Building energy management method and device based on large model

Abstract

The invention discloses a building energy management method and device based on a large model. The method comprises the following steps: constructing and training a sequence-based building energy management modeling framework used for feature extraction; constructing and training a building adaptive decision model for generating a decision; building data are received in real time, corresponding state representation is generated through the building energy management modeling framework, the corresponding state representation is input into the building self-adaptive decision model, corresponding control actions are generated, and the corresponding control actions are executed and fed back. According to the invention, under the condition of training less building data, various unknown building environments can be successfully coped with, and the adaptability and accuracy of the energy management system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of building energy management, and particularly to a building energy management method and device based on a large model. Background Art

[0002] Buildings are the main source of global electricity consumption and emissions. Therefore, developing an efficient building energy management system (BEMS) is crucial for improving energy efficiency and ensuring indoor comfort. In recent years, artificial intelligence (AI)-based methods, especially deep reinforcement learning (DRL), have received extensive attention in BEMS. However, the generalization ability of AI methods in different building environments poses challenges, making it difficult for them to be effectively migrated and applied between different building types. Although existing control methods can operate effectively in specific environments, when applied to different buildings, they often exhibit poor generalization performance, limiting their effectiveness and universality in practical applications.

[0003] Traditional building energy management methods, such as model predictive control (MPC) and dynamic programming (DP), often face problems of accurate modeling and high computational costs when dealing with complex building thermodynamics and environmental changes. Although deep reinforcement learning (DRL) methods overcome the limitations of these traditional methods, they still suffer from poor generalization. Existing DRL methods usually rely on training for specific buildings and are difficult to achieve good control effects on unknown buildings. In addition, although transfer learning has been proposed as a method to improve generalization, when there are significant differences in the state-action space or operation constraints between the source environment and the target environment, the effect of transfer learning will be greatly reduced. Therefore, how to improve the generalization ability of building energy management systems, especially when facing different building environments, remains an urgent challenge to be solved. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a building energy management method and device based on a large model to improve the adaptability and accuracy of the building energy management system.

[0005] To solve the above technical problem, the embodiments of the present invention propose a building energy management method based on a large model, including: Constructing and training a sequence-based building energy management modeling framework for feature extraction; constructing and training a building adaptive decision model for generating decisions; Receiving building data in real time, generating corresponding state representations through the building energy management modeling framework, inputting the corresponding state representations into the building adaptive decision model, generating corresponding control actions, executing the corresponding control actions, and providing feedback.

[0006] Accordingly, an embodiment of the present invention further provides a building energy management device based on a large model, including a building energy management modeling framework and a building adaptive decision-making model. Among them, The building energy management modeling framework receives building data in real time, generates corresponding state representations, and inputs the corresponding state representations into the building adaptive decision-making model; The building adaptive decision-making model generates corresponding control actions according to the input, executes the corresponding control actions and gives feedback.

[0007] The beneficial effects of the present invention are as follows: By utilizing the few-shot learning ability of the large language model, the present invention learns general strategies in different building energy management, and solves the problem of poor adaptability of existing methods in the face of various building environments; The present invention can successfully cope with various unknown building environments with less building data training, improving the adaptability and accuracy of the energy management system. The present invention combines a sequence-based building energy management modeling framework, transforms the building energy management problem into a unified sequence decision-making task, so as to better capture the general control strategies that can be learned by the large model from building energy data. The present invention proposes an architecture of a building adaptive decision-making Transformer, and learns general building energy management strategies by fine-tuning the large language model. The building adaptive decision-making Transformer contains a building adaptive module, which enables the model to have strong adaptability in the face of different building environments, further improving the generalization effect of downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a schematic structural diagram of the building energy management system according to an embodiment of the present invention.

[0009] Figure 2 is a schematic structural diagram of the building adaptive decision-making model according to an embodiment of the present invention.

[0010] Figure 3 is a bar chart of the experimental results according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0012] In the embodiments of the present invention, if there are directional indications (such as up, down, left, right, front, back...), they are only used to explain the relative position relationship and movement situation between components in a certain specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0013] In addition, in the present invention, descriptions such as "first" and "second" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature.

[0014] The present invention is applied to a building energy management system (BEMS), and uses a sequential building energy management modeling framework (SeqBECMF) and a building adaptive decision-making model (BADT) to improve the generalization ability of the present invention in different buildings. The building energy management method based on a large model according to an embodiment of the present invention includes: Construct and train a sequence-based building energy management modeling framework for feature extraction; construct and train a building adaptive decision-making model for generating decisions; Receive building data in real time, generate corresponding state representations through the building energy management modeling framework, input the corresponding state representations into the building adaptive decision-making model, generate corresponding control actions, execute the corresponding control actions and provide feedback.

[0015] As Figure 1 shown, the building energy management system (BEMS) of the present invention works by coordinating three main components: an energy storage system (ESS), electrical consumption, and power sources. The ESS includes a battery energy storage system (BESS) and a thermal energy storage system (TESS); electrical equipment includes non-adjustable loads, heating equipment, cooling equipment, and domestic hot water (DHW) equipment; power sources include the power grid and a photovoltaic (PV) system. By effectively managing these components, the BEMS aims to achieve the following goals: maintaining thermal comfort, reducing power consumption, improving energy efficiency, and ensuring the resilience of the building during power outages.

[0016] At each moment within, the operation period is , where T is the total number of time steps. The control agent executes an optimization model and generates control actions based on the measurement data of the system and the observation data set by the user . These control actions include charging and discharging operations of the energy storage system and the power output of electrical equipment, thereby ensuring the efficient coordination of various energy resources.

[0017] The operation of the ESS is controlled by its state of charge (SoC), which is defined as the ratio of the current energy level to the total capacity. The control action represents the percentage of the system capacity for charging ( ) or discharging ( ), where . For the BESS, the operation of the SoC is restricted by a safety range , and usually takes a value of , to prevent overcharging or deep discharging. For TESS, the SoC can operate across the entire range without significant risk of damage. The evolution formula for the SoC of the ESS is shown below: ; where, represents the state of charge at time , is the maximum charge and discharge power, is the total energy storage capacity, and are the charge and discharge efficiencies respectively, is the self-discharge rate.

[0018] BESS and TESS can be further subdivided. BESS only includes electrical energy storage and provides charge and discharge support for electrical equipment such as space heating, space cooling, and domestic hot water equipment. TESS includes three specialized thermal energy storage types: cooling, heating, and domestic hot water storage. Cooling storage provides energy for cooling equipment (such as air-conditioning heat pumps); heating storage provides energy for heating equipment (such as heat pumps and electric heaters); domestic hot water storage specifically provides energy for domestic hot water systems.

[0019] For electrical energy storage (controlled by BESS constraints), the SoC is expressed as , with an energy capacity of and a maximum charge and discharge power of . BESS operates within a safe SoC range , typically taking a value of to avoid overcharging or deep discharging. The control action represents the percentage of the system's maximum capacity for charging ( ) or discharging ( ), where .

[0020] For thermal energy storage, the SoC is divided into cooling , heating and domestic hot water storage , with energy capacities of respectively, and maximum power limits of respectively. TESS can operate across the full SoC range without risk of damage. The control actions for TESS represent the maximum capacity percentages for cooling, heating, and domestic hot water storage respectively, where .

[0021] Electrical equipment includes an HVAC system for space cooling, heating, and domestic hot water heating, which is typically modeled as an adjustable heat pump. The power consumption of these devices depends on control actions, which are defined as a proportion of the nominal power of the device. The power consumption of the cooling, heating, and domestic hot water devices is as follows: ; where 、 and are the control actions of the cooling, heating, and domestic hot water devices, respectively, 、 and are the nominal powers of the device. The control action is normalized, where 0 represents no operation and 1 represents the device operating at maximum power.

[0022] The main energy sources of the building include the power grid and the photovoltaic (PV) system. Under normal circumstances, the power grid provides an unrestricted power supply and preferentially meets the energy demands of the building. During a power outage, the building relies on the PV system and the energy storage system (ESS) as backup power sources to maintain basic operations. The PV system generates on-site power while the ESS, including battery energy storage and thermal energy storage, provides additional flexibility by discharging the stored energy. In the present invention, whether a power outage occurs is determined by a stochastic outage model based on the reliability index of the distribution system to ensure a realistic simulation of the grid reliability.

[0023] At each time step the power provided by the PV system depends on the solar power generation capacity and environmental conditions. The net power contribution of the energy storage system is calculated as the sum of the outputs of the electrical energy storage 、cooling energy storage 、heating energy storage and hot water supply energy storage : ; The total power consumption of the building can be calculated by the following formula: , where , and represent the power consumed by the cooling, heating, and hot water supply devices, respectively. Represents non-adjustable loads, including household appliances such as sockets and lighting. By leveraging grid power, photovoltaic systems, and energy storage systems, buildings can effectively balance energy supply and demand, reduce dependence on external energy, and optimize energy use under various operating scenarios.

[0024] The optimization problem in BEMS aims to achieve a balance between thermal comfort, minimizing power consumption, and ensuring a smooth energy consumption curve. The objective function quantifies this trade-off by considering three key factors: thermal comfort, total power consumption, and power consumption fluctuations.

[0025] Thermal comfort is represented by the discomfort time steps and quantifies the proportion of time steps during which the indoor temperature deviates from the desired set temperature by more than the allowed tolerance . It is defined as: , where is the total number of time steps during the operating period. This term ensures minimizing the thermal discomfort time during BEMS operation.

[0026] Total power consumption considers the total energy consumption of all building systems, including heating, ventilation, and air conditioning (HVAC), energy storage, and other household appliances, and its expression is:

[0027] where represents the power consumption of the building at each time step .

[0028] Power consumption fluctuations quantify the smoothness of the building's energy consumption curve. It measures the degree of abrupt changes in load demand, which can have a negative impact on grid stability and increase operating costs. Power consumption fluctuations are defined as: , where and are the power consumptions of the building at consecutive time steps and respectively. A smaller value indicates a smoother energy consumption curve with fewer abrupt changes.

[0029] Weighted objective: The overall objective function can be expressed as: ,

[0030] where the weights satisfy: ,

[0031] Among them, the weight factors , and respectively control the relative importance of thermal comfort, energy consumption, and smoothness (fluctuation) in the optimization objective. Specifically, minimizing thermal discomfort is given priority, focusing on reducing power consumption, aiming to reduce fluctuations to obtain a smoother energy consumption curve. The values of these parameters can be adjusted according to different operation priorities. Usually, the range of these values is from 0 to 1, and higher values correspond to greater emphasis on the relevant objectives.

[0032] This multi-objective optimization problem ensures the effective operation of the BEMS under different conditions while balancing competing objectives such as comfort, efficiency, and stability.

[0033] The sequence-based building energy management modeling framework (SeqBECMF) integrates states, actions, and rewards in the form of time series into the MDP framework, avoiding explicit modeling of state transitions or assuming the Markov property. To enable large language models (LLMs) to learn meaningful and general control strategies and generate expected actions during testing, this framework focuses on the entire trajectory, capturing past and current conditions. This approach tightly connects the immediate state-action relationship with long-term goals such as thermal comfort, energy efficiency, and load stability, optimizing the overall effect of building energy management.

[0034] 1. Action: At each time step the action includes decisions for multiple building subsystems, involving cooling, heating, domestic hot water, and energy storage. The action is defined as: where each component controls one aspect of building energy use or storage (e.g., cooling / heating equipment, water heating, and battery management).

[0035] ; where, , , , , , and respectively represent the control variables of the cooling equipment, heating equipment, domestic hot water equipment, cooling energy storage, heating energy storage, domestic hot water energy storage, and electrical energy storage. By controlling these actions, the BEMS can distribute energy between equipment and energy storage systems, enabling the building to respond to changing energy demands while maintaining an efficient and comfortable living environment.

[0036] 2. Status: The status consists of the information received by the agent from the environment, enabling it to make real-time decisions. The selection of state variables should comprehensively reflect the environmental information while avoiding irrelevant content. In the BEMS problem, the state space is represented as: This observation space contains four categories of information, specifically as follows: .

[0037] 2.1. Time Status: ; Time-related observations include month , day of the week and hour . These provide the basis for seasonal changes, occupancy patterns, and daily energy usage variations.

[0038] 2.2. Weather Status: Weather-related observations include indoor temperature , outdoor temperature , direct solar radiation and diffuse solar radiation . These factors affect the building's heat load and energy demand.

[0039] 2.3. Energy Status: ; Energy-related observations include solar power generation , non-adjustable load , and various energy demands (cooling demand , heating demand , hot water demand ). It also includes the state of different energy storage systems (SoC): domestic hot water storage , electrical energy storage , cooling energy storage , heating energy storage . These observations assist in energy allocation and storage management.

[0040] 2.4. Other Status: ; Other status includes the number of occupants , carbon intensity and power outage status . These provide additional context for managing energy demand during high carbon emissions or power outages.

[0041] 3. Status: The reward function aims to balance three key objectives: (1) alleviating the discomfort of occupants by penalizing the deviation of the indoor temperature from the set point; (2) minimizing the total energy consumption; (3) smoothing the sudden changes in building loads. Its formula is as follows: ; where the coefficients 、 and are consistent with those defined in the optimization objectives, reflecting their roles in prioritizing thermal comfort, energy efficiency, and load stability. 、 are the indoor temperature and the set value of the indoor comfort temperature at time t, respectively.

[0042] 4. Sequence Modeling: Traditional DRL methods predict actions based on the current state and implicitly learn rewards over time. However, Transformer-based architectures excel in sequence data because they can capture long-term dependencies more effectively than isolated state points. To leverage this advantage, SeqBECMF redefines the trajectory of the Markov decision process (MDP) as sequence data and reformulates the reward structure using the concept of going back to the future. This approach provides the model with future rewards to be achieved, enabling it to make more informed decisions, considering not only the current state but also the long-term reward potential. SeqBECMF explicitly incorporates future performance into the decision-making process, using a target reward , capturing the cumulative reward from the current time step to the end of the trajectory, defined as: ; where represents the reward obtained at time step t'. It represents the cumulative reward from time step to the end. This enables the model to optimize its actions with an eye on future outcomes rather than just focusing on immediate rewards.

[0043] To fully utilize the sequence modeling capabilities of large models based on the Transformer architecture, the trajectory is redefined such that the model can effectively capture meaningful patterns: ; where each triple represents the target reward, the observed state, and the corresponding action at time step . represents the trajectory of control decisions.

[0044] This trajectory structure unifies information from past states and actions, current observations, and future goals into a modeling framework. By explicitly conditioning actions on , SeqBECMF enables the Transformer to align immediate decisions with long-term goals. By unifying data from different buildings into a modeling framework, SeqBECMF enables the Transformer to learn control strategies applicable to different building environments. This approach leverages the flexibility of the Transformer-based architecture and can be generalized to different types of buildings and hardware configurations without the need for extensive retraining.

[0045] As an implementation, the framework of the Building Adaptive Decision-making Model (BADT) is as Figure 2 shown. The BADT framework aims to provide adaptive and efficient energy management for the building environment. BADT integrates multiple key components. The following will explain the main components of the BADT framework in detail according to the Figure 2 inference steps.

[0046] The Observation Forecaster module provides prediction information through time series prediction, thereby enhancing decision-making capabilities. The forecaster generates one-hour-ahead predictions of key variables, enabling the agent to predict future conditions and proactively manage energy. Its input is steps of historical state data and generates the following predictions: ; where represents the prediction function, is the predicted observation value, represents the historical data sequence composed of and . The extended state at time is defined as: ; where is the predicted observation value, represents the state in the extended

[0047] By incorporating these predictions into the state, the present invention can understand environmental and load dynamics such as external temperature, photovoltaic power generation, and energy demand in advance, thereby realizing more intelligent and forward-looking energy strategy adjustments and improving energy efficiency and living comfort.

[0048] LLM backbone network: The LLM BackBone of the BADT framework is responsible for processing the input sequence and generating control actions. It consists of several key components: an embedding layer, layer normalization, Transformer layers, and a decoder.

[0049] 1) Input Tokens are defined as: ; Among them, each contains a triple of the target reward , state and action . The input has a total of tokens.

[0050] 2) Embeddings Layer The BADT framework uses the embedding layer to integrate the target reward, state, and action information at each time step. The embedding of position is defined as: , , , Among them, , , and represent the embedding functions of the target reward, state, action, and position encoding respectively.

[0051] The embeddings of all time steps will be concatenated into a sequence: .

[0052] 3) Layer Normalization: To stabilize the training and normalize the input data, layer normalization (Layer Norm) is applied to the concatenated embedding sequence: .

[0053] 4) Transformer Layers: The normalized input sequence is processed through stacked Transformer layers to generate context-aware latent variables: .

[0054] Each Transformer block includes multi-head self-attention (MHA), residual connections, layer normalization, and a feed-forward neural network (FFN). In the MHA mechanism, the query, key, and value matrices are calculated as , and , where , and are learnable weights. The self-attention is calculated as follows: .

[0055] The outputs of multiple attention heads will be concatenated together: .

[0056] After MHA, the residual connection adds the input back, followed by layer normalization:

[0057] The normalized output will pass through a feed-forward neural network (FFN) to apply token-level transformations: ; where , , , are learnable parameters. Next, another residual connection and normalization operation are as follows: .

[0058] These steps are repeated in all Transformer blocks to generate the final latent variable sequence .

[0059] 5) Linear decoder: The latent variable generates control actions through a Linear Decoder: , where and are learnable parameters.

[0060] Buildings usually exhibit different device characteristics, such as differences in energy storage capacity and maximum output power, resulting in heterogeneous action spaces. Although the action has been normalized to the standardized range , due to the existence of operation constraints between buildings, the action ranges of different buildings are often affected by device parameters and are not consistent. For example, some buildings may operate within [-0.5, 0.5], while others utilize [-0.7, 0.7]. Adopting a unified action space may severely limit the model's ability to generalize in different building environments. This diversity of action spaces makes it difficult for the model to generalize because different buildings require different control strategies to adapt to their unique constraints.

[0061] To address the problem of different action spaces, the BADT of the present invention introduces a Building Adapter module as a key component in the BADT framework. The BA enhances the generalization ability of the model by dynamically adjusting and scaling the original actions according to the building-specific metadata (such as energy storage capacity and power).

[0062] The adaptation process combines the original actions with the building metadata (such as storage capacity and rated power, etc.), and passes the synthesized vector to a deep neural network (DNN) to generate scaled actions . This process ensures that the actions can be customized according to the unique operation constraints of each building. This relationship can be defined as: .

[0063] By adopting this adaptation process, the actions are tailored according to the equipment parameters of the building (such as rated power), so as to better adapt to the specific action space of each building. This method ensures a more efficient and scalable solution, enabling the model to better adapt to the diverse action spaces of different buildings and significantly improving its generalization performance.

[0064] The present invention incorporates the prior knowledge of experts by collecting the interaction data of pre-trained deep reinforcement learning (DRL) methods with different building environments. These data contain expert interaction trajectories and are processed according to the SeqBECMF framework to form a supervised learning dataset . During the training process, a pair is sampled from the dataset , where is the th Input token in the data, and is the action made according to expert experience as the training label. The BADT model generates actions as follows: .

[0065] The training loss is the mean squared error (MSE) between the predicted action and the expert action : , where is the batch size. The parameters of the entire BADT are represented by and are updated by the gradient descent method with a learning rate of : .

[0066] The building energy management device based on a large model according to an embodiment of the present invention includes a building energy management modeling framework and a building adaptive decision-making model. Among them, The building energy management modeling framework receives building data in real time, generates corresponding state representations, and inputs the corresponding state representations into the building adaptive decision-making model; The building adaptive decision-making model generates corresponding control actions according to the input, executes the corresponding control actions, and gives feedback.

[0067] As an implementation method, the building energy management modeling framework inputs states, actions, and rewards in the form of a time series into the building adaptive decision-making model. Among them, at each time step on, action satisfies: ; Among them, , , , , , and respectively represent the control variables of the refrigeration equipment, heating equipment, domestic hot water equipment, refrigeration energy storage, heating energy storage, domestic hot water energy storage, and electrical energy storage; state satisfies: ; ; ; ; is the time state, is the month, week type, is the hour; is the weather state, including the indoor temperature , outdoor temperature , direct solar radiation and diffuse solar radiation ; is the energy state, including solar power generation , non-adjustable load , refrigeration demand , heating demand , hot water demand , domestic hot water energy storage , electrical energy storage , Refrigeration energy storage , Heating energy storage ; For other states, including the number of occupants U t , Carbon intensity and power outage status ; Reward function Satisfies: ; , and Are the weight factors for controlling thermal comfort, energy consumption, and smoothness respectively, , Are the indoor temperature and the set value of the indoor comfort temperature at time t respectively; Is the total power consumption of the building.

[0068] As an implementation, the building energy management modeling framework satisfies: ; ; Is the target reward; Represents the reward obtained at time step t', and represents the cumulative reward from time step to the end; Each triple Represents the target reward, the observed state, and the corresponding action at time step ; Represents the trajectory of the control decision.

[0069] As an implementation, the building adaptive decision model includes an observation predictor module, an LLM backbone network, and a building adaptive module. The observation predictor module provides prediction information through time series prediction. The LLM backbone network is responsible for processing the input sequence and generating control actions. The building adaptive module dynamically adjusts and scales the original actions according to the building metadata; The observed values predicted by the observation predictor module satisfy: ; Among them, Represents the prediction function, Is the predicted observed value, Represents the historical data sequence composed of and .

[0070] As an implementation, the LLM backbone network consists of an embedding layer, layer normalization, Transformer layers, and a linear decoder. Among them, the embedding layer concatenates the embeddings of all time steps into a sequence ; layer normalization normalizes the sequence to obtain a sequence ; the Transformer layers process the normalized sequence to generate a latent variable sequence ; the linear decoder generates control actions based on the latent variable sequence .

[0071] Experimental design and result analysis of the present invention: 1. Experimental design.

[0072] To evaluate the effectiveness of BADT, the simulation used the NeurIPS CityLearn Challenge dataset, which contains six buildings with different thermodynamic characteristics, covering 2,207 hourly time steps from June to August 2018. Four buildings were used for training, and the other two buildings were used for testing to evaluate the generalization ability of the model. The training dataset came from a pre-trained deep reinforcement learning (DRL) strategy. The simulation was conducted in the OpenAI Gym environment, running on an Nvidia RTX3090 server, and the device configuration information used is shown in Table 1.

[0073]

[0074] The BADT framework uses the GPT-2 Small model with 124 million parameters as the backbone network. This model can effectively capture temporal dependencies by modeling past state, action, and reward sequences, thus enabling adaptive control of the building energy management system (BEMS). For comparison, four widely used DRL methods were used as benchmarks: DDPG, Twin Delayed Deep Deterministic Policy Gradient (TD3), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC). Among them, DDPG and TD3 perform excellently in continuous control, and TD3 can solve the overestimation bias problem. PPO is widely used due to its robustness in policy learning, and SAC provides high sample efficiency for stochastic policies.

[0075] The evaluation of BEMS control performance is based on four key scores: thermal discomfort score , total power consumption score , electricity consumption smoothness score and overall performance score. These scores comprehensively evaluate thermal comfort, energy efficiency, and grid stability. Thermal discomfort score Consistent with the definition of the objective function, expressed as a percentage. Total power consumption score Based on the total power consumption As defined, it is normalized by dividing by the baseline power consumption where represents the result obtained without applying active control. Power consumption smoothness score Based on As defined, it is normalized by dividing by the baseline power consumption smoothness where represents the result without active control. Finally, the overall performance score is calculated as the weighted sum of these normalized scores.

[0076] The parameter settings of the experiment are shown in Table 2. The algorithm neural network uses a multi-layer perceptron as the basic framework. Among them, for the actor network, the input dimension is 8 and the output dimension is 3. For the critic network, the input dimension is (8, 3) and the output dimension is 1. The neural network has 2 hidden layers, and the number of neurons in each layer is 256.

[0077]

[0078] Experimental results and analysis: 1. Generalization performance evaluation across different metrics.

[0079] This experiment evaluated the performance of BADT in optimizing energy management objectives, including its application in unseen building configurations. As Figure 3 shown, BADT was compared with four DRL algorithms on four key metrics. The results showed that BADT achieved higher scores, fewer hours of thermal discomfort, smoother fluctuations, and lower power consumption on both the training set and the test set, outperforming the DRL baseline algorithms. The numerical annotations above the bar charts show the specific performance. BADT demonstrated stronger stability and minimal performance degradation on unseen buildings. This strong generalization ability is attributed to BADT's ability to utilize the few-shot learning ability of large language models (LLMs). By training only on data from four buildings, BADT effectively learned control strategies that could generalize well to different building environments.

[0080] 2. Ablation study of the building adaptation module.

[0081] This ablation study explored the impact of the BA module on the model performance by comparing the performance of BADT with its baseline algorithm DT (without the BA module) on the training set and the test set. The experimental results are shown in Table 3, providing important insights into the role of the BA module in the model performance.

[0082] On the training set, BADT outperforms DT (without the BA module) on multiple key metrics. Specifically, BADT has a significantly lower score (0.429 vs. 0.645), and a significantly reduced number of hours of thermal discomfort (UH) (0.082 vs. 0.202), indicating that the BA module enhances the model's ability to meet energy demands, thus improving overall performance. These results fully demonstrate the significant role of the BA module in optimizing energy demand management and enhancing model efficiency.

[0083] On the test set, BADT also outperforms DT (without the adapter) in reducing the number of hours of thermal discomfort (0.186 vs. 0.496), showing the BA module's stronger generalization ability to unseen building environments. Although the performance score of BADT slightly decreases (0.473 vs. 0.698), this result further verifies that the BA module helps maintain the efficiency and stability of the model, even when dealing with new data. Despite the decrease in the performance score during the test phase, the decrease is smaller compared to DT, indicating that the BA module enhances the model's generalization ability and robustness in new environments.

[0084] However, the addition of the BA module also introduces trade - offs, especially in terms of carbon emissions (CE) and variability (RA). On the training set, BADT has higher carbon emissions (0.9583 vs. 0.910) and variability (0.939 vs. 0.907). This is because the BA module enables the model to better match energy demands, thereby increasing overall energy consumption and grid fluctuations. In contrast, DT (without the adapter) pays less attention to energy efficiency and comfort, resulting in lower carbon emissions and more stable variability. The increase in carbon emissions and variability reflects the trade - off between energy efficiency and stability while ensuring comfort and energy demand satisfaction.

[0085] In summary, the introduction of the BA module significantly enhances the model's ability to meet energy demands and generalize to unseen buildings, as demonstrated by the lower number of hours of thermal discomfort and the smaller decrease in performance during the test phase. However, the BA module also has negative impacts on carbon emissions and variability. Therefore, when applying this module, it is necessary to carefully consider the trade - offs it brings and further optimize to balance multiple objectives such as energy efficiency, comfort, and environmental impact.

[0086]

[0087] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalent scope.

Claims

1. A building energy management method based on large models, characterized in that, Including: Construct and train a sequence-based building energy management modeling framework for feature extraction; Construct and train a building adaptive decision-making model for generating decisions; Receive building data in real time, generate corresponding state representations through the building energy management modeling framework, input the corresponding state representations into the building adaptive decision-making model, generate corresponding control actions, execute the corresponding control actions and provide feedback.

2. The building energy management method based on a large model according to claim 1, characterized in that, The building energy management modeling framework inputs the state, action, and reward into the building adaptive decision-making model in the form of a time series. Among them, at each time step on, Action Satisfy: ； Among them, , , , , , and respectively represent the control variables of the refrigeration equipment, heating equipment, domestic hot water equipment, refrigeration energy storage, heating energy storage, domestic hot water energy storage, and electrical energy storage; Status Satisfy: ； ； ； ； is the time status, is the month, the week type, is the hour; is the weather condition, including indoor temperature , outdoor temperature , direct solar radiation and diffuse solar radiation ; is the energy state, including solar power generation , non-adjustable load , cooling demand , heating demand , hot water demand , domestic hot water energy storage , electrical energy storage , cooling energy storage , heating energy storage ; For other states, including the number of occupants U t , carbon intensity and power outage status ; Reward function Satisfy: ； , and are the weighting factors for controlling thermal comfort, energy consumption, and smoothness, respectively. , are the indoor temperature and the set value of the indoor comfort temperature at time t, respectively. is the total power consumption of the building.

3. The building energy management method based on a large model according to claim 2, wherein The building energy management modeling framework satisfies: ； ； is the target reward; represents the reward obtained at time step t', and represents the cumulative reward from time step to the end; each triple represents at time step the target reward, the observed state, and the corresponding action at; represents the trajectory of the control decision.

4. The building energy management method based on a large model according to claim 2, wherein The building adaptive decision-making model includes an observation predictor module, an LLM backbone network, and a building adaptation module. The observation predictor module provides prediction information through time series prediction. The LLM backbone network is responsible for processing the input sequence and generating control actions. The building adaptation module dynamically adjusts and scales the original actions according to building metadata; The observed values predicted by the observation predictor module satisfy: ； Among them, represents the prediction function, is the predicted observed value, represents the historical data sequence composed of and .

5. The building energy management method based on a large model according to claim 4, characterized in that, The LLM backbone network consists of an embedding layer, layer normalization, Transformer layers, and a linear decoder. Among them, the embedding layer concatenates the embeddings of all time steps into a sequence ; layer normalization normalizes the sequence to obtain the sequence ; the Transformer layers process the normalized sequence to generate a latent variable sequence ; the linear decoder generates control actions based on the latent variable sequence .

6. An architecture energy management device based on a large model, characterized in that, Including a building energy management modeling framework and a building adaptive decision-making model, where The building energy management modeling framework receives building data in real time, generates corresponding state representations, and inputs the corresponding state representations into the building adaptive decision-making model; The building adaptive decision-making model generates corresponding control actions according to the input, executes the corresponding control actions and provides feedback.

7. The building energy management device based on a large model according to claim 6, characterized in that, The building energy management modeling framework inputs states, actions, and rewards into the building adaptive decision-making model in the form of a time series, where, at each time step on, Action Satisfy: ； Among them, , , , , , and respectively represent the control variables of the refrigeration equipment, heating equipment, domestic hot water equipment, refrigeration energy storage, heating energy storage, domestic hot water energy storage and electrical energy storage; Status Satisfy: ； ； ； ； is the time status, is the month, the week type, is the hour; is the weather condition, including indoor temperature , outdoor temperature , direct solar radiation and diffuse solar radiation ; is an energy state, including solar power generation , non-adjustable load , cooling demand , heating demand , hot water demand , domestic hot water energy storage , electrical energy storage , cooling energy storage , heating energy storage ; For other states, including the number of occupants U t , carbon intensity and power outage status ; Reward function Satisfy: ； , and are the weighting factors for controlling thermal comfort, energy consumption, and smoothness, respectively. , are the indoor temperature and the set value of the indoor comfort temperature at time t, respectively. is the total power consumption of the building.

8. The building energy management device based on a large model according to claim 7, characterized in that, The building energy management modeling framework satisfies: ； ； is the target reward; represents the reward obtained at time step t', and represents the cumulative reward from time step to the end; each triple represents at time step the target reward, the observed state, and the corresponding action at; represents the trajectory of the control decision.

9. The building energy management device based on a large model according to claim 7, characterized in that, The building adaptive decision-making model includes an observation predictor module, an LLM backbone network, and a building adaptation module. The observation predictor module provides prediction information through time series prediction. The LLM backbone network is responsible for processing the input sequence and generating control actions. The building adaptation module dynamically adjusts and scales the original actions according to building metadata; The observed values predicted by the observation predictor module satisfy: ； Among them, represents the prediction function, is the predicted observed value, represents the historical data sequence composed of and ​ 10. The building energy management device based on a large model according to claim 9, characterized in that, The LLM backbone network consists of an embedding layer, layer normalization, Transformer layers, and a linear decoder. Among them, the embedding layer concatenates the embeddings of all time steps into a sequence ; layer normalization normalizes the sequence to obtain the sequence ; the Transformer layers process the normalized sequence to generate a latent variable sequence ; the linear decoder generates control actions based on the latent variable sequence .

Images