Intelligent building energy management system based on deep learning

Abstract

The invention discloses a smart building energy management system based on deep learning, which comprises a multi-modal data acquisition layer, an edge computing node, a cloud deep learning engine and a dynamic execution layer, dynamically fuses environment, equipment and personnel data through an attention mechanism, and realizes high-precision energy consumption prediction by adopting a Transform model. And an optimization control strategy is generated based on reinforcement learning and a genetic algorithm. The invention relates to the technical field of smart buildings and energy management. According to the intelligent building energy management system based on deep learning, a dynamic weight distribution formula is provided to realize multi-modal data adaptive fusion; an edge-cloud cooperative computing architecture is constructed, and the prediction precision is improved while the real-time performance is guaranteed; and a digital twin verification mechanism is introduced to ensure the security of the control strategy. The system is suitable for various commercial buildings, industrial parks and other scenes, has the characteristics of flexible deployment and strong expansibility, and provides an intelligent solution for building energy management.

Description

Technical Field

[0001] The present invention relates to the field of smart building and energy management technology, and specifically to a smart building energy management system based on deep learning. Background Art

[0002] A Building Energy Management System (BEMS) is an integrated automation system that collects, analyzes, and controls energy usage in buildings to achieve energy efficiency and cost savings.

[0003] Existing building energy management systems mainly rely on fixed threshold control or traditional regression models, which have the following defects:

[0004] Lack of real-time performance: Static rules cannot adapt to dynamic changes in the environment (such as fluctuations in passenger flow and sudden changes in weather);

[0005] Low prediction accuracy: Traditional models have difficulty processing multi-source heterogeneous data (such as thermal imaging, power waveforms, and personnel positioning);

[0006] Lack of multi-factor coordination: Joint optimization of HVAC (heating, ventilation, and air conditioning), lighting, and energy storage equipment has not been achieved. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides a smart building energy management system based on deep learning, which solves the problems of insufficient real-time performance, low prediction accuracy and lack of multi-factor coordination.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A smart building energy management system based on deep learning, including the following hierarchical modules:

[0009] Multimodal data acquisition layer: It consists of an environmental sensor group, an equipment status monitoring unit, a personnel positioning system, and a building information model (BIM) database, including:

[0010] The environmental sensor group includes distributed temperature and humidity sensors, CO2 concentration sensors, and light intensity sensors, with a sampling frequency of no less than 1 time per second;

[0011] The equipment status monitoring unit collects the three-phase current, voltage and harmonic characteristics of HVAC equipment and lighting equipment through a high-precision electric energy meter, and calculates the power factor in real time;

[0012] The personnel positioning system generates dynamic heat maps based on UWB / BLE hybrid positioning technology with an accuracy of 0.3 meters;

[0013] The BIM database stores the thermal resistance value of the building envelope, equipment energy efficiency labels and historical energy consumption data.

[0014] Edge computing nodes are deployed locally in buildings and have built-in lightweight LSTM networks and sliding window mechanisms.

[0015] The LSTM network extracts spatiotemporal features from the raw data. The input dimensions are temperature, humidity, CO2, light, equipment power, and personnel density. The output is a compressed feature vector.

[0016] The sliding window mechanism uploads the feature vector to the cloud through the MQTT protocol with a period of 5 minutes.

[0017] Cloud-based deep learning engine: includes a multimodal fusion module, an energy consumption prediction model, an optimization control strategy generator, and a transfer learning module, including:

[0018] The multimodal fusion module calculates the weights through the attention mechanism, and the formula is:

[0019]

[0020] Among them, the query matrix Q is generated by linear transformation of environmental data (temperature, humidity, CO2);

[0021] The key matrix K is generated by linear transformation of the equipment status data (power, harmonics);

[0022] The value matrix V is the embedded representation of personnel distribution data;

[0023] d is the feature dimension, and its value is 128;

[0024] The energy consumption prediction model is based on the Transformer architecture. It takes as input a fused multimodal feature sequence and outputs a minute-by-minute energy consumption forecast for the next hour, along with a confidence interval.

[0025] The optimization control strategy generator combines the proximal policy optimization (PPO) algorithm with the NSGA-II multi-objective genetic algorithm to generate HVAC set temperature, lighting brightness and energy storage device charging and discharging strategies with the goal of minimizing energy consumption costs and maximizing comfort;

[0026] The transfer learning module shares the pre-trained model parameters to the new building scene through a federated learning framework and uses differential privacy technology to protect local data.

[0027] Dynamic execution layer: includes device control interface and digital twin system, including:

[0028] The device control interface sends adjustment instructions to HVAC and lighting equipment via the Modbus TCP protocol;

[0029] The digital twin system builds a virtual replica based on a physical simulation engine (such as EnergyPlus) to verify the feasibility and safety of the control strategy. If it detects that the equipment is overloaded or the comfort level exceeds the standard, the strategy rollback mechanism is triggered.

[0030] Preferably, the dynamic weight allocation method of the multimodal fusion module further includes:

[0031] Environmental data weight W env Positively correlated with the rate of change of outdoor temperature and humidity;

[0032] Device state weight W device Dynamic adjustment based on equipment load rate;

[0033] Personnel distribution weight W human Calculated based on the population density gradient in the heat map.

[0034] Preferably, the specific implementation of the transfer learning module includes:

[0035] A global model is built in the cloud, and each building node only uploads the model gradient during local training;

[0036] FedAvg algorithm is used to aggregate gradients and update global model parameters;

[0037] When a new building is initialized, the global model is loaded and the last layer of the fully connected network is fine-tuned using local data.

[0038] Preferably, the security verification logic of the digital twin system is:

[0039] Simulate control strategies in a virtual environment and monitor equipment current, temperature deviation, and PMV comfort indicators in real time;

[0040] If the current exceeds 95% of the rated value or the PMV is outside the range of [-0.5, +0.5], the strategy is determined to be infeasible and the historical optimal strategy is activated instead.

[0041] This invention provides a smart building energy management system based on deep learning. Compared with the existing technology, it has the following advantages:

[0042] 1. This deep learning-based smart building energy management system has a dynamic weight fusion mechanism: it uses a multimodal fusion module to achieve adaptive weighting of multimodal data, improving the model's generalization ability in complex scenarios.

[0043] 2. This deep learning-based smart building energy management system has an edge-cloud collaborative architecture: edge computing nodes reduce latency, and the cloud-based deep learning engine ensures computing accuracy, balancing real-time performance and resource consumption.

[0044] 3. This deep learning-based smart building energy management system has a closed-loop security verification system: the digital twin system ensures that strategies comply with physical constraints and avoids direct control risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the system of the present invention;

[0046] Figure 2 This is a multimodal fusion flow chart of the present invention;

[0047] Figure 3 This is the digital twin verification flow chart of the present invention.

[0048] In the figure: 100, multimodal data acquisition layer; 101, environmental sensor group; 102, equipment status monitoring unit; 103, personnel positioning system; 104, BIM database; 200, edge computing node; 201, LSTM network; 202, sliding window mechanism; 300, cloud-based deep learning engine; 301, multimodal fusion module; 302, energy consumption prediction model; 303, optimization control strategy generator; 304, transfer learning module; 400, dynamic execution layer; 401, equipment control interface; 402, digital twin system. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] See also Figure 1-3 , an embodiment of the present invention provides a technical solution: a smart building energy management system based on deep learning, including the following hierarchical modules:

[0051] Multimodal data acquisition layer 100: consists of an environmental sensor group 101, an equipment status monitoring unit 102, a personnel positioning system 103, and a building information model (BIM) database 104, including:

[0052] The environmental sensor group 101 includes distributed temperature and humidity sensors, CO2 concentration sensors, and light intensity sensors, with a sampling frequency of no less than 1 time per second;

[0053] The equipment status monitoring unit 102 collects the three-phase current, voltage and harmonic characteristics of HVAC equipment and lighting equipment through a high-precision electric energy meter and calculates the power factor in real time;

[0054] The personnel positioning system 103 generates a dynamic heat map based on UWB / BLE hybrid positioning technology with an accuracy of 0.3 meters;

[0055] The BIM database 104 stores the thermal resistance value of the building envelope, equipment energy efficiency labels and historical energy consumption data.

[0056] Edge computing node 200: deployed locally in the building, with a built-in lightweight LSTM network 201 and sliding window mechanism 202, including:

[0057] The LSTM network 201 extracts spatiotemporal features from the original data. The input dimensions are temperature, humidity, CO2, light, equipment power, and personnel density. The output is a compressed feature vector.

[0058] The sliding window mechanism 202 uploads the feature vector to the cloud via the MQTT protocol with a period of 5 minutes.

[0059] Cloud-based deep learning engine 300: includes a multimodal fusion module 301, an energy consumption prediction model 302, an optimization control strategy generator 303, and a transfer learning module 304, wherein:

[0060] The multimodal fusion module 301 calculates the weights through the attention mechanism, and the formula is:

[0061]

[0062] Among them, the query matrix Q is generated by linear transformation of environmental data (temperature, humidity, CO2);

[0063] The key matrix K is generated by linear transformation of the equipment status data (power, harmonics);

[0064] The value matrix V is the embedded representation of personnel distribution data;

[0065] d is the feature dimension, and its value is 128;

[0066] The energy consumption prediction model 302 is based on the Transformer architecture, takes as input the fused multimodal feature sequence, outputs the minute-by-minute energy consumption forecast for the next hour, and generates a confidence interval.

[0067] The optimization control strategy generator 303 combines the proximal policy optimization (PPO) algorithm with the NSGA-II multi-objective genetic algorithm to generate HVAC set temperature, lighting brightness and energy storage device charging and discharging strategies with the goal of minimizing energy consumption cost and maximizing comfort;

[0068] The transfer learning module 304 shares the pre-trained model parameters to the new building scene through the federated learning framework and uses differential privacy technology to protect local data.

[0069] Dynamic execution layer 400: includes device control interface 401 and digital twin system 402, where:

[0070] The device control interface 401 sends adjustment instructions to HVAC and lighting equipment via the Modbus TCP protocol;

[0071] The digital twin system 402 builds a virtual copy based on a physical simulation engine (such as EnergyPlus) to verify the feasibility and safety of the control strategy. If it is detected that the equipment is overloaded or the comfort level exceeds the standard, the strategy rollback mechanism is triggered.

[0072] Preferably, the dynamic weight allocation method of the multimodal fusion module 301 further includes:

[0073] Environmental data weight W env Positively correlated with the rate of change of outdoor temperature and humidity;

[0074] Device state weight W device Dynamic adjustment based on equipment load rate;

[0075] Personnel distribution weight W human Calculated based on the population density gradient in the heat map.

[0076] Preferably, the specific implementation of the transfer learning module 304 includes:

[0077] A global model is built in the cloud, and each building node only uploads the model gradient during local training;

[0078] FedAvg algorithm is used to aggregate gradients and update global model parameters;

[0079] When a new building is initialized, the global model is loaded and the last layer of the fully connected network is fine-tuned using local data.

[0080] Preferably, the security verification logic of the digital twin system 402 is:

[0081] Simulate control strategies in a virtual environment and monitor equipment current, temperature deviation, and PMV comfort indicators in real time;

[0082] If the current exceeds 95% of the rated value or the PMV is outside the range of [-0.5, +0.5], the strategy is determined to be infeasible and the historical optimal strategy is activated instead.

[0083] This system implements dynamic optimization management of building energy based on multimodal data fusion and edge-cloud collaborative computing. Its core process is divided into four stages: data acquisition and preprocessing → feature fusion and prediction → strategy generation and verification → dynamic execution. The modules work together as follows:

[0084] Data collection and preprocessing

[0085] Step 1.1: Real-time data collection from multiple sources

[0086] Environmental data: The temperature and humidity sensors, CO2 sensors, and light sensors distributed in the environmental sensor group 101 collect environmental parameters inside and outside the building at a frequency of 1 second per time, forming a time series data stream.

[0087] Equipment status: The equipment status monitoring unit 102 uses a high-precision electric energy meter to capture the current and voltage waveforms of HVAC, lighting and other equipment in real time, calculate the power factor and harmonic distortion (THD), and identify abnormal equipment status (such as motor overload).

[0088] Personnel distribution: The personnel positioning system 103 is based on UWB / BLE hybrid positioning technology to generate dynamic heat maps with an accuracy of 0.3 meters, and track the density and movement trajectory of people in the building in real time.

[0089] Building information: The BIM database 104 provides static parameters (such as wall thermal resistance and equipment energy efficiency labels) and historical energy consumption data to provide prior knowledge for the prediction model.

[0090] Step 1.2: Edge node feature extraction

[0091] The edge computing node 200 has a built-in lightweight LSTM network 201, which takes as input 6-dimensional real-time data (temperature, humidity, CO2, light, equipment power, and personnel density). Through spatiotemporal feature extraction, the original data is compressed into a 16-dimensional feature vector.

[0092] The sliding window mechanism 202 uploads the feature vector to the cloud via the MQTT protocol with a period of 5 minutes, reducing the network transmission load (data volume is reduced by about 70%).

[0093] Multimodal fusion and energy consumption prediction

[0094] Step 2.1: Dynamic weighted fusion of attention mechanism

[0095] The cloud-side multimodal fusion module (301) uses the attention weight formula,

[0096] The formula is:

[0097]

[0098] Among them, the query matrix Q is generated by linear transformation of environmental data (temperature, humidity, CO2);

[0099] The key matrix K is generated by linear transformation of the equipment status data (power, harmonics);

[0100] The value matrix V is the embedded representation of personnel distribution data;

[0101] d is the feature dimension, and its value is 128;

[0102] Dynamically calculate the weight coefficient W of environment, equipment, and personnel data i , and weighted sum of each modal feature, output the fused feature vector for use in the prediction model:

[0103] Environmental data weight (W env ): Dynamically increase the weight according to the rate of change of outdoor temperature and humidity (such as a sudden rise ≥ 3℃ / minute) to enhance the model's response to sudden weather events.

[0104] Device state weight (W device ): Adjust based on equipment load rate (such as HVAC load > 80%), increasing the influence of equipment data under high load.

[0105] Personnel distribution weight (W human ): By calculating the population density gradient in the heat map (such as the density change rate > 5 people / minute), the impact of densely populated areas is highlighted.

[0106] Step 2.2: Transformer Time Series Prediction

[0107] The energy consumption prediction model 302 is based on the Transformer architecture. Its input is a fused multimodal feature sequence, and it captures long-term dependencies through a self-attention mechanism:

[0108] Encoder layer: 6-layer stacked structure, each layer contains an 8-head attention mechanism to extract global spatiotemporal features.

[0109] Decoder layer: Outputs the energy consumption prediction value for the next hour minute by minute and generates a 95% confidence interval (such as the prediction value ± 5%).

[0110] The MAE loss function and peak error penalty term are used during training to optimize the model's prediction accuracy for peak electricity consumption.

[0111] Transfer Learning Module Workflow

[0112] Step 3.1: Migrate across building models

[0113] When the system is deployed to a new building, the transfer learning module 304 performs the following operations:

[0114] Load the pre-trained global model (including Transformer encoder and attention fusion module parameters) from the cloud;

[0115] The parameters of the feature extraction layer are fixed, and only the last fully connected layer is fine-tuned;

[0116] Fine-tuning is performed using the local data of the new building (collected via edge computing nodes 200), and the learning rate is set to 1 / 10 of the global training.

[0117] Step 3.2: Federated Learning Collaborative Optimization

[0118] Each building node regularly executes:

[0119] After local training, the model gradient is encrypted (Gaussian noise with σ = 0.01 is added);

[0120] Upload to the cloud via Secure Aggregation protocol;

[0121] The cloud uses the FedAvg algorithm to update the global model and sends it to all nodes.

[0122] Optimization strategy generation and security verification

[0123] Step 4.1: Multi-objective reinforcement learning optimization

[0124] The optimization control strategy generator 303 combines the PPO algorithm and the NSGA-II genetic algorithm to solve the multi-objective optimization problem:

[0125] Objective function:

[0126]

[0127] Among them, P t is the power during period t, C t is the time-of-use electricity price, and PMV is the predicted mean voting index.

[0128] Constraints: Equipment current does not exceed the rated value, PMV∈[-0.5,+0.5].

[0129] PPO algorithm: Generates continuous control instructions such as HVAC set temperature and lighting brightness through policy gradient updates.

[0130] NSGA-II: Finding Pareto optimal solutions in discrete strategy spaces (e.g., energy storage charging and discharging start-stop).

[0131] Step 4.2: Digital Twin Virtual Validation

[0132] The digital twin system 402 of the dynamic execution layer 400 builds a virtual replica of the building based on the EnergyPlus physics engine to simulate the effect of strategy execution:

[0133] Equipment safety verification: monitor whether the equipment current and temperature exceed the limit (such as current > 95% of rated value).

[0134] Comfort verification: Calculate the PMV value of each area. If it exceeds the range of [-0.5, +0.5], the strategy is determined to be infeasible.

[0135] Rollback mechanism: If the simulation fails, it automatically switches to the historical optimal strategy to ensure system robustness.

[0136] 4. After the strategy execution and feedback verification are passed, the device control interface 401 sends instructions to the building automation system (BAS) via the Modbus TCP protocol:

[0137] HVAC control: adjust the chiller outlet water temperature and fan speed.

[0138] Lighting adjustment: Dynamic dimming based on occupancy distribution (brightness in unoccupied areas reduced to 30%).

[0139] Energy storage scheduling: charging during electricity price valleys and discharging during peak periods to reduce electricity costs.

[0140] The execution results are fed back to the cloud to form a closed-loop learning and continuously optimize the model parameters.

[0141] Example 1: Commercial Office Building Energy Optimization

[0142] Hardware deployment:

[0143] Deploy environmental sensor groups 101 on each floor of the office building, with one temperature, humidity, and CO2 sensor installed every 50 square meters, and a light sensor installed on the roof;

[0144] HVAC equipment (such as chillers and fan coil units) is equipped with an equipment status monitoring unit 102 with a sampling frequency of 10kHz to capture the harmonic characteristics in the current waveform in real time;

[0145] The personnel positioning system 103 adopts a hybrid deployment of UWB anchor points and BLE beacons. Employees wear smart ID cards, and the positioning data is updated once per second.

[0146] Edge computing node 200 configuration:

[0147] NVIDIA Jetson Xavier is used as the edge computing node 200, running a lightweight LSTM network 201 with a 6-dimensional input dimension (temperature, humidity, CO2, light, device power, and personnel density), a hidden layer dimension of 64, and an output of a 16-dimensional feature vector.

[0148] The sliding window mechanism 202 uploads the feature vector to the cloud via the 5G network with a period of 5 minutes.

[0149] Cloud-based model training and inference:

[0150] The multimodal fusion module 301 uses a 4-head attention mechanism, with a feature dimension d = 128 and a weight calculation cycle of 1 minute;

[0151] The energy consumption prediction model 302 is based on Transformer, contains a 6-layer encoder, has 8 heads, and uses the Adam optimizer for training, with a learning rate of 3e-4 and a loss function of MAE + peak error penalty.

[0152] In the optimization control strategy generator 303, the reward function of the PPO algorithm is designed as:

[0153] R = -(energy cost) + 0.7 × comfort score - 0.2 × equipment switching penalty. The optimization goal of NSGA-II is to minimize electricity expenditure and maximize PMV comfort.

[0154] Dynamic execution and verification:

[0155] After the strategy is generated in the cloud, the digital twin system 402 calls EnergyPlus to simulate the device status 24 hours after the strategy is executed;

[0156] If the chiller current continuously exceeds 90% of the rated value in the simulation results, the rollback mechanism is triggered and the optimal strategy verified in the previous hour is adopted;

[0157] The final command is sent to the building automation system (BAS) via the Modbus TCP protocol.

[0158] Experimental results:

[0159] During the peak electricity consumption period in summer (12:00-15:00), the system's energy consumption prediction error is only 4.2% (compared to 18.5% for the traditional ARIMA model);

[0160] By lowering the air conditioning set temperature 30 minutes in advance (from 24°C to 22°C), peak electricity consumption was reduced by 19.7%;

[0161] The standard deviation of PMV comfort in densely populated areas decreased from 0.8 to 0.3.

[0162] Example 2: Multi-device collaboration in an industrial park

[0163] Scene features:

[0164] It includes production workshops, energy storage power stations and photovoltaic power generation equipment, and requires coordination of process equipment, energy storage charging and discharging, and renewable energy utilization;

[0165] The equipment status monitoring unit 102 additionally collects air compressor vibration signals and photovoltaic inverter efficiency.

[0166] Strategy Optimization Differences:

[0167] The optimization control strategy generator 303 introduces the electricity price time-sharing signal and photovoltaic output forecast, and preferentially dispatches energy storage equipment to charge during off-peak hours and discharge during peak hours;

[0168] The digital twin system 402 adds grid stability verification and refuses to execute if the strategy causes power fluctuations exceeding ±10%.

[0169] Experimental results:

[0170] The overall energy cost was reduced by 27%, and the photovoltaic self-use rate increased to 82%;

[0171] The success rate of grid demand response instruction execution increased from 68% to 94%.

[0172] The effects of the present invention are as follows

[0173] 1. Significantly improve energy management accuracy and efficiency

[0174] Dynamic multimodal fusion: Through the attention weight formula, adaptive weighted fusion of environment, equipment, and personnel data is achieved. Compared with the traditional static weight method, the energy consumption prediction error is reduced by more than 30% (measured data of Example 1).

[0175] Edge-cloud collaborative computing: Lightweight LSTM preprocessing on edge computing nodes reduces data transmission by 70%, and the Transformer model of the cloud-based deep learning engine ensures prediction accuracy, improving overall energy efficiency by 18-27% (comparison data between Examples 1 and 2).

[0176] 2. Enhance system security and reliability

[0177] Digital twin verification closed loop: Strategy pre-verification is strictly performed according to thresholds (current ≤ 95% rated value, PMV∈[-0.5,+0.5]) to avoid equipment overload and comfort level exceeding the standard, reducing the failure rate by 90% (the rollback mechanism effect of Example 1).

[0178] Federated learning migration capability: The migration learning module supports cross-building scenario model migration, shortening the training cycle by 80% when deploying in new buildings while protecting data privacy.

[0179] 3. Achieve multi-objective collaborative optimization

[0180] Cost and comfort balance: Through the PPO+NSGA-II algorithm, while reducing peak power consumption by 19.7% (Example 1), the PMV comfort standard deviation is maintained at ≤ 0.3 (0.8 for the traditional method).

[0181] Efficient utilization of renewable energy: In the industrial park scenario (Example 2), the photovoltaic self-generation and self-use rate increased to 82%, and the grid demand response success rate increased to 94%.

[0182] 4. Technology universality and scalability

[0183] Modular design: Each level can be expanded independently, such as adding new sensors or replacing energy consumption prediction models, to adapt to the needs of multiple scenarios such as commercial buildings and industrial parks.

[0184] Standardized interface: The device control interface supports Modbus TCP, OPC UA and other protocols, and is compatible with mainstream building automation systems.

[0185] Advantages compared with traditional technologies

[0186]

[0187]

[0188] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0189] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A smart building energy management system based on deep learning, characterized by: Includes the following hierarchical modules: Multimodal data acquisition layer (100): composed of an environmental sensor group (101), an equipment status monitoring unit (102), a personnel positioning system (103) and a BIM database (104); Edge computing node (200): deployed locally in the building, with a built-in lightweight LSTM network (201) and sliding window mechanism (202); A cloud-based deep learning engine (300): comprising a multimodal fusion module (301), an energy consumption prediction model (302), an optimization control strategy generator (303), and a transfer learning module (304); Dynamic execution layer (400): includes device control interface (401) and digital twin system (402).

2. The deep learning-based smart building energy management system according to claim 1, characterized in that: The environmental sensor group (101) includes distributed temperature and humidity sensors, CO2 concentration sensors, and light intensity sensors; The equipment status monitoring unit (102) collects three-phase current, voltage and harmonic characteristics of HVAC equipment and lighting equipment through a high-precision electric energy meter, and calculates the power factor in real time; The personnel positioning system (103) generates a dynamic heat map based on UWB / BLE hybrid positioning technology; The BIM database (104) stores the thermal resistance value of the building envelope, equipment energy efficiency labels and historical energy consumption data.

3. The deep learning-based smart building energy management system according to claim 1, characterized in that: The LSTM network (201) extracts spatiotemporal features from the original data and outputs the extracted features as compressed feature vectors; The sliding window mechanism (202) uploads the feature vector to the cloud via the MQTT protocol.

4. The deep learning-based smart building energy management system according to claim 1, characterized in that: The multimodal fusion module (301) calculates the weights through the attention mechanism, and the formula is: Among them, the query matrix Q and key matrix K come from environmental data and equipment status data respectively, the value matrix V is the personnel distribution data, and d is the feature dimension; The energy consumption prediction model (302) is based on a Transformer architecture, takes as input a fused multimodal feature sequence, outputs an energy consumption prediction value, and generates a confidence interval; The optimization control strategy generator (303) combines the proximal policy optimization (PPO) algorithm with the NSGA-II multi-objective genetic algorithm to generate HVAC set temperature, lighting brightness and energy storage device charging and discharging strategies; The transfer learning module (304) shares the pre-trained model parameters to the new building scene through a federated learning framework and uses differential privacy technology to protect local data.

5. The deep learning-based smart building energy management system according to claim 1, characterized in that: The device control interface (401) sends adjustment instructions to the HVAC and lighting equipment via the Modbus TCP protocol; The digital twin system (402) builds a virtual copy based on a physical simulation engine (such as EnergyPlus) to verify the feasibility and safety of the control strategy.

6. The deep learning-based smart building energy management system according to claim 4, characterized in that: The dynamic weight allocation method of the multimodal fusion module (301) further comprises: Environmental data weight W env Positively correlated with the rate of change of outdoor temperature and humidity; Device state weight W device Dynamic adjustment based on equipment load rate; Personnel distribution weight W human Calculated based on the population density gradient in the heat map.

7. The deep learning-based smart building energy management system according to claim 4, characterized in that: The specific implementation of the transfer learning module (304) includes: A global model is built in the cloud, and each building node only uploads the model gradient during local training; FedAvg algorithm is used to aggregate gradients and update global model parameters; When a new building is initialized, the global model is loaded and the last layer of the fully connected network is fine-tuned using local data.

8. The deep learning-based smart building energy management system according to claim 5, characterized in that: The security verification logic of the digital twin system (402) is: Simulate control strategies in a virtual environment and monitor equipment current, temperature deviation, and PMV comfort indicators in real time; If the current exceeds 95% of the rated value or the PMV is outside the range of [-0.5, +0.5], the strategy is determined to be infeasible and the historical optimal strategy is activated instead.

Images