Building-level energy prediction and control based on BIM and IIoT technologies

Abstract

This paper presents the NOECS framework based on Building Information Modeling (BIM) and Internet of Things (IoT) technologies for the collaborative optimization of building energy efficiency and thermal comfort. The framework integrates BIM spatial semantics with multi-source IoT sensor data and utilizes Neural Ordinary Differential Equations (ODE) and Transformer models to accurately predict temperature, humidity, and Predicted Mean Vote (PMV). Reinforcement learning is applied for HVAC control optimization. The model first binds sensors and devices effectively through BIM floor plans and room partitions, organizing state vectors by region to support subsequent visualization and operational decision-making. Experimental results show that NOECS achieves a Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) of approximately 1.0 °C for temperature prediction, 2.3% for humidity, and 2.7–2.9% for relative humidity. The MAE/RMSE for PMV remain below 0.2, demonstrating the model’s high accuracy and stability. In terms of energy consumption optimization, NOECS achieves energy savings of 92.3% and 90.7% on the BEC and BIM-BEM datasets, respectively, with energy utilization rates of 93.8% and 89.3%, significantly outperforming traditional baseline models like CNN-LSTM. Additionally, NOECS demonstrates superior computational performance: the training time on the BEC dataset is 24.5s with an inference time of 118.7 ms and a parameter count of 228.3 M, while on the BIM-BEM dataset, training takes 26.7s, inference is 124.4ms, and the parameter count is 232.7 M. Compared to DERN and CNN-LSTM, the computational cost is significantly reduced. Overall, NOECS achieves remarkable results in energy conservation and comfort assurance through efficient prediction and control optimization, while exhibiting excellent deployment feasibility, offering a replicable technical path for intelligent building energy optimization.

Introduction

Energy consumption and carbon emissions during the operation phase of buildings have been a central point in global energy and climate discussions. The International Energy Agency (IEA) and the Global Alliance for Buildings and Construction/United Nations Environment Programme (UNEP) have both highlighted the substantial contribution of buildings to global energy demand and associated emissions. Concurrently, the growth of operational energy consumption and the pressure to reduce emissions coexist, rendering the optimisation of building operations, represented by HVAC systems, a pivotal lever in achieving carbon neutrality and sustainability goals^1,2.

In real-world scenarios, enhancing the energy efficiency of buildings must address both the reduction of energy consumption and the necessity of meeting human thermal comfort requirements within indoor environments. In the event that control strategies prioritise energy savings without taking comfort into consideration, they frequently result in complaints, diminished productivity and rebound effects in secondary energy consumption. In terms of thermal comfort evaluation, the PMV/PPD model proposed by Homod³ laid the foundation for modern thermal comfort engineering calculations, which has been adopted by standard systems as an important method for evaluating moderate thermal environments. Standards like ISO 7730 and ASHRAE 55 have further provided acceptable ranges and evaluation processes based on PMV/PPD and combinations of environmental and individual factors, offering a basis for the quantification of comfort and its engineering applicability^4,5.

Therefore, intelligent building operation control for real-world scenarios is essentially a multi-objective optimization problem centered around the trade-off between energy consumption and comfort. The challenge lies in the fact that indoor thermal environments are influenced by the coupling of multiple factors such as external weather, envelope structure, occupant behavior, and equipment status, resulting in strong time variability and non-linearity. Furthermore, sensor data is prone to noise, missing values, and time delays, which exacerbates the uncertainty in prediction and control⁶.

In recent years, the integration of BIM and IoT has provided a new technological path to address these challenges. BIM can represent space, components, materials, and system topology in a semantic manner, providing a structured framework for cross-system data alignment and visual analysis⁷. IoT sensors collect operational data such as temperature, humidity, illumination, energy consumption, and equipment status at high frequency, facilitating the transition of buildings from static design to dynamic perception⁸. In this context, research has proposed IoT-BIM-based frameworks for thermal comfort monitoring and optimization. For example, Shi et al.⁹ introduced Dynamic PMV, which links BIM scenes with IoT data to optimize and visualize indoor thermal comfort, demonstrating the potential of integrating BIM semantics and real-time data for comfort control. Bassir et al.¹⁰ further combined BIM, IoT, and machine learning to construct thermal comfort models focused on user experience, highlighting the complementary value of data-driven methods to traditional rule-based strategies.

However, existing work still has several limitations: First, many methods separate prediction and control into offline modeling and online rule tuning, making it difficult to form an end-to-end closed loop, which results in poor adaptability to time-varying disturbances and seasonal transitions¹¹. Second, mainstream time-series models often rely on discrete-time step regression, which, while capable of fitting short-term trends, typically lacks the capacity to model continuous-state dynamics inherent in building energy efficiency, limiting the ability to characterize system inertia and lag¹². Third, the comfort-energy trade-off usually depends on manually set weights or empirical thresholds, making it difficult to achieve stable and generalizable optimal strategies under complex operating conditions¹³.

Deep learning and reinforcement learning provide the methodological foundation to address these shortcomings. In time-series prediction, Transformer¹⁴ enhances the modeling ability of long-range dependencies through self-attention, becoming an important paradigm for sequence modeling. However, pure discrete prediction might still overlook the continuous dynamic structure of building systems. The Neural ODE proposed by Worsham et al.¹⁵ parameterizes the state derivative with neural networks and completes forward evolution using an ODE solver, enabling the model to describe system dynamics in a continuous deep form and providing a unified framework for continuous-time energy efficiency modeling.

In decision optimization, reinforcement learning learns strategies through interaction with the environment, shifting from prediction to active control¹⁶. Notably, Proximal Policy Optimization (PPO) by Schulman et al.¹⁷ has been widely adopted for its simplicity and training stability, while Soft Actor-Critic (SAC), proposed by Jebrane et al.¹⁸, enhances exploration and robustness through a maximum entropy mechanism, providing effective solutions for continuous action space control. On the other hand, reinforcement learning research for building control has gradually formed a path from simulation to real-world applications. For instance, EnergyPlus, an open building energy consumption simulation engine, has long been used for energy consumption modeling and control strategy validation^18,19. Meanwhile, standardized training interfaces for building reinforcement learning are emerging, providing support for algorithm evaluation and reproducibility²⁰.

Nevertheless, integrating BIM semantics constraints, IoT multi-source perception, continuous dynamic prediction, and reinforcement learning control into an interpretable, reproducible, and deployable closed-loop system remains rare. Especially in research focused on PMV as the core comfort indicator and joint optimization of energy consumption and energy utilization efficiency, gaps still exist in areas such as loose model coupling, arbitrary reward design, insufficient cross-seasonal generalization, and inadequate engineering feasibility description.

Based on the above background and research gaps, this paper focuses on the collaborative optimization of building thermal comfort and energy efficiency by constructing a closed-loop intelligent decision-making framework that integrates BIM and IoT. Beyond simply combining Neural ODE²¹, Transformer, and reinforcement learning, the framework introduces a dual-path encoding mechanism with learnable gating to enhance feature representation and control precision²², while providing high-resolution predictions and fine-grained control for total building energy consumption, including HVAC, lighting, and major electrical loads. Indoor and outdoor multi-point sensor data (temperature, humidity, illumination, energy consumption, and related equipment information) serve as the input foundation, with BIM-provided spatial semantics and device associations integrated as data organization and decision constraints. The framework emphasizes closed-loop consistency from prediction to decision feedback, achieves low computational overhead suitable for edge deployment despite 228–232 million parameters, and maintains robust performance even under high sensor missing rates. In the experimental section, a systematic evaluation of energy-saving rates, energy utilization efficiency, and PMV prediction errors is conducted with comparative models and visual analysis, providing a rigorous and practically deployable evidence chain for building-level energy management.

NOECS framework: BIM–IoT-driven energy consumption prediction and thermal comfort control method

For the collaborative optimization of building thermal comfort and energy efficiency under BIM–IoT integration, this paper establishes a unified framework of multi-source data fusion—continuous dynamic modeling—long-term multi-step prediction—reinforcement learning closed-loop control, as shown in Fig. 1.

Fig. 1.

Overall flow diagram.

At time t, the system obtains observations from indoor and outdoor sensors, HVAC and energy consumption sides, and forms the state combined with BIM semantic information. The prediction layer uses Neural ODE to characterize the continuous-time evolution of energy efficiency and thermal environment states, and Transformer to model long-range dependencies across time scales, outputting predictions of energy consumption and comfort (PMV or sPMV) for the next H steps. The control layer injects prediction results as look-ahead features into the MDP state and uses reinforcement learning to learn the optimal HVAC strategy, enabling the system to reduce energy consumption and improve energy utilization efficiency while satisfying comfort constraints¹.

The core of the proposed method lies in coupling continuous dynamic consistency (ODE), long-range dependency representation (Transformer) and multi-objective decision optimization (RL) through unified state representation and closed-loop feedback, thereby enhancing robustness and deployability under the conditions of strong disturbances, strong nonlinearity and large hysteresis in real buildings.

System and data fusion

The input layer consists of three components: environmental perception (indoor and outdoor temperature, relative humidity, illumination, etc.), device and energy consumption (HVAC operational status and sub-component energy consumption/power), and BIM static semantics (room/function zone, spatial topology, orientation, envelope structure properties, device-area associations, etc.). Let the continuous variable set be denoted as , the discrete device state as or one-hot encoding, and the BIM semantic embedding as . The raw observations can then be written as:

1

Due to the inconsistencies in sampling frequency, missing values, and noise across multi-source data, this paper first unifies the resampling along the time axis, setting a consistent time step , resulting in an equidistant sequence . Continuous variables are aggregated using window-based aggregation or interpolation reconstruction, while discrete states are aligned using nearest-neighbor alignment or interval ratio encoding. To explicitly represent missing information, a missing mask is introduced, where indicates that the i continuous variable is missing at time t The observation with missing values is denoted as , and thus, the input features can be expanded as:

2

To avoid instability during training caused by differences in dimensionality, continuous variables are standardized using the training set statistics (,):

3

Additionally, energy-related variables can be scaled using logarithmic or quantile scaling to mitigate the impact of long-tailed peaks. Outlier handling is achieved by combining physical range clipping with local statistical detection, ensuring more stable feature distributions and reducing the risk of the model learning sensor noise patterns²³.

In terms of spatial modeling, this paper uses the region division and sensor placement in BIM to establish the mapping , and the mapping of HVAC operation ranges is incorporated into the state organization. This approach ensures that the state is not merely a stack of time-series data but includes structured expressions with regional semantics. Let there be k regions, with the observations for region k aggregated into , then the global state can be written as:

4

The aggregation can use derived features such as regional averages, maxima/minima, indoor-outdoor temperature difference , and humidity difference to enhance distinguishability and reduce noise sensitivity. At the end of this section, a floor plan with sensor placement is shown to illustrate which regions are observed, how indoor-outdoor coupling enters the state vector, and the correspondence between devices and regions, ensuring that the state definition is engineering-interpretable.

The output of this data fusion module is a unified state sequence used for both prediction and control. Its advantages lie in: explicit missing value modeling, which improves robustness to incomplete sensor data; BIM semantics enhancement, which makes the model more capable of learning spatial attribution and cross-region generalization patterns; and providing a consistent input interface for subsequent prediction and control tasks.

Neural ODE: continuous energy efficiency dynamics

The thermal environment and energy consumption processes in buildings exhibit typical continuous-time characteristics: indoor temperature and humidity respond to external disturbances and HVAC adjustments with thermal inertia and lag, and energy consumption does not jump at discrete time steps but rather continuously changes with the load, superimposed by local fluctuations caused by the on/off switching of equipment²⁴. If RNNs or convolutional structures are directly used to regress future values at discrete time steps^23,24, the model often needs to approximate this continuous dynamics through hidden state memory, which can lead to inconsistent predictions when the sampling interval changes, data is missing, or external conditions rapidly switch²⁵. Motivated by this, this paper introduces Neural ODE to model the continuous evolution of the energy efficiency and thermal environment system, allowing the model to learn the continuous change patterns of the state in the form of differential equations, thus achieving more consistent dynamic expressions over time.

Specifically, let represent the continuous-time hidden state (which can be understood as a low-dimensional representation of the indoor thermal and energy consumption states), and represent external drives and control inputs (such as outdoor temperature and humidity, illumination, equipment load surrogates, as well as HVAC set points and fan speeds). Neural ODE approximates the system dynamics using a learnable derivative network :

5

Given the initial hidden state at the starting point of the window, the ODE solver can integrate it forward to any time t, thereby obtaining a continuous hidden trajectory . To relate the hidden state to observable quantities, a decoding mapping is introduced to output the desired predictions (such as energy consumption, temperature, humidity, PMV, etc.):

6

In terms of implementation, is obtained from the fused state vector encoding, and consists of external disturbances and control inputs, ensuring that the ODE explicitly reflects changes in external inputs during the integration process. This approach allows the model to preserve the structure of continuous-time evolution while also providing predictions synchronized with the sensor readings at discrete sampling points.

As shown in Fig. 2, the core of ODENet is a shared-parameter derivative network and a numerical solver for integration. The connection between the ODENet block and the ODE solver demonstrates that the same set of parameters θ is repeatedly called throughout the time integration process, thus maintaining dynamic consistency across different time scales. Meanwhile, the ODE solver advances the hidden state from to outputting discrete samples of the continuous trajectory²⁶. This approach is more aligned with the properties of continuous physical systems, naturally expressing thermal inertia and lag responses. It also ensures that the model can still produce smooth and consistent state evolution through integration, even in the presence of irregular sampling or short-term missing data, thereby reducing the abrupt changes and drifts common in discrete models under missing data scenarios.

Fig. 2.

Neural ODE.

The input to this module is the BIM–IoT fused state sequence and external drives/control inputs. The modeling approach involves learning the evolution of the hidden state in continuous time and implementing time advancement via the solver. The output is a dynamically consistent hidden state, available for subsequent Transformer and control layers, along with the corresponding short-term prediction signals²⁷. Its main advantage lies in its ability to express continuous dynamic processes and its robustness to irregular sampling, providing stable dynamic priors for subsequent long-term predictions and closed-loop control.

Transformer: time-series prediction

Building operational data not only contains short-term continuous dynamics but also exhibits significant long-term dependencies and periodic structures, such as daily cycles (working hours/non-working hours), weekly cycles, and seasonal trends, while also superimposing disturbances from weather and occupancy behaviors. Relying solely on local dynamics often struggles to capture these long-range patterns. Therefore, this paper further employs the Transformer model for multi-step prediction, using it as the core structure to capture long-term dependencies and cross-variable coupling, complementing the Neural ODE hidden states²⁸.

Let the historical window length be T, and the state sequence be . The Transformer establishes dependencies across the entire window using self-attention, and its basic form can be written as:

7

Here, the attention mechanism allows the model to focus on the most relevant segments in the historical window when predicting future energy consumption or PMV, such as the influence of morning equipment startup on afternoon peaks or the delayed impact of outdoor humidity changes on indoor humidity and PMV. To inject the continuous dynamic information provided by the ODE into the attention network, this paper combines the ODE hidden state as an auxiliary channel with the raw state features before feeding them into the Transformer. This allows the model to learn long-term dependencies while benefiting from both continuous dynamics priors and data-driven associations²⁹.

At the output stage, the Transformer predicts a multi-variable sequence for the next H steps, covering energy consumption curves (hourly/daily), as well as temperature, humidity, and thermal comfort indicators (PMV or sPMV). Training uses multi-task joint regression, enabling the model to learn the coupling patterns of environment, energy consumption, and comfort within the same representation space. In terms of evaluation metrics, this paper uses MAE, RMSE, MAPE, and to quantitatively assess prediction performance. These metrics provide a comprehensive evaluation from the perspectives of absolute error, squared error more sensitive to peaks, relative error, and goodness-of-fit, and they are consistent with the metrics in subsequent comparative experimental figures.

Reinforcement learning control strategy

Building on the prediction module, which provides future energy consumption and comfort trends, this paper models HVAC optimization as a Markov Decision Process (MDP)³⁰ and uses reinforcement learning to learn the optimal control strategy under different disturbance conditions. Let the state space be , and the action space be . At time t, the agent observes state , outputs action , and the environment transitions to , returning a reward . This paper emphasizes the integrative and anticipatory nature of state construction: not only includes the current indoor and outdoor temperature, humidity, energy consumption, equipment status, and BIM semantic encoding but also incorporates the short-term trends predicted by the forecasting module (future H-step energy consumption/PMV predictions) as anticipatory features into the state. This allows the strategy to transition from reactive control to predictive, foresight-based control³¹.

The action corresponds to the actual controllable quantities of the HVAC system (such as set temperature, fan speed/valve opening, on/off switching, mode changes, etc.), with specific dimensions determined by the system implementation. The reward function is designed to express the trade-off between energy savings and comfort, combining the energy consumption term with the comfort term, while adding penalties for comfort violations and drastic changes in actions to enhance engineering applicability. A general and interpretable form is:

8

where represents energy consumption or power, represents the degree of deviation from neutrality, r is the comfort threshold (comfort zone), and the last term penalizes frequent adjustments and policy oscillations. The meaning of this design is that: the strategy must optimize energy consumption while controlling PMV deviation and minimizing comfort violations; at the same time, the strategy is encouraged to adjust HVAC smoothly, aligning better with equipment lifespan and real-world operational logic.To ensure consistency, all reward components (energy, PMV deviation, and action smoothness) are normalized to comparable scales. The weights for each term (w_E, w_PMV, w_Action) are chosen based on engineering priorities, prioritizing occupant comfort while maintaining smooth control. Sensitivity analysis confirms that the selected weights balance energy efficiency and thermal comfort without favoring extreme energy savings at the expense of comfort.

The policy learning uses the Actor–Critic framework, where the Actor generates actions (or action distributions), and the Critic evaluates the state/action value and guides the policy update³². As shown in Fig. 3, the framework consists of state input, Actor outputs actions, environmental feedback on energy consumption and comfort metrics, reward function calculation, and Critic value estimation and policy update, forming a typical closed-loop. Compared to traditional rule-based control, this approach can form adaptive strategies through interactive learning under complex disturbances. In contrast to control based solely on current observations, this paper incorporates prediction outputs into the state, allowing the strategy to anticipate future risks (such as peak energy consumption and comfort violations) and thus perform peak shaving and energy savings while maintaining comfort.

Fig. 3.

RL control agent.

Experimental deployment and evaluation system design based on NOECS

BIM–IoT platform deployment and data pipeline configuration

To ensure the deployability and reproducible evaluation of NOECS in real building operation scenarios, this paper first establishes a BIM–IoT data acquisition and processing pipeline for the operational stage, which enables the fusion of indoor and outdoor environmental observations, HVAC operating status, and energy consumption metrology on a unified timeline, and further converts them into state vectors directly usable by the model.

The system inputs mainly come from three types of channels: The first is temperature and relative humidity continuously collected by indoor and outdoor environmental sensors (extensible to illumination, etc.), used to characterize the building’s thermal-humidity boundary conditions and their time-varying disturbances. The second is operation information and control variables from the HVAC side, including set temperature, fan speed/valve opening, on/off status, and mode switching, to describe controllable inputs and equipment working conditions. The third is energy consumption data provided by electricity meters or sub-metering interfaces, used to reflect system load and energy-saving effects.

Due to differences in sampling periods and communication delays among various devices, this paper adopts a unified timestamp as the alignment benchmark, resamples multi-source data, and synchronizes them to a uniform sampling interval. This ensures that the state vector of each time slice is composed of environmental observations, equipment status, and energy consumption measurements from the same time step, avoiding spurious correlations and training biases caused by cross-time splicing.

The introduction of BIM semantic information elevates sensor observations from pure time series to state representations with spatial indexing³³. Sensors are associated with rooms or zones, and Table 1 summarizes the BIM features used, including room ID, adjacency, and sensor-device associations, with types and encoding methods. Discrete features are embedded via one-hot or learnable vectors, while adjacency is structure-aware. Combined with IoT data, these features form state vectors for each time slice, providing spatial context for the Neural ODE, Transformer, and RL modules.Based on BIM floor plans and room/functional zone divisions, this paper constructs the association between sensors and rooms or regions. Meanwhile, state vectors are organized according to adjacency relationships and equipment influence ranges, so that model inputs are structurally encoded by spatial partitions³⁴, supporting subsequent visual interpretation and connection with operation and maintenance decision-making. Table 2 summarizes the BEC and BIM-BEM datasets used in this study, including building type, climate, sensors, sampling rate, and train/validation/test splits, providing an overview of the data supporting model training and evaluation.

Table 1.

BIM features used for building-level state vector construction and their encoding methods.

Feature	Type	Encoding
Room ID	Discrete	Learnable embedding or one-hot
Room adjacency/spatial partition	Graph/structure-aware	Adjacency encoding
Sensor-device association	Discrete/graph	Embedded as part of state vector

Table 2.

Summary of BEC and BIM-BEM datasets methods.

Dataset	Building type	Location/Climate	Sensors/Variables	Sampling rate (min)	Train period	Val period	Test Period
BEC	Commercial Building	Temperate climate, City, Country	Indoor/Outdoor Temperature, Humidity, HVAC states, Energy meters	5	Jan 1–Mar 31, 2024	Apr 1–Apr 30, 2024	May 1–May 31, 2024
BIM-BEM	Multi-zone Office	Temperate climate, City, Country	Indoor/Outdoor Temperature, Humidity, HVAC states, Energy meters, BIM spatial partitions	5	Feb 1–Apr 30, 2024	May 1–May 31, 2024	Jun 1–Jun 30, 2024

The sensor layout and zone numbering are shown in Fig. 4. Multiple indoor points cover main functional areas, and outdoor observation points are set to characterize external disturbance inputs. Such paired indoor-outdoor observation configuration provides a foundation for indoor-outdoor coupling analysis and state modeling³⁵. Furthermore, one-month monitoring results of temperature and relative humidity are presented in Fig. 5. The outdoor curve shows larger fluctuation amplitude and sharper variations, reflecting strong disturbances from meteorological conditions. The indoor curve is relatively smooth under the thermal inertia of the building envelope and HVAC regulation, but still exhibits obvious time-varying fluctuations and non-stationarity.

Fig. 4.

Indoor sensor distribution.

Fig. 5.

Monthly temperature and relative humidity monitoring graph.

Such data pattern—strongly disturbed inputs and slowly responding outputs—determines that pure discrete modeling may become unstable in long-term prediction and missing-data scenarios. The design of continuous dynamic characterization and long-dependency prediction adopted in this paper is precisely based on the above data characteristics and engineering pipeline constraints.

Task definition, comparison baselines, and evaluation metrics

In terms of evaluation setup, this paper performs a unified validation of NOECS across two main tasks: multi-step prediction and closed-loop control. For the prediction task, we use a sliding window to construct supervised samples: given the multi-source state sequence for the past T steps as input, the model outputs the multi-variable prediction for the next H steps. The term is employed to denote the fused state at time t incorporating a multitude of parameters such as indoor and outdoor temperature and humidity, energy consumption, HVAC state, and BIM semantic embeddings, among others. The NOECS model is trained with a sliding window of 30 time steps and a prediction horizon of 10 steps. The AdamW optimizer is used with an initial learning rate of 8 × 10⁻⁴ and a weight decay of 0.1. Batch size is set to 64, and models are trained for up to 100 epochs with early stopping based on validation loss. Training and evaluation are conducted on a NVIDIA Jetson AGX Xavier with 16 GB LPDDR4 memory, ensuring reproducibility and demonstrating efficiency under embedded deployment conditions.

For the closed-loop control task, Proximal Policy Optimization (PPO) is employed to optimize building-level actions. The reward function combines energy savings, comfort deviation (PMV), and action smoothness, with each term normalized and weighted according to engineering priorities. Actions are constrained by HVAC operational limits, and comfort is constrained by a PMV threshold. Training occurs in a building-level closed-loop simulation environment incorporating multi-point indoor/outdoor sensor data, BIM spatial partitions, and HVAC device states, allowing evaluation of prediction and control under realistic operational conditions.

The prediction task can then be formalised as follows:

9

This includes both the future energy consumption sequence and thermal comfort-related variables (such as indoor temperature and humidity, PMV, or sPMV), supporting forward-looking assessments at the daily/hourly level. To ensure fairness in comparison, all baseline models and NOECS use the same set of input features and the same training/validation/testing time splits. Training is conducted under consistent training budgets (training epochs, early stopping strategy, or maximum number of iterations) to ensure that performance differences primarily reflect model structure and modeling capabilities³⁶.

The comparison baselines cover typical deep time-series forecasting frameworks and multi-task structures, including CNN-LSTM³⁷, pure Transformer, BERT-TS³⁸, DERN³⁹, MSCNN⁴⁰, and MTL-BET⁴¹. BERT-TS is a Transformer-based time series forecasting model leveraging bidirectional encoder representations for sequential prediction, DERN is a deep energy forecasting network using residual connections for multi-step prediction, and MTL-BET is a multi-task learning framework integrating energy and thermal prediction tasks.These baselines and NOECS are trained and tested on the same dataset and under the same evaluation protocol, with the same prediction step length H and input window T to avoid any non-comparability caused by differences in prediction difficulty.

In terms of evaluation metrics, prediction accuracy is measured using MAE, RMSE, MAPE, and , with statistics provided separately by variable and scenario (e.g., indoor temperature, outdoor humidity, energy consumption, PMV/sPMV, etc.), to prevent single summary statistics from masking differences. In consideration of the true values and the values predicted for a specific variable in the test set, , the metrics are defined as:

10

where ε is a small constant to avoid division by zero, and y is the sample mean. In this paper, the above metrics are reported separately for each variable, and can further be aggregated by indoor/outdoor, region, or time periods (peak/non-peak) to reflect the differences in prediction difficulty and error distribution.

The closed-loop control evaluation focuses on energy savings and the satisfaction of comfort constraints. The Energy Saving Rate (ESR) is used to measure the reduction in energy consumption relative to the baseline strategy. Let represent the baseline energy consumption (e.g., original strategy/control algorithm), and represent the control strategy energy consumption, then:

11

where E_baseline represents the total energy consumption under the baseline building-level control strategy, including HVAC, lighting, and major electrical loads, and E_NOECS represents the total energy under our proposed control framework. The baseline corresponds to conventional scheduled HVAC and lighting control in the BEC and BIM-BEM datasets, and the ESR is computed over the entire evaluation period, reflecting cumulative energy reduction.

The Energy Utilization Rate (EUR) describes the proportion of energy consumption that contributes to effective comfort or service output. It can be generally expressed as:

12

In this definition, the numerator counts the cumulative number of time steps where the indoor PMV falls within the comfort range [− 0.5, 0.5], representing effective comfort satisfaction. The denominator is the total building energy consumption, including HVAC, lighting, and major electrical loads. This formulation makes EUR interpretable as the ratio of delivered comfort per unit energy consumed. Different studies define EUR differently (e.g., using comfort duration, effective operation duration, or penalizing deviations), but here it is consistently calculated according to this definition and reported uniformly in the results section.The baseline corresponds to conventional scheduled control of HVAC and lighting. This ensures that the reported energy savings reflect full-building energy reduction rather than a subset.

Meanwhile, the satisfaction of comfort constraints is evaluated through the proportion of PMV falling within the comfort range or the extent of violation. For instance, comfort satisfaction rate is defined as:

13

where τ is the comfort threshold, and is the indicator function, used to measure whether the control strategy maintains comfort while saving energy.

The PMV comfort threshold is set to ± 0.5, corresponding to the range recommended by ASHRAE Standard 55 for acceptable indoor thermal comfort. Values of PMV outside this range incur proportional penalties in the reward function, guiding the RL agent to maintain occupant comfort while optimizing energy efficiency. This ensures that the reward design has an operational justification and aligns with established comfort standards.

The closed-loop control experiments are conducted in a building-level simulation environment that integrates multi-point indoor and outdoor sensor data, BIM spatial partitions, and HVAC device states. Counterfactual energy consumption under RL actions is computed within the simulator, which has been calibrated using historical building operation data to ensure realistic energy and comfort dynamics. This simulation-based evaluation allows safe testing of various control strategies and provides consistent and reproducible measurements for both NOECS and baseline models.

NOECS prediction-control collaborative optimization results and analysis

Load patterns and seasonal difference analysis

This section provides testable conclusions from three perspectives: daily load structure, cross-seasonal differences, and sub-component energy consumption composition. The goal is to clarify the key time periods and critical load pools that should be the focus of subsequent prediction evaluation and control optimization. First, the daily energy consumption variation on an hourly scale is shown in Fig. 6 (predicted daily energy consumption changes). Based on hourly sequences, the minimum value occurs at 23:00, with = 37.6 kWh/h; and the maximum value occurs at 11:00, with = 137.6 kWh/h, corresponding to a peak-to-valley difference of = 100.0 kWh/h kWh/h and a peak-to-valley ratio of = 3.7. This magnitude indicates a significant scheduling potential in the scene, with energy-saving potential concentrated primarily during the daytime high-load periods. Dividing the entire day into typical operational segments, the cumulative energy consumption is approximately 287.4 kWh (13.5% of the total) from 0:00 to 5:00 (nighttime), about 1436.6 kWh (67.5%) from 6:00 to 17:00 (main operating period), and about 405.4 kWh (19.0%) from 18:00 to 23:00 (decline phase). Therefore, the 6:00–17:00 period alone accounts for about two-thirds of the total energy consumption. For subsequent control strategies to generate substantial benefits, stable energy savings must be achieved during this period, rather than merely optimizing the nighttime baseline.

Fig. 6.

Energy consumption variation on a daily scale.

Further examination of the load change rate reveals distinct patterns of rapid ascent, high-level plateau, and rapid decline throughout the day. The maximum increase occurs between 10:00 and 11:00, with a single-hour increment of approximately + 26.2 kWh/h; the maximum decrease occurs between 20:00 and 21:00, with a single-hour decrease of about − 26.5 kWh/h. These two high-gradient periods are the most sensitive intervals for prediction errors and control actions: underestimating the load during the ascent phase can result in a delayed response from the control system, while overestimating the load during the decline phase may trigger unnecessary forced load shedding and induce oscillation in actions. Meanwhile, from 8:00 to 17:00, a sustained high-level plateau characteristic is observed, with a mean of about 125.5 kWh/h and a standard deviation of around 9.0 kWh/h. This indicates that daytime operations are not isolated peaks but long plateaus with small fluctuations. This necessitates control objectives to focus more on smooth optimization under constraints, rather than short-term peak shaving to achieve localized reductions in energy consumption.

Figure 7 compares electricity consumption between February and August. The August curve (green) exhibits higher magnitude and more frequent fluctuations, with several high-magnitude peaks occurring within short periods. External disturbances and cooling load drivers are more pronounced in this period, significantly increasing the variance in load. In contrast, lower levels in February (blue) show regular periodic peaks and phase shifts, with load following rhythm-driven and segmented adjustment patterns. This difference implies that the risk-reward structure of the same control strategy varies across seasons: the main challenge in summer is ensuring robustness and comfort constraint satisfaction during high peaks and strong fluctuations, while winter emphasizes efficiency improvements in alignment with the operational rhythm. It is important to note that the later part of the August curve is truncated. To perform a strict cross-seasonal statistical analysis, a unified valid time window should be used or explicit handling of missing data should be applied to avoid bias caused by differences in sample length.

Fig. 7.

Energy consumption in february and august.

Figure 9 shows the percentage distribution. Figure 8 shows the total annual energy consumption. It is approximately 689,958.8 kWh/year. The highest contributing components are: technical equipment (111898.7 kWh/year, 16.2%), domestic hot water (101097.0 kWh/year, 14.7%), lighting (92658.2 kWh/year, 13.4%), fans (89957.8 kWh/year, 13.0%) and room heating (83,881.9 kWh/year, 12.2%). These five components account for about 69.5% of the total energy consumption. HVAC-related components (heating, ventilation and heating, ventilation and cooling, cooling, fans, pumps) account for approximately 51.9% of total energy consumption. This demonstrates that HVAC control has sufficient leverage to alter the overall energy consumption. However, technical equipment, hot water, and lighting—non-HVAC dominated loads—account for nearly 48.1%, meaning that the nighttime baseline and energy-saving limits will be constrained by uncontrollable loads. Figure 9 further visually confirms this structure: technical equipment and hot water form the largest energy consumption block, followed by lighting and fans, with HVAC-related components occupying about half of the total consumption structure, providing a direct basis for subsequent control benefit attribution.

Fig. 9.

Annual energy consumption share by component.

Fig. 8.

Annual energy consumption by component.

NOECS multivariable prediction performance comparison and error source analysis

This section focuses on evaluating the performance of NOECS in temperature, humidity, and PMV prediction, particularly whether it is reliable enough to support subsequent control decisions. We perform a thorough analysis of the model’s prediction accuracy and its impact on control strategies through time-series consistency analysis, quantitative comparison, and visual presentation.

First, Fig. 10 shows the comparison between Prediction and Actuality on a monthly scale. The key aspect of this figure is to verify the model’s tracking ability under different operating conditions: it maintains stable alignment during steady phases, does not exhibit significant phase lag or systematic bias during abrupt transitions, and replicates the main fluctuation patterns even during high-frequency oscillations. In other words, NOECS not only fits the mean level but also captures the change process, which directly determines whether control strategies will suffer from action bias or comfort boundary risks due to state estimation errors during load increases, decreases, and disturbance enhancement periods.

Fig. 10.

Indoor indicator prediction vs. actual comparison.

Based on this, Fig. 11; Table 3 quantify the visual consistency from Fig. 10 into measurable indicators and baseline comparisons. In terms of absolute error, NOECS maintains a MAE/RMSE of approximately 1.0 for temperature, 2.3% and 2.7–2.9% for humidity, and keeps PMV’s MAE/RMSE stable at around 0.2. These results show two key points: first, the environmental variable prediction errors are within the range of engineering usability; second, and more importantly, the PMV is not significantly amplified by the compounded temperature and humidity errors, indicating that the model’s mapping from environmental variables to comfort indicators is stable. This allows the model to provide reliable inputs for the comfort constraints and reward terms in subsequent reinforcement learning.

Fig. 11.

Temperature/Humidity/PMV comparison with other models (bar chart).

Table 3.

Comparison of indoor/outdoor temperature/humidity/PMV.

	CNN-LSTM		NOECS
	Outside	Inside	Outside	Inside
MAE
Temperature(°C)	0.867	0.896	0.368	0.371
Humidity(%)	2.337	2.342	0.783	1.167
PMV	0.165	0.228	0.079	0.102
MAPE
Temperature(°C)	3.03	4.454	1.377	1.378
Humidity(%)	3.926	4.569	1.561	2.433
PMV	23.375	26.521	6.567	12.353
RMSE
Temperature(°C)	0.979	1.012	0.441	0.543
Humidity(%)	2.896	2.67	1.367	1.766
PMV	0.183	0.22	0.075	0.091
MAPE
Temperature(°C)	0.993	0.938	1.079	1.007
Humidity(%)	0.935	0.852	1.034	0.948
PMV	1.022	0.912	1.072	0.941

Regarding relative error, the MAPE for temperature and humidity remains in the low single-digit percentage range, demonstrating robustness in terms of amplitude scale. However, the higher MAPE for PMV mainly arises from the denominator effect when PMV is close to zero. Therefore, this paper uses MAE/RMSE as the primary indicator for the usability of comfort predictions, with MAPE serving as a secondary reference to avoid misjudgments due to the potential amplification of relative errors.

To evaluate the individual contribution of each module in the NOECS framework, we conducted a comprehensive ablation study with four experimental groups: Transformer-only, ODE-only, the complete ODE+Transformer model, and a No-RL baseline. Figure 12 compares the prediction accuracy across these groups, using temperature MAE, temperature RMSE, and PMV MAE. The results show that the complete model consistently achieves the lowest prediction errors, demonstrating that Neural ODE effectively captures continuous dynamics, Transformer models long-range dependencies, and RL contributes to predictive control. Figure 13 presents the closed-loop control performance of the same groups, quantified by ESR (Energy Saving Rate) and EUR (Energy Utilization Rate). The complete model achieves the highest ESR and EUR, while the No-RL baseline shows no closed-loop performance, confirming that RL is essential for effective control and that both ODE and Transformer modules enhance the quality of the predictive state representation for decision-making. Together, these results indicate that each module contributes meaningfully to the overall performance, and the combined architecture provides substantial gains in both prediction accuracy and closed-loop control efficiency rather than merely increasing computational complexity.

Fig. 12.

Ablation study: prediction accuracy of different models.

Fig. 13.

Ablation study: closed-loop control performance of different models.

Figure 14 illustrates NOECS prediction performance under varying sensor missing rates (10–50%) using the proposed mask mechanism. As the missing rate increases, temperature MAE, temperature RMSE, and PMV MAE show a slight upward trend, reflecting moderate performance degradation. Nevertheless, the overall errors remain low across all metrics, demonstrating that NOECS maintains robust prediction even under substantial sensor data gaps and effectively mitigates the impact of missing IoT inputs.Missing sensor values are handled using a mask-based approach, where the mask indicates missing entries to the model. The same mechanism is applied to all baseline models for consistent comparison.

Fig. 14.

Prediction performance degradation of NOECS under different sensor missing rates.

Finally, Fig. 15 maps the prediction results to operational decision-making scenarios, illustrating the interpretability and practical applicability of the model outputs. The selection of three typical time points (7:00, 14:00, and 19:00) enables the establishment of a correlation between the indoor temperature, average humidity, and PMV prediction results with the BIM visualisation interface. This creates a coherent information chain between numerical predictions, comfort status, improvement suggestions, and control actions. The significance of this display is that the model not only provides predictive accuracy in offline evaluation but also serves actual operational management through time-segmented and visualized methods. This provides intuitive support for the execution of subsequent closed-loop control strategies and human review.

Fig. 15.

Indoor indicator prediction for each time period of the day.

Control benefits, operational efficiency, and deployment feasibility evaluation

This section verifies the closed-loop value of NOECS from three perspectives: the stability of control benefits, the sources of the benefits, and the ease of system deployment. Firstly, with regard to the benefits of control, the two indicators (Energy Saving Rate/Energy Utilization Rate) provided in Fig. 16; Table 4 demonstrate that NOECS achieves the optimal or tied-best performance across both datasets. For instance, on the BEC dataset, NOECS attains an ESR of 92.3% and an EUR of 93.8%, which are 0.4 and 1.2% points higher than the top-performing CNN-LSTM (91.9%, 92.6%), respectively. This finding suggests that the observed gains cannot be attributed to random fluctuations resulting from changing models or metrics. Instead, it is more likely that these gains represent a substantial enhancement in efficiency under consistent comparison conditions. In the BIM-BEM dataset, NOECS continues to demonstrate its superiority, with an ESR of 90.7%, which is higher than the 90.3% achieved by CNN-LSTM. Additionally, the EUR of 89.3% surpasses that of other methods, such as CNN-LSTM with 89.0%, further substantiating the stability of the framework’s benefits across various scenarios and data distributions, and not merely on a specific dataset.

Fig. 16.

Comparison of energy utilization efficiency and energy saving rate across models. (a) BEC Dataset (b) BIM-BEM Dataset

Table 4.

Comparison of energy utilization efficiency and energy saving rate across models.

	DERN	MSCNN	BERT-TS	Transformer	MTL-BET	CNN-LSTM	Ours
BEC Dataset
Energy saving rate	86.693	87.014	88.270	88.804	91.156	91.931	92.305
Energy utilization rate	85.812	86.052	88.778	88.056	90.327	92.625	93.801
BIM-BEM Dataset
Energy saving rate	83.992	84.310	85.690	86.432	88.130	90.305	90.703
Energy utilization rate	84.204	85.000	85.955	86.273	88.077	88.979	89.324

Energy consumption attribution can be directly linked to controllable load pools and the maximum benefit achievable. As shown in Figs. 8 and 9, annual energy consumption is not evenly distributed but concentrated in a few components. This provides a clear conclusion: the control strategy theoretically affects at most about half of the total energy consumption. Therefore, the potential for energy saving improvements primarily comes from enhancing the efficiency of HVAC-related components, rather than expecting significant reductions in less controllable loads like technical equipment or hot water. At the same time, fans (13.0%) and heating (12.2%) are high-impact components in the controllable load pool, determining the primary sources and priorities for control optimization benefits.

Deployment feasibility is also demonstrated through evidence of lower costs for the same benefits. Figure 17 shows that NOECS achieves the lowest training, inference, and parameter costs across both datasets: 24.5s training time, 118.7ms inference time, and 228.3 M parameters on the BEC dataset; 26.7s training time, 124.4ms inference time, and 232.7 M parameters on the BIM-BEM dataset. Compared to CNN-LSTM, NOECS reduces training time by about 14.6–14.7%, inference time by 1.9–2.1%, and parameter count by 4.9–5.0% across both datasets. Compared to DERN, training time is reduced by 42.7–47.1%, inference time by 15.6–16.9%, and parameter count by 21.2–22.8%. Combining this with NOECS’s lead in energy savings and energy utilization rate (Figs. 4, 5, 6, 7 and 8), we can draw a stronger conclusion: NOECS achieves its performance gains through lower inference latency and smaller parameter sizes. Its primary bottleneck for online deployment lies not in model computation but in data collection and control execution. This also means that embedding NOECS into a BIM visualization platform for near-real-time inference is feasible from an engineering standpoint and has the resource margin to expand further into multi-zone/multi-floor scenarios.Table 5 further compares NOECS with parameter-matched baseline models. Although the total number of parameters is similar, the complete NOECS model consistently outperforms models lacking one or more modules (Transformer, ODE, or RL), indicating that its performance benefits stem from the architectural design rather than parameter count alone.

Fig. 17.

Performance comparison across different models on datasets.

Table 5.

Comparison of parameter-matched baselines and NOECS performance.

Model	Parameters (M)	Modules included	ESR (%)	EUR (%)
Transformer-only (scaled to 230 M)	230	Transformer	90.2	91.5
ODE+Transformer (scaled to 230 M)	230	Transformer + ODE	91	92
NOECS (ODE+Transformer + RL)	230	Transformer + ODE + RL	92.3	93.8

While Fig. 15 demonstrates relatively short training and inference times, the NOECS model contains approximately 230 million parameters, which may pose challenges for typical edge nodes. On an NVIDIA Jetson AGX Xavier (16 GB LPDDR4 memory), inference for a single sample requires approximately 0.7–0.8 GB of GPU memory and 120–150 ms, indicating that near-real-time deployment is feasible for modern embedded devices. For more resource-constrained environments, further optimizations such as FP16/INT8 quantization, input tiling, or reduced batch sizes can reduce memory footprint and latency while maintaining predictive and control performance.

Conclusion

This paper focuses on the three main aspects of “Building Energy Efficiency Prediction–Comfort Control–Collaborative Optimization of Energy Savings and Comfort,” proposing the NOECS framework based on BIM and IoT, and conducting systematic design and experimental validation. The conclusions are primarily reflected in the following three areas:

1. Building Energy Efficiency and Comfort Prediction: The NOECS model significantly improves multivariable prediction accuracy. By combining Neural ODE with Transformer models, the framework can reliably predict indoor temperature, humidity, and comfort indicators such as PMV. Experimental results show that the MAE/RMSE for temperature is approximately 1.0 °C, for humidity is around 2.3% and 2.7–2.9%, and for PMV, the MAE/RMSE remains below 0.2. Compared to traditional baseline models (such as CNN-LSTM and DERN), NOECS achieves lower prediction errors across multiple metrics, validating the model’s robustness under various operational conditions, and effectively supporting subsequent control decisions.

2. Control Optimization and Energy Consumption Analysis: This paper attributes control benefits from the perspective of individual energy consumption components. Annual energy consumption data shows that HVAC-related loads account for approximately 51.9%, providing a critical control target for energy optimization. Furthermore, NOECS demonstrates outstanding performance in energy savings and energy utilization efficiency. On the BEC dataset, the ESR is 92.3%, with an EUR of 93.8%; on the BIM-BEM dataset, the ESR is 90.7%, and EUR is 89.3%. These results indicate that NOECS can achieve significant energy savings optimization under different data distributions, confirming its feasibility for practical applications.

3. System Deployment and Practical Application: NOECS showcases low computational overhead and high deployment efficiency. On the BEC and BIM-BEM datasets, the training times are 24.5s and 26.7s, respectively, with inference times of 118.7ms and 124.4ms, and the parameter counts are 228.3 M and 232.7 M. Compared to other baseline methods (such as CNN-LSTM and DERN), NOECS demonstrates significant advantages in computational cost. This highlights that NOECS not only performs well in terms of energy savings but also has practical feasibility for online deployment.

Limitations

The current study has several limitations. Cross-season generalization is not fully validated, and performance may vary under different climates. Sensor failures or missing data can affect accuracy, despite mask-based handling. Domain shifts across buildings may require additional calibration, and RL policies need careful consideration for safe deployment in real HVAC systems.

In summary, this paper demonstrates the effectiveness of applying the NOECS model to building energy efficiency prediction and comfort control, providing a replicable technical pathway for the collaborative optimization of energy savings and comfort. Future work will further extend to multi-zone, multi-floor scenarios, explore cross-seasonal and cross-building-type adaptability, and optimize real-time control and intervention strategies to achieve more efficient building operation management and energy-saving goals.

Author contributions

B.T. conceived the study, designed the experiments, performed the research, analyzed the data, and wrote the manuscript. B.T. prepared all figures and tables. The author reviewed and approved the final version of the manuscript.

Funding

No Funding.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.