind speed, grid voltage, and grid current) into the RBAGCN model progressively. The total forecasting error was not more than 1.5%, which validates the fact that the factors that were left out were not much redundant, but at best were loosely related to the target load. Also, a variance inflation factor (VIF) test showed that there was a high multicollinearity (VIF 78) among some of the dropped attributes, and that they had overlapping information content. Therefore, the constrained set of features including five variables not only minimized the dimensional complexity but also enhanced the generalization of the models and minimized the chances of overfitting.

Prediction using relational bi-level aggregation graph convolutional network

In this section, prediction using RBAGCN is discussed. It is a graph-based deep learning model designed to perform accurate short-term electricity load forecasting in SBMGs, enabling more efficient DM and enhanced battery performance²⁰. RBAGCN represents input data as a graph, where nodes correspond to selected features (e.g., Three Phase Discharged Power, Battery Discharged Power, Time of Day, Solar Voltage, and Temperature). Edges represent the underlying relationships and correlations between these features. The model captures both spatial relationships (across different features) and temporal dynamics (across time) using graph convolutional operations. The bi-level aggregation mechanism in RBAGCN enhances the learning process by aggregating information at two relational levels. First, it considers local feature interactions within immediate neighbours of each node. Second, it captures higher-level global relationships by further aggregating from broader graph structures. This two-tier aggregation improves model’s ability to extract meaningful patterns that influence load behavior, such as the interplay between solar input and battery discharge across different times of day. Though RBAGCN primarily aggregates relational data of spatially correlated features, it also has support of temporal dynamics, creating node representations that are time indexed. All node feature vectors correspond to time steps, and temporal adjacency is added with autocorrelation weights between successive intervals (t, t + 1, t + 2). Those connections enable the graph-convolution layers to learn both spatial feature interactions, as well as, time transition patterns at the same time. A gated recurrent embedding is added after the global aggregation layer, a sequence of spatial embeddings is merged with a temporal context and the regression output is obtained. Consequently, RBAGCN is a spatio-temporal predictor, which may be used to trace periodic changes in load and an abrupt shift in energy-usage in SBMG dataset²¹. RBAGCN is trained and validated using techniques such as K-fold cross-validation and hyperparameters optimization, ensuring high generalization capability and avoiding over fitting. By effectively modeling complex relationships in the data, RBAGCN delivers precise load forecasts that support EM, reduce peak demand, and contribute to battery longevity in SBMGs environments. The proposed RBAGCN includes four layers which are organized in a dual-aggregation hierarchy.

Layer 1–2 (Local Aggregation):

The feature of each node is then subject to a graph convolution that has ReLU activation with 64 as the hidden size. The adjacency matrix A is an adjacency matrix that represents the correlation between PDO-selected variables. Local interactions are obtained by means of the normalized mean of the neighbors.

Layer 3–4 (Global Aggregation):

The computation of softmax weighted sum of node embeddings by all local clusters is aggregated by a global attention mechanism. This allows the model to indicate important dependencies, such as the dependency between solar voltage and discharge power, in the whole graph.

The last, 32 unit, linear activation layer transforms the aggregated embeddings into the load prediction. In L2 regularization prevent overfitting in a dropout of 0.2. The Adam optimizer (learning rate 0.001, batch size of 32, 300 epochs) and mean squared error loss are used to perform training. This architecture is able to learn both local fine-grained features and global patterns thus ensuring temporal and spatial correlation of Smart Building Microgrid data is effectively. The Residual Graph Attention Convolutional Neural Network (RBAGCN) is used in the current research as a spatio-temporal regression approach to short-term electricity load prediction in Smart Building Microgrids. A Genetic Gradient Optimizer (GGO) is used to optimize the weight parameters of the RBAGCN to increase the stability of convergence and the ability of generalization to non-stationary loads conditions.

Optimization using greylag goose optimization

GGO controls the parameters of RBAGCN. GGO is an algorithm based on nature; it replicates the controlled and coordinated migration of Greylag geese in a V-formation. This is the action that motivated the maximization of the RBAGCN parameters that enhanced the accuracy of the short-term electricity load forecast of Smart Building Microgrids (SBMGs). GGO is highly dimensional, which is why it is highly suitable in optimization and tuning of deep learning models to their best performance. The important parameters of RBAGCN, including learning rates, weight coefficients, aggregation factors, etc. are tuned during GGO optimization. Such modifications enable the model to respond to the alterations in the energy requirements and surrounding circumstances in the microgrid. In the implementation of GGO, a goose is a solution candidate. The geese are guided by the most promising leader, but also they search in new directions when this may result in a better overall solution. It is an exploratory behavior that increases the search space and minimizes the chances of trapping in a local optimum. GGO tests a wide range of parameter settings and at each iteration, the best is taken according to the fitness of predictive performance. GGO beats the constraints of more traditional optimization techniques by applying an adaptative and cooperative search. It will increase the ability of RBAGCN to adjust to the dynamic patterns of loads and renewable energy. GGO optimizer is applied in this work to optimize the small, fixed-dimensional set of RBAGCN hyper-parameters: convolutional weight matrices W1-W4, global aggregation coefficient α, as well as the learning rate η. Each goose in the population represents a parameter vector

1

Unlike traditional hyperparameter tuning which does not change the hyperparameters dynamically as the training progresses, GGO adjusts dynamically the hyperparameters of the training processes to maximize the convergence robustness in the face of non-stationary loads changes.

The optimization goal is to minimize a composite fitness function F, defined as

2

where β₁ = 0.6, β₂ = 0.3, and β₃ = 0.1 weight accuracy and error contributions. Such weighting gives a tradeoff between accuracy (through R² ) and robustness (through MSE and MAE). In every iteration, GGO optimizes the candidate solutions based on the evaluation of the fitness function F on all the parameter vectors, and then, it retains the configuration that provides the smallest forecasting error. The optimal solution (minimum F) takes the place of the existing RBAGCN parameter set and an adaptively tuned model is obtained that converges more quickly and is more likely to generalize to the different load conditions in the SBMG dataset. With the use of the capabilities of GGO, the total accuracy, efficiency, and reliability of the load forecasting and decision-making system are enhanced. The GGO bio-inspired movement equations are not described in detail, since the algorithm is only applied as a convergence-stabilizing optimizer, and is not theoretical work. Flowchart of GGO for optimizing RBAGCN parameters is shown in Fig. 2(a).

Fig. 2.

(a) Flowchart of GGO. (b) Flowchart of Proposed GGO–RBAGCN Methodology.

Simulation environment and implementation

All simulations of the proposed GGO 0 -RBAGCN framework were implemented in MATLAB R2022b (64-bit) in a system with Intel® Core™ i7 processor, 16 GB of memory and Windows 11. Deep Learning Toolbox was utilized to implement the RBAGCN block, whereas the Global Optimization Toolbox to implement the GGO block. The steps of the workflow were: data acquisition, FRIF preprocessing, PDO features, RBAGCN prediction and GGO optimization. This MATLAB-based simulation environment schematic is given in Fig. 2(b), and ensures the results of the experiment can be replicated.

Result with discussion

The research findings of proposed methodology are covered in this section. The MATLAB platform was used to implement proposed GGO-RBAGCN technique. The obtained outcome of proposed GGO-RBAGCN approach was analysed with existing methods like RNN, ABMO-ANN, RNN-LSTM, GA-DNN, and RNN-GRU respectively.

Performance measures

The dataset would consist of about 1,05,000 samples of the time-indexed variations, whereby the samples are taken with a constant sampling rate of five minutes as other current methodologies in the short-term load forecasting and smart building energy management have been reported in other parts of the world. A 5-min time resolution forms a good compromise between capturing the fast changes in demand, the nature of battery charge–discharge cycles, and intermittency of the sun yet still maintaining computational tractability of deep-learning-based forecasting models. The dataset in question contains approximately 365 days of uninterrupted work of a Smart Building Microgrid at the given sampling resolution to include the fluctuation of demand on a daily, weekly, and a seasonal basis. The source of all measurements was provided by a realistic Smart Building Microgrid implemented in realistic residential, and commercial loads, along with the time-varying demand changes according to occupancy and daily battery cycling processes, and solar intermittency. The percentage of missing samples in the raw data set was not more than 2% and was mainly due to temporary sensor communication failures; which were later filled in the preprocessing stage by the interpolation of the neighboring time-step values. The data set was not shuffled but instead divided in sequence in order to maintain temporal causality and avoid data leakage. The samples used were divided into seventy percent training set, fifteen percent validation set and the remaining fifteen percent test set. The error measures (MAE, MAPE, and MSE) were evaluated on the normalized load values obtained through a min max scaling process and in this respect, the error measures represent relative errors that are scale free and are not actual deviations measured in physical quantities (kW).

MAE

The average absolute difference between actual and forecasted load values for each data point is measured by MAE. With a smaller MAE signifying more accurate predictions, it displays the model’s average prediction error in the same units as energy load.

3

where, denotes actual value of load at time denotes predicted value of load at time presents total number of data points.

MAPE

MAPE represents the prediction error as a %age of actual values; it is scale-independent and appropriate for cross-dataset comparisons across load scales. Prediction accuracy is higher when the MAPE is lower.

4

where, presents actual value of load at time denotes predicted value of load at time presents total number of data points.

MSE

Larger errors are severely penalized by MSE, which computes average of squared deviations between actual and projected values. It offers a quadratic indicator of prediction accuracy; higher performance is shown by reduced MSE values.

5

where, presents actual value of load at time denotes predicted value of load. presents total number of data points.

Accuracy

Error metrics like MAE, MAPE, and MSE are commonly used to evaluate the accuracy of load forecasting models; lower values denote better performance. However, these metrics, particularly MAPE, which has a direct correlation with prediction accuracy in percentage, can be used to develop a common way to describe accuracy as a percentage.

6

From Equ.6, the accuracy represents how close the predictions are to the actual values, expressed as a %. A higher accuracy (closer to ) indicates a more precise model. The accuracy numbers shown in this paper are the mean of ten fold cross-validation. The training–testing partitions were the same in all the baseline models. To every run research obtained Accuracy, MAE, MAPE, and MSE; the results averaged to an end result Accuracy of 98.3% error of 0.42. One-way ANOVA and HSD by Tukey showed that the value is significantly greater than the best available model, RNN-LSTM (91.7%) with p value less than 0.05. Therefore, the accuracy reported is a strong and statistically proven mean performance, but not a one-time performance. The forecast accuracy in this paper is [Accuracy = 100-MAPE], thus measuring the percentage based closeness of the predictions to the normalised set of data, no longer indicating the matter of absolute power-estimation accuracy.

Performance analysis

From Figs. 3, 4, 5, 6, 7, 8, 9, 10 and 11, the average with the standard deviation of the important indices, i.e., Accuracy, MAE, MAPE, and MSE, are mentioned based on ten separate experiments. The statistical significance of the results was proved by one-way ANOVA and Tukey HSD post-hoc test at the 95% confidence interval. The GGO-RBAGCN model attains substantial improvement in comparison to the optimal existing baseline (RNN-LSTM): Accuracy 98.3% ± 0.42%, MAE 0.0164 ± 0.0011 and MSE 0.0069 ± 0.0008 (p less than 0.05). These quantitative and statistical tools complement the visual observations and provide the support of the evidence of the superiority of the model. The Figs. 3, 4, 5, 6, 7, 8, 9 and 10 outline the operating scenario of the SBMG illustrating the variability of loads, renewable intermittency and battery dynamics, which all characterize the nonlinear scenario under which performance forecasting is assessed.

Fig. 3.

Analysis of SBMG’s Electric Load.

Fig. 4.

PV output power Analysis.

Fig. 5.

Wind Output Power Analysis.

Fig. 6.

Analysis of Overall Load demand.

Fig. 7.

Analysis of VRES Total Power.

Fig. 8.

Analysis of EV Battery Activity at the Building on a Typical Day.

Fig. 9.

Analysis of Irregular EV Battery Activity at the Building on a Typical Day.

Fig. 10.

Analysis of Battery Capacity Degradation.

Fig. 11.

Analysis of Accuracy.

Figure 3 represents the electric load profile of SBMG. The load begins at approximately 1450 kW at midnight and remains the same until approximately 6 a.m. After this, the demand decreases down abruptly to an average of 520 kW close to 7 a.m. Consumption is independent on the day with a short-lived peak of consumption at 2500 kW at 10 a.m. and at 2600 kW at midday and again falls to below 800 kW after 8 p.m. Such trend is a demonstration of the average daily cycle of a smart building: minimal usage at night, increase during the working day, and decrease in the evening. This kind of fluctuations is essential in proper load forecasting models and to make efficient plans to discharge batteries²².

Figure 4 presents the analysis of PV output power. The PV output is a definite bell like curve, which starts generating power beyond 6 h. The power peaks at approximately 960 kW at around 11 h–12 h around the time of the peak of the solar irradiance. The production remains at this high level over a few hours after which it gradually drops to zero at approximately 6 h. This daily predictability makes solar energy a good option of providing EM and charging of batteries during the daytime. The proper prediction of this trend contributes to the enhanced integration of renewables and decreases the use of grid power²³.

The analysis of the wind output power is presented in Fig. 5. The production of wind power is very erratic during the day. The early morning values range between 500 and 1500 kw and some cases of near zero generation are between 6 and 12 h. The production is not steady in the afternoon, as it is higher and lower now and then. Nevertheless, at about 5 -h the power slowly increases to a peak of approximately 1900 h of almost 11 h. It is this unpredictability of wind power that leads to the relevance of checking and adaptable demand-side administration approaches in the aim of maintaining a steady and balanced supply of energy, particularly to those systems that store energy in batteries to counterbalance renewable fluctuations²⁴.

The overall load demand analysis is depicted in Fig. 6. The peak load is at 12 0–15 0, and it is approximately 2600 0W. Its minimum is approximately at 7 h and is approximately 520 kW. Once optimized, the load is always maintained at a range of 1100–1250 kW. This is considerably within the range of 1200 kW minimum and 2200 kW maximum. The optimized demand reduces the load variation and ensures the load remains in an efficient range, which enhances the predictability and minimizes energy storage stress to facilitate a sustainable energy balance in the SBMG²⁵.

The total power presented by variable renewable energy sources (VRES) is represented at a time in Fig. 7. The curve of generation is extremely changeable. The peak power is about 1750 kW at 11 h. Other significant peaks of approximately 5 h and 10 h each are over 1250 kW. The minimum values, about 6 h and 18 h, are down to about 250 kW. Load matching VRES availability in semi-buried microgrids (SBMGs) minimizes waste, increases the proportion of clean-energy, and allows the operation even without the main grid²⁶.

Figure 8 depicts the EV battery activity at the building on a day in a typical day. The largest discharge event is at 10 h and its power output is almost 2700 kW. At 11 h, there is another discharge of 2000 kW. Charging becomes optimum at 16 0 h with input reaching about 1700 kW at this point. The charging and discharging actions are not exceeding 500 kW during the off peak (early morning 2 h 6 h) and bad weather (20 h-24 h). This regulated action expresses an effective battery-management plan. Storage and release of energy is planned to create balance between load and availability of renewable energy and will ultimately improve the performance of the smart building integrated battery system²⁷.

The analysis of irregular EV battery activity at the building on a regular day is presented in Fig. 9. The release rate peaks up to approximately 310 kW at 17 h and once more to approximately 230 kW at 22 h. Conversely, the charging remains relatively low of less than 70 kW. This discrepancy implies that the peaks in discharge are not accompanied by the appropriate periods of charging resulting in inefficiencies in power flow and stress in the battery of SBMG. Proper forecasting of loads and synchronized charge–discharge management are needed to even out these swings and maintain the operations at the same level²⁸.

Figure 10 indicates the decline in battery capacity during the course of time. The battery can retain only 80% of its initial capacity after approximately 1,300 days without any improvements. Under our scheme, the capacity reduces at a slower rate retaining 80% of the capacity beyond 2,900 days. This exceeds the actual life of the battery by more than twice the battery life and testifies to the advantages of optimized scheduling, reduced depth of discharge and controlled energy cycling. By making the energy storage have a longer life enhance the reliability and economic feasibility of the demand-side management in Smart Building Microgrids (SBMGs).

An analysis of the accuracy is shown in Fig. 11. The forecasting methods were compared and it is achieved that the proposed method will reach the best accuracy of 98.3% which is much better compared to the other methods compared. The lowest accuracy is registered by RNN at 80.1, followed by ABMO-ANN at 83.4, RNN-LSTM at 91.7, GA-DNN at 89.4 and RNN-GRU at 90.6. These findings indicate the weakness of conventional and hybrid models to explain variations in the patterns of power-use and renewable-generation. This approach is more accurate since the structured-forecasting framework increases scheduling, battery wear and DSM performance in SBMGs eventually resulting in more stability and energy-allocation efficiency of the smart buildings. Even though explicit root-mean-square error (RMSE) or loss convergence curves are not provided, the learning procedure stability is supported by consistently measured error measures obtained in training, validation and cross-validation partitions, thus suggesting strong convergence properties.

Table 4 presents a comparative analysis of various forecasting models using three performance metrics: MAE, MAPE, and MSE. The reduction in prediction variance and an overall mistaking improvement in all the measures of errors are an indirect indication of the effectiveness of GGO in stabilizing the convergence of the parameter when the object is being trained.. The GGO-RBAGCN model demonstrates the highest accuracy with the lowest values across all metrics MAE of 0.0164, MAPE of 0.0128, and MSE of 0.0069. The RNN method yields MAE of 0.0180, MAPE of 0.0341, and MSE of 0.0093, showing relatively higher errors. The dynamics of convergence were also stable in the training process, as the validation error was much similar to the training error during the epochs, thus indicating the lack of overfitting. ABMO-ANN records MAE of 0.0193, MAPE of 0.0246, and MSE of 0.0077, while RNN-LSTM gives MAE of 0.0196, MAPE of 0.0211, and MSE of 0.0082. The GA-DNN model scores MAE of 0.0210, MAPE of 0.0194, and MSE of 0.0074, and RNN-GRU results in MAE of 0.0216, MAPE of 0.0154, and MSE of 0.0071. These results confirm the robustness of proposed model in delivering minimal prediction error, making it most effective for load forecasting in SBMG environments. The chosen baseline models (RNN, LSTM, GRU, GA-, DNN, and ABMO-, ANN) are standard and popular benchmarks in the literature of short-term load prediction and the smart-building energy management. Even though the performance of state-of-the-art graph-based and attention-based models like STGCN, T -GCN, GAT, and Transformer models has been shown to be effective in large-scale grid scale forecasting, they either require high-quality spatial topologies and/or a large amount of computational resources to be deployed. Conversely, the suggested RBAGCN also uses a small feature-based relational graph, which makes it more suitable in real-time SBMG applications that have restrictions given by the data dimensionality. To reduce overfitting, sequential data partitioning strategy was used without shuffling and the generalization ability of the model was tested using ten-fold cross-validation, dropout regularization (the rate was 0.2) and L2 weight regularization in RBAGCN framework. To make a fair and repeatable comparison, all baseline models were created using the MATLAB R2022b. All of the models were trained on the same pre-processed data and 70:30 training–testing split. The RNN model had two hidden layers that had 64 neurons each and was optimized using Adam at a learning rate of 0.001. Its ABMO-ANN contained one hidden layer of 100 neurons, a weight tuning based on ABMO, and a sigmoid activation. The RNN-LSTM model had two 128-cell LSTM units, a dropout of 0.2, and a tanh activation. The GA-DNN had three dense layers with 128, 64, and 32. Its weights were solved using a Genetic Algorithm using crossover probability of 0.8 with a mutation rate of 0.05. The architecture was the RNN-GRU, with two layers that contained 128 GRUs each and had ReLU activation. Each of the models was trained in 300 epochs and a batch of 32 with the loss function being the mean-square error (MSE). Exactly the same conditions were used in the assessment of the proposed GGO-RBAGCN model and this guarantees that performance of all methods is relatively similar, thus avoiding any bias in the evaluation.

Table 4.

The comparison of MAE, MAPE, and MSE of proposed and existing methodologies.

Methods	MAE	MAPE	MSE
GGO-RBAGCN (Proposed)	0.0164	0.0128	0.0069
RNN	0.0180	0.0341	0.0093
ABMO-ANN	0.0193	0.0246	0.0077
RNN-LSTM	0.0196	0.0211	0.0082
GA-DNN	0.0210	0.0194	0.0074
RNN-GRU	0.0216	0.0154	0.0071

To test the strength of the proposed GGO-RBAGCN model, a tenfold cross-validation process was also done in a systematic way to the training part of the dataset. The consistency of performance across tenfold cross-validation further confirms stable validation behavior without significant performance degradation The smart-building microgrid load and generation profile dataset had been divided into 70% training and 30% testing, which all the benchmark models had the same share to be fair. The framework had a maximum validation of 97.8 to 98.4% with a total average of 98.3%. This good performance is due to a combination of Greylag Goose Optimization to hyperparameter optimization and better capability of the RBAGCN structure to learn both time–space features, which gives the highest predictive accuracy and the less error variance compared to current models.

Table 5 presents the accuracy percentages of various methods applied to a specific task, with the GGO-RBAGCN method outperforming all other techniques, achieving an accuracy of 98.3%. In comparison, traditional methods such as RNN, ABMO-ANN, and RNN-LSTM show lower performance, with accuracy values of 79.6%, 83.4%, and 91.7%, respectively. The GA-DNN and RNN-GRU methods achieve slightly better results than RNN and ABMO-ANN, with accuracy of 89.4% and 90.6%, respectively. The proposed GGO-RBAGCN method demonstrates significantly superior accuracy, indicating its effectiveness over other machine learning models for the given application.

Table 5.

Comparison of Accuracy with proposed and existing methods.

Methods	Accuracy (%)
GGO-RBAGCN (Proposed)	98.3
RNN	80.1
ABMO-ANN	83.4
RNN-LSTM	91.7
GA-DNN	89.4
RNN-GRU	90.6

Ablation study of FRIF, PDO, GGO, and RBAGCN components

The analysis included a component-level ablation study to compare the value each module contributes to the proposed forecasting model. This analysis was conducted by sequentially eliminating a component of the pipeline and keeping the other parts intact. Identical datasets, a sequential data partitioning scheme, and a tenfold cross-validation protocol were used to evaluate all the variants of ablation in order to ensure that the experiments were consistent. Table 6 shows the quantitative results of the current paper, and Fig. 12 shows the performance difference between the various configurations. Without FRIF preprocessing, the prediction errors increased, which is evidence of the noise and pure temporal fluctuations that are not removed by the FRIF filters on the raw sensor signals. The omission of PDO-based feature selection added redundancy to the input variables and in turn, raised the error metrics and reduced the generalization performance. The elimination of GGO optimization reduced rates of convergence and magnified the error of the forecast, which represents the learning process sensitivity to constant hyperparameter values in non-stationary load regimes. Replacing the RBAGCN predictor with a strictly sequential backbone led to the most significant performance loss, which supports the lower ability of sequence models to learn the inter-variable networks between energy-based features.

Table 6.

Component-Wise Ablation Results of the Proposed Framework.

Configuration	Accuracy (%)	MAE	MAPE	MSE
FRIF + PDO + RBAGCN + GGO (Full Model)	98.3	0.0164	0.0128	0.0069
w/o FRIF	96.9	0.0189	0.0167	0.0094
w/o PDO	97.1	0.0181	0.0152	0.0087
w/o GGO (Fixed Hyperparameters)	97.0	0.0184	0.0159	0.0090
w/o RBAGCN (LSTM Backbone)	91.7	0.0196	0.0211	0.0082

Fig. 12.

Forecasting accuracy under different ablation configurations.

Results and discussions

Extensive improvements in energy prediction, battery optimization, and system efficiency overall are demonstrated by the proposed model of the integrated load forecasting and decision-making in smart building microgrids (SBMGs). Electric load profile is also very dynamic with a minimum of 520 kW at 7 h up to 2600 kW between 12 and 15 h. It is hence necessary to accurately forecast. PV and wind production introduce even further complexity: the solar power reaches its maximum at 960kw towards 11–12am, whereas wind reaches its maximum at 1900kw irregularly. Once optimized, the load is maintained at between 1100 to 1250 kW which reduces the pressure on storage systems. The maximum discharge of the battery occurs at 2700 kW at 10 h and the maximum charge occurs at 1700 kW at 16 h in line with the renewable availability. Abnormal battery performance, including discharges as high as 310 kW, and charge as low as 70 kW, is evidence of inefficiency. Capacity-degradation analysis indicates that the approach doubles the battery life, which grows to more than 2900 days. GOGO-RBAGCN model has a 98.3% forecasting accuracy, which is better than RNN (80.1%), ABMO-ANN (83.4%), and LSTM (91.7%). It also has the lowest MAE (0.0164), MAPE (0.0128), and MSE (0.0069) and this confirms that it is stable and reliable in terms of demand scheduling and energy optimization. The causative reasons behind its superior performance are twofold: (i) the bi-level relational aggregation in RBAGCN is the ability to capture the local feature dependencies (e.g. short-term voltage-current correlations) and the global relationships (renewable generation and storage response); and (ii) the adaptive weight tuning of GGO is the ability to maintain an optimal exploration–exploitation balance in the course of the training. The combination of these mechanisms enables the model to deal with non-linear temporal changes to a better extent as compared to pure sequences such as RNN or LSTM, which lose memories of events over time. The FRIF + PDO preprocessing pipeline has an added advantage of only high impact and low redundancy features being used to make predictions, which further improves convergence and minimizes the effects of over-fitting. The study notes limitations. A one-site dataset was tested in only one climatic region, and to validate multi-regionality a multi-regional test should be done. The size of the population of GGO influences runtime, implying lightweight variants or distributed implementations of the code can be introduced. Expansions of the model to probabilistic forecasting and multi-objective optimization with balanced accuracy, computational cost and real-timeadaptability to future work will be achieved.

Analysis of resources and performance

Despite the fact that the GGO-RBAGCN framework is very precise, there is moderate level of computational overhead due to its multi stage structure. The implementation of the model was done in MATLAB R2023b on an Intel Core i9 (3.4 GHz) and 32 GB RAM. Training time per iteration was approximately 2.8 s on average and convergence was reached after 60–70 iterations. GGO has the largest contribution to runtime as it is based on population exploration yet it converges faster than other optimizers (PSO, GA). RBAGCN inference is fast, at latency of less than 0.8 s, and near-real-time load forecasting and scheduling is possible. Further enhancement of the performancFe is possible through parallel execution of GGO populations and decreasing the degree of aggregation. The strategies ensure that the framework remains computationally efficient, maintaining the accuracy needed to conduct real-time SBMG operations.

Practical implication and scalability

GGO-RBAGCN model can be used effectively in micro grids with large scale deployment. It works on a graph-based architecture and thus it can be scaled by hierarchical aggregation of nodes and thus it allows inclusion of more distributed energy resources, sensors, and load clusters without the need to train them extensively. The metaheuristic optimization at GGO is able to operate both in parallel and in distributed modes hence supporting edge and cloud-based networks of smart buildings. Parallel processing and reduced population strategies can be used to reduce the impact of training overhead. FRIF-PDO preprocessing pipeline is implementable on the IoT edge devices as on-site data conditioning so that the latency is minimal. Such characteristics render the hybrid framework to be flexible to large-scale smart microgrids and future renewable-storage-EV ecosystems based on Industry 4.0 standards.

Conclusion

A new hybrid method uses GGO and RBAGCN to address the process of integrated load forecasting and decision making in SBMGs. The model considers improper forecasting, inefficiency in storage, and battery degradation, which is an effective way to maximize energy consumption and increase battery life. It includes sophisticated preprocessing through FRIF and PDO based feature selection. The GGO-RBAGCN has an accuracy of 98.3% which is higher than the traditional models (RNN, ABMO-ANN, LSTM) and provides MAE of 0.0164, MAPE of 0.0128 and MSE of 0.0069. These findings affirm its efficiency in raising the forecasting accuracy and effectiveness of operations in SBMGs. Multi-stage framework will add more computational complexity by way of preprocessing FRIF and PDO selection, and optimization of GGO-based RBAGCN. The required length of training and resources in hardware are more than with conventional architectures. Whereas the accuracy improvement is worth the overhead in research contexts, in practice the performance of model compression, parallel execution, or smaller GGO populations will be required to run the models on edge devices in real time. This points to an instance of a trade-off between predictive accuracy and computational scalability, to be optimized in the future by research. The major weakness is that it might be specific to certain features of the input, implying that it cannot be applied to other types of buildings or climate zones. The future work will build on the proposed framework reflecting a broad range of intelligent building designs and operating conditions and incorporate adaptive learning capabilities to tackle the reality variability of the environment and dynamically changing user behaviour. Though the current research is founded on one real-life dataset of SBMG, future studies will compare the model on multi-building datasets and cross-regional datasets to determine its generalizability of application. Besides that, the next stage of research will involve direct visualization of the training convergence and optimization dynamics to better demonstrate the behavior and stability of learning.

Author contributions

B. Shamreen Ahamed: conceptualization, data curation, methodology, writing – original draft; Dhanya D.: software, validation, formal analysis; M. SivaramKrishnan: supervision, methodology, project administration, corresponding author; Arun Sriniva T.: investigation, resources, visualization; Ahmed Adnan Hadi: validation, critical review, editing; Shafat Khan: software support, technical validation; Wahaj Ahmad Khan: writing – review & editing, funding acquisition, correspondence.

Funding

The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Research Project under grant number RGP2/441/46.

Data availability

The data supporting the findings of this study can be accessed from the corresponding author upon reasonable request.                                                                                                                                                                                                                                                                                                                                                                             Acknowledgement: The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Research Project under grant number RGP2/441/46.

Declarations

Competing interests

The authors declare no competing interests.