Building energy management and optimization design method, system, equipment and medium

Abstract

The invention relates to a building energy management and optimization design method, system and device and a medium. The method comprises the following steps: based on an original data set of a target building, performing calculation through a preset gating circulation unit network to obtain a dynamic environmental protection weight factor; performing feature extraction processing on the equipment operation parameters, the environmental sensor readings and the user comfort feedback data in the original data set to obtain corresponding features, and performing fusion processing on the features and the dynamic environmental protection weight factors through a cross-modal attention mechanism to generate a unified state vector; and inputting the unified state vector into a preset strategy network for decision reasoning processing, and outputting an equipment control action set. According to the method, through a strategy network of dynamic environmental protection weight factor and cross-modal data fusion, the low-carbon, economic and comfortable target collaborative optimization precision of building energy management is improved, and the intelligent level of building energy management and the adaptability to energy use scene changes are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of building energy management technology, and in particular relates to building energy management and optimization design methods, systems, equipment and media. Background Technology

[0002] With the global energy crisis and the advancement of "dual carbon" goals, the level of precision and intelligence in building energy management has become a focus of industry attention, as buildings are the core carriers of energy consumption and carbon emissions. Existing building energy management methods have achieved sensor-based energy consumption data collection and basic optimization. For example, by monitoring equipment operating parameters and adjusting equipment start-up and shutdown to reduce energy consumption, or by combining user feedback to initially optimize indoor comfort, some solutions have also incorporated grid carbon intensity data to take into account low-carbon goals.

[0003] However, existing building energy management technologies still have significant shortcomings: First, the integration of carbon intensity factors often uses fixed weights, failing to dynamically adjust the priority of environmental protection goals based on real-time changes in grid carbon intensity gradients, renewable energy output, and actual building energy consumption status. This results in insufficient precision in the synergistic optimization of low-carbon, economic, and comfort goals. Second, the processing of multi-source data, such as equipment operating parameters, environmental sensor readings, and user comfort feedback, is mostly done through independent analysis or simple splicing, lacking a deep fusion mechanism for cross-modal data, making it difficult to form a unified representation reflecting the overall state of the building energy system. Third, the reasoning of the strategy network during the decision-making process is often based on a single fixed goal, such as optimizing only energy consumption, without incorporating dynamic environmental weights to achieve dynamic balance of multiple goals. Furthermore, there is a lack of an update mechanism for the strategy network, resulting in poor adaptability to changes in building energy consumption scenarios. Summary of the Invention

[0004] Therefore, it is necessary to provide building energy management and optimization design methods, systems, equipment and media to address the above-mentioned technical problems, aiming to improve the overall optimization effect of energy management in terms of economy, environmental protection and comfort.

[0005] Firstly, this application provides building energy management and optimization design methods, including:

[0006] The original dataset of the target building is obtained. Based on the original dataset, the dynamic environmental protection weight factor is obtained by calculation through a pre-defined gated cyclic unit network. The original dataset includes real-time carbon intensity of the power grid, time-of-use electricity price, equipment operating parameters, environmental sensor readings and user comfort feedback data.

[0007] Feature extraction is performed on the equipment operating parameters, environmental sensor readings and user comfort feedback data in the original dataset to obtain the corresponding features. The features are then fused with the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector.

[0008] The unified state vector is input into a preset strategy network for decision-making and reasoning processing, and the output is a set of device control actions.

[0009] In one embodiment, based on the original dataset, a dynamic environmental protection weight factor is obtained through calculation processing using a pre-defined gated recurrent unit network, including:

[0010] Gradient calculation is performed based on time-series data of real-time carbon intensity of the power grid to obtain the historical gradient of real-time carbon intensity of the power grid.

[0011] Power calculation is performed based on the equipment operating parameters in the original dataset to obtain the real-time power purchase of the building.

[0012] The operating parameters of the local renewable energy power generation equipment of the target building are collected, and the output is calculated based on the operating parameters of the local renewable energy power generation equipment to obtain the local renewable energy output.

[0013] The user comfort feedback data in the original dataset is parsed and processed to obtain a spatial comfort matrix. The mean of the spatial comfort matrix is ​​calculated to obtain the overall average comfort of the building.

[0014] The real-time carbon intensity of the power grid, the historical gradient of the real-time carbon intensity of the power grid, the real-time power purchase of the building, the local renewable energy output, and the overall average comfort of the building are input into a preset gated loop unit network. The network is then processed by an activation function to output a dynamic environmental protection weight factor.

[0015] In one embodiment, feature extraction is performed on the device operating parameters, environmental sensor readings, and user comfort feedback data in the original dataset to obtain corresponding features. These features are then fused with a dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector, including:

[0016] The time-series data corresponding to the device operating parameters in the original dataset are processed by a one-dimensional convolutional network to extract features and obtain the device operating feature vector.

[0017] The environmental feature vector is obtained by extracting features from the environmental sensor readings in the original dataset through an encoder network.

[0018] A graph neural network is used to extract features from the user comfort feedback data in the original dataset to obtain a user comfort feature vector.

[0019] Using the device operation feature vector as the query basis, and the environmental feature vector and the user comfort feature vector as key-value pairs, a cross-modal attention mechanism is used to perform fusion processing to generate fused features;

[0020] Construct a trainable weight matrix, and perform a dot product operation on the dynamic environmental protection weight factor and the trainable weight matrix to obtain the dot product result.

[0021] The dot product operation result and the fused features are processed by element-wise multiplication to generate a unified state vector.

[0022] In one embodiment, the preset policy network is a near-end policy optimization network, which is constructed based on the policy gradient algorithm and includes a policy network module and a value network module.

[0023] The unified state vector is input into a preset policy network for decision-making and reasoning processing, and the output is a set of device control actions, including:

[0024] Cost accounting is performed based on the time-of-use electricity price and equipment operating parameters in the original dataset. The current energy cost is obtained by multiplying the real-time power in the equipment operating parameters with the time-of-use electricity price for the corresponding time period and then summing the results.

[0025] Emissions are calculated based on the real-time carbon intensity of the power grid and equipment operating parameters in the original dataset. The current carbon emissions are obtained by multiplying the power consumption in the equipment operating parameters by the real-time carbon intensity of the power grid and then summing the results.

[0026] The overall average comfort level of the building was calculated based on user comfort feedback data in the original dataset.

[0027] Obtain the energy cost weight and construct an economic incentive item based on the energy cost weight and the current energy cost; obtain the carbon emission weight and construct an environmental incentive item based on the dynamic environmental protection weight factor, the carbon emission weight, and the negative current carbon emissions; obtain the comfort weight and multiply the comfort weight by the overall average comfort level of the building to obtain a comfort incentive item.

[0028] The economic incentives, environmental incentives, and comfort incentives are added together to generate the incentive value.

[0029] The unified state vector is input into the proximal policy optimization network, and processed by the policy network module and the value network module respectively, outputting the corresponding initial action distribution and state value.

[0030] By combining reward value and state value, the initial action distribution is optimized to obtain an optimized action distribution, and a set of device control actions is output based on the optimized action distribution.

[0031] In one embodiment, the user comfort feedback data in the original dataset is parsed to obtain a spatialized comfort matrix, including:

[0032] The feedback text in the user comfort feedback data is processed by sentiment analysis using a natural language processing model to obtain sentiment polarity scores.

[0033] The physiological indicators in user comfort feedback data are processed using a machine learning model to map comfort levels, resulting in a comfort index; the physiological indicators include body temperature and heart rate.

[0034] Acquire the user's indoor location data, and based on the indoor location data, associate the emotional polarity score and comfort index with the building zoning of the target building to obtain the association results;

[0035] A spatial comfort matrix is ​​constructed based on the correlation results.

[0036] In one embodiment, the method further includes:

[0037] Execution instructions are generated based on the set of equipment control actions and sent to the equipment corresponding to the target building. The execution instructions are used to instruct the equipment to perform the corresponding operation according to the set of equipment control actions.

[0038] After collecting a set of equipment control actions from multiple sources, a new original dataset is obtained. The new original dataset includes real-time carbon intensity of the new power grid, new time-of-use electricity price, new actual operating parameters of the new equipment, new environmental sensor readings, and new user comfort feedback data.

[0039] Feature extraction is performed on the new original dataset to obtain and construct a new unified state vector based on the corresponding new features;

[0040] Based on the new original dataset, the real-time carbon intensity gradient of the new power grid, the real-time power purchase of new buildings, the output of new local renewable energy, and the overall average comfort of new buildings were calculated.

[0041] The parameters of the pre-defined gated cyclic unit network are fine-tuned by taking the real-time carbon intensity gradient of the new power grid, the real-time power purchase of new buildings, the output of new local renewable energy, and the overall average comfort of new buildings as inputs, and using the historical best scheduling results as labels, to obtain the updated gated cyclic unit network parameters.

[0042] In one embodiment, the spatial comfort matrix is ​​constructed using the following formula:

[0043]

[0044] in, Let be the spatialized comfort matrix at time t, with dimension . , The number of building zones for the target building. Let t be the total number of users who participate in comfort feedback. The location weight for the i-th user is used to characterize the importance of the building zone where the user is located. Let be the emotional polarity score of the i-th user at time t. Let be the comfort index of the i-th user at time t.

[0045] Secondly, this application also provides a building energy management and optimization design system, including:

[0046] The raw data processing module is used to obtain the raw dataset of the target building. Based on the raw dataset, it performs calculations through a preset gated cyclic unit network to obtain dynamic environmental protection weight factors. The raw dataset includes real-time carbon intensity of the power grid, time-of-use electricity price, equipment operating parameters, environmental sensor readings, and user comfort feedback data.

[0047] The feature fusion-driven module is used to extract features from the equipment operating parameters, environmental sensor readings and user comfort feedback data in the original dataset, obtain the corresponding features, and fuse each feature with the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector.

[0048] The strategy decision execution module is used to input a unified state vector into a preset strategy network for decision reasoning processing and output a set of device control actions.

[0049] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0050] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0051] The aforementioned building energy management and optimization design methods, systems, equipment, and media first acquire multi-source raw datasets and generate dynamic environmental protection weight factors through a gated cyclic unit network, solving the problems of fixed low-carbon weights and poor adaptability to environmental protection objectives in traditional methods. Second, features are extracted from the multi-source data and combined with the dynamic environmental protection weight factors for cross-modal attention fusion, generating a unified state vector. This overcomes the shortcomings of independent processing and insufficient fusion depth of multi-source data, improving the accuracy of state representation. Finally, the unified state vector is input into a preset strategy network to output a set of device control actions, addressing the shortcomings of single-objective decision-making and multi-objective collaborative optimization, and enhancing the comprehensive optimization effect of energy management in terms of economy, environmental protection, and comfort. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart of a building energy management and optimization design method provided as an exemplary embodiment of the present invention;

[0054] Figure 2 A flowchart of a method for obtaining dynamic environmental protection weight factors is provided as an exemplary embodiment of the present invention;

[0055] Figure 3 A schematic diagram of a building energy management and optimization design system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] In one embodiment, such as Figure 1 As shown, a building energy management and optimization design method is provided. This embodiment illustrates the application of this method to a building energy system. In this embodiment, the method includes the following steps:

[0058] S101: Obtain the original dataset of the target building. Based on the original dataset, perform calculations through a pre-defined gated cyclic unit network to obtain dynamic environmental protection weight factors. The original dataset includes real-time carbon intensity of the power grid, time-of-use electricity price, equipment operating parameters, environmental sensor readings, and user comfort feedback data.

[0059] Specifically, the optimization decisions for building energy systems need to respond simultaneously to dynamic changes on the grid side, equipment side, environment side, and user side. Therefore, it is possible to first obtain the original dataset of the target building, which includes real-time carbon intensity of the grid, time-of-use pricing, equipment operating parameters, environmental sensor readings, and user comfort feedback data. Real-time carbon intensity of the grid directly reflects the environmental attributes of current grid power generation and is a core input for low-carbon target optimization. Time-of-use pricing provides a quantitative basis for energy cost control. Equipment operating parameters, such as the power, start / stop status, and operating time of equipment like air conditioners, lighting, and water pumps, characterize the energy consumption level and operating status of the equipment. Environmental sensor readings, such as indoor temperature and humidity, light intensity, and CO2 concentration, reflect the building's internal environmental conditions. User comfort feedback data is directly related to the user experience and is a key reference for comfort optimization. Subsequently, based on the above original dataset, dynamic environmental weight factors can be constructed using a pre-defined gated recurrent unit (GRU) network. The gated recurrent unit network is a recurrent neural network structure that can effectively handle long-term and short-term dependencies in time-series data. In this embodiment, the GRU network learns the time series features in the original dataset and dynamically adjusts the environmental protection weight factor, thereby reflecting the priority of low-carbon goals and other goals such as economy and comfort at different time points and under different operating conditions.

[0060] S102: Feature extraction is performed on the equipment operating parameters, environmental sensor readings and user comfort feedback data in the original dataset to obtain the corresponding features. The features are then fused with the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector.

[0061] Specifically, equipment operating parameters, environmental sensor readings, and user comfort feedback data belong to different modalities, with significant differences in their data structure, feature distribution, and impact mechanisms on energy optimization. Equipment operating parameters are mostly time-series continuous data, with core features reflecting dynamic information such as power fluctuation patterns and start-stop cycles. Environmental sensor readings are spatially distributed data, with key features being the spatial gradient and temporal rate of change of parameters such as temperature, humidity, and illumination. User comfort feedback data are mostly discrete subjective or physiologically related data, with core features being the intensity and distribution characteristics of comfort needs. Using a single method to process multimodal data can easily lead to feature distortion or loss of key information. Therefore, feature extraction can be performed on each of the three types of data separately, thereby maximizing the preservation of core representational information for each modality and laying the foundation for subsequent fusion processing. The extracted features and dynamic environmental weight factors can then be fused using a cross-modal attention mechanism to generate a unified state vector. The cross-modal attention mechanism is a data fusion technique that can effectively handle the differences and correlations between data from different modalities. In this embodiment, based on this mechanism, the contribution of each feature in the fusion process can be dynamically adjusted by calculating the attention weight between each feature and the dynamic environmental protection weight factor. The attention weight can be calculated based on the similarity or correlation between the feature and the weight factor. Then, based on this weight, the features can be fused into a unified state vector through weighted summation. This unified state vector can comprehensively and accurately reflect the overall state of the building energy system, providing high-quality input data for subsequent decision-making and reasoning.

[0062] S103: Input the unified state vector into the preset strategy network for decision-making and reasoning processing, and output a set of device control actions.

[0063] Specifically, the policy network is a deep learning-based decision-making model that outputs the optimal control strategy based on the input state vector. In this embodiment, the policy network is trained by learning from a large amount of historical and simulated data to achieve a decision-making model that can synergistically optimize low-carbon, economic, and comfort goals. Furthermore, the unified state vector includes multi-source data information after feature extraction and fusion processing, as well as dynamically adjusted environmental weight factors, enabling the policy network to fully consider the actual operating state of the building and changes in the external environment during the decision-making process. During decision inference, the unified state vector serves as the network input. After feature transformation and mapping in the network's hidden layers, it can directly output a set of control actions for each building device. This set of control actions can include device start / stop commands, power adjustment parameters, and operating mode switching commands, thereby enabling intelligent and dynamic optimization of the building's energy system and significantly improving the overall efficiency of energy management.

[0064] The aforementioned method first calculates dynamic environmental protection weight factors based on the original dataset of the target building through a gated cyclic unit network, overcoming the shortcomings of fixed weights in existing technologies and achieving a dynamic balance between low-carbon, economic, and comfort goals. Second, it extracts features from equipment operating parameters, environmental sensor readings, and user comfort feedback data, and fuses them through a cross-modal attention mechanism to generate a unified state vector, solving the problem of independent analysis or simple splicing of multi-source data and improving the depth and accuracy of data fusion. Finally, the unified state vector is input into a policy network for decision-making inference, outputting a set of equipment control actions, addressing the lack of a policy network update mechanism, enhancing adaptability to changes in energy consumption scenarios, and significantly improving the refinement and intelligence of building energy management.

[0065] In one embodiment, such as Figure 2 As shown, based on the original dataset, dynamic environmental protection weight factors are obtained through calculation and processing using a pre-defined gated recurrent unit network, including:

[0066] S201: Gradient calculation is performed based on time-series data of real-time carbon intensity of the power grid to obtain the historical gradient of real-time carbon intensity of the power grid;

[0067] S202: Based on the equipment operating parameters in the original dataset, perform power calculation to obtain the real-time power purchase of the building;

[0068] S203: Collect the operating parameters of the local renewable energy power generation equipment of the target building, perform output calculation based on the operating parameters of the local renewable energy power generation equipment, and obtain the local renewable energy output;

[0069] S204: Parse and process the user comfort feedback data in the original dataset to obtain a spatial comfort matrix. Calculate the mean of the spatial comfort matrix to obtain the overall average comfort level of the building.

[0070] S205: Input the real-time carbon intensity of the power grid, the historical gradient of the real-time carbon intensity of the power grid, the real-time power purchase of the building, the local renewable energy output, and the overall average comfort of the building into a preset gated loop unit network, process it through the activation function, and output a dynamic environmental protection weight factor.

[0071] Specifically, the real-time carbon intensity of the power grid is not a static value; its trend over time directly affects the priority of environmental protection goals. For example, when carbon intensity shows an upward gradient, future carbon emission costs will increase, requiring an earlier increase in environmental protection weight to mitigate subsequent high emission risks. Conversely, when carbon intensity shows a downward gradient, environmental protection priorities can be adjusted appropriately while ensuring comfort and economic efficiency. In this embodiment, continuous real-time carbon intensity time-series data of the power grid for a preset duration, such as one hour, can be selected from the original dataset to construct a carbon intensity time series. Then, the first-order difference method is used to calculate the difference in carbon intensity between adjacent moments. Finally, the average of each difference within the preset duration is taken to obtain the historical gradient of the real-time carbon intensity of the power grid. This gradient parameter quantifies the rate and direction of carbon intensity change, providing dynamic characteristics of carbon intensity for the GRU network and compensating for the decision-making lag caused by relying solely on the current carbon intensity value. Furthermore, the real-time electricity purchase power of a building directly reflects the scale of energy consumption from the power grid and is the core basis for carbon emission calculation. Its product with the real-time carbon intensity of the power grid can approximately characterize the current carbon emission rate of the building. Furthermore, the device operating parameters in the original dataset include information such as the real-time power and operating status of each electrical device. Therefore, by summarizing the real-time power of all operating devices, the total power consumption of the building can be obtained. Then, by deducting the power generated and consumed by local renewable energy sources, the building's real-time power purchase can be calculated. If renewable energy data has not yet been collected, the power generated and consumed by local renewable energy sources is assumed to be 0.

[0072] Specifically, the output scale of local renewable energy sources such as photovoltaic, wind power, and energy storage systems directly affects the building's dependence on the power grid, and thus the total carbon emissions. In other words, the higher the renewable energy output, the lower the demand for electricity from the grid, and the lower the carbon emissions. Therefore, the adjustment logic for environmental weighting needs to be adjusted accordingly. For example, taking photovoltaic equipment, operating parameters such as irradiance of the photovoltaic panels, module temperature, and inverter efficiency can be collected first. Then, calculations can be performed based on the equipment's power characteristic model, as shown in the following formula:

[0073]

[0074] in, This indicates the real-time output of the photovoltaic equipment. Indicates real-time irradiance. This represents the conversion efficiency of the photovoltaic module, with a value range of [0,1]. Indicates the total effective area of ​​the photovoltaic module. This represents the power loss due to temperature, and its value is positively correlated with the temperature of the photovoltaic module. It can be calculated using data from the module's temperature sensor combined with a preset loss model. For wind power equipment, the real-time output can be calculated using the corresponding wind power characteristic model (built based on parameters such as wind speed and rotor speed). This output parameter reflects the utilization level of clean energy in the building.

[0075] Furthermore, optimizing environmental protection goals cannot come at the expense of user comfort. The overall average comfort of the building is a key constraint for balancing environmental protection and comfort. Therefore, when comfort falls below a preset threshold, even with a higher environmental weight, it is necessary to prioritize adjusting equipment operation to improve comfort. Illustratively, this embodiment can perform structured analysis on user comfort feedback data, mapping discrete feedback information such as comfortable, fairly comfortable, and uncomfortable to quantitative scores such as 1-5. Then, combining this with building spatial zoning information, the feedback scores of each user are associated with their respective zoning, constructing a spatialized comfort matrix with zoning as rows and time slices as columns. Each element in the matrix represents the comfort score of a zoning at a specific time. Finally, the arithmetic mean of all elements in the spatialized comfort matrix is ​​taken to obtain the overall average comfort of the building. This overall average comfort quantifies the overall comfort level of the building, providing a quantitative basis for comfort constraints in the GRU network and ensuring that the calculation of dynamic environmental weight factors is always performed above the comfort baseline.

[0076] The real-time carbon intensity of the power grid, its historical gradient, the building's real-time electricity purchases, local renewable energy output, and the building's overall average comfort level are comprehensively covered from the dimensions of power grid environmental attributes, building energy consumption characteristics, and user comfort constraints. These factors comprehensively encompass the core factors influencing the priority of environmental protection goals and can form the input vector for a GRU network. During the pre-training phase, the GRU network learns the mapping rules between different input combinations and environmental weights. During inference, the historical information retention ratio of each input parameter can be controlled by updating the gate, and noisy data can be filtered by resetting the gate. This allows for in-depth mining of the temporal characteristics and correlation rules of the input vector. Finally, the network output is input into a Sigmoid activation function, which maps the output value to the [0,1] interval, resulting in a dynamic environmental weight factor. This factor can adapt in real-time to fluctuations in power grid carbon intensity, changes in building energy load, fluctuations in renewable energy output, and user comfort needs, solving the problem that traditional fixed weights cannot dynamically balance multiple objectives. This provides environmental guidance for subsequent multimodal data fusion and decision-making inference.

[0077] In another embodiment, the user comfort feedback data in the original dataset is parsed to obtain a spatialized comfort matrix, including:

[0078] The feedback text in the user comfort feedback data is processed by sentiment analysis using a natural language processing model to obtain sentiment polarity scores.

[0079] The physiological indicators in user comfort feedback data are processed using a machine learning model to map comfort levels, resulting in a comfort index; the physiological indicators include body temperature and heart rate.

[0080] Acquire the user's indoor location data, and based on the indoor location data, associate the emotional polarity score and comfort index with the building zoning of the target building to obtain the association results;

[0081] A spatial comfort matrix is ​​constructed based on the correlation results.

[0082] Specifically, user text feedback often suffers from semantic ambiguity and individual differences in expression. Therefore, natural language processing (NLP) techniques can be used to quantify semantic representation. For example, the BERT sentiment analysis model, with its bidirectional Transformer architecture, can accurately capture the semantic relationships within the text context. First, the feedback text is preprocessed: stop words are removed and words are split using word segmentation tools. The segmented results are then converted into WordPiece units, and positional and segment encodings are combined to generate word embedding vectors. During the model inference phase, these embedding vectors are processed by a 12-layer Transformer encoder to extract semantic features. A fully connected layer then outputs probability distributions for negative, neutral, and positive sentiments, ultimately converting these probability distributions into sentiment polarity scores. This transformation effectively eliminates semantic ambiguity in subjective feedback by converting unstructured text into standardized quantitative scores. Physiological indicators, including body temperature and heart rate, are objective representations of human comfort. Their numerical fluctuations have a non-linear relationship with comfort levels. Illustratively, physiological indicators can be Z-score standardized to eliminate dimensional differences. Subsequently, a random forest regression model was used to construct the mapping relationship between physiological indicators and the comfort index. This model adapts to nonlinear associations through ensemble learning of multiple decision trees. During model training, standardized body temperature and heart rate can be used as input features, and user-annotated comfort scores (1-5 points) can be used as labels to optimize hyperparameters such as the number of decision trees and maximum depth. During the inference phase, the comfort index can be output, compensating for individual biases in subjective textual feedback and achieving a collaborative representation of the subjective and objective comfort state.

[0083] Specifically, due to the spatial heterogeneity of user comfort needs, ignoring spatial attributes can lead to local discomfort being masked by global averaging. Therefore, ultra-wideband (UWB) positioning technology can be used to obtain three-dimensional coordinates by measuring the time-of-flight difference between the positioning tag and the anchor point within the building. The target building is pre-divided into zones, defining the planar coordinate range of each zone to form a zone number-coordinate range mapping table. Subsequently, a coordinate matching algorithm can be used to determine the zone to which the user's positioning coordinates belong, while introducing positioning weights to characterize the importance of the zone, such as a conference room having a higher weight than a corridor. By associating the user's emotional polarity score and comfort index with the target building's building zones and corresponding weights, a structured association result can be obtained. Based on this result, a spatial comfort matrix can then be constructed using the following formula to achieve a spatial-temporal structured expression of comfort data:

[0084]

[0085] in, Let be the spatialized comfort matrix at time t, with dimension . , The number of building zones for the target building. Let t be the total number of users who participate in comfort feedback. The location weight for the i-th user is used to characterize the importance of the building zone where the user is located. Let be the emotional polarity score of the i-th user at time t. Let be the comfort index of the i-th user at time t.

[0086] The matrix constructed by this formula intuitively presents the spatial-temporal two-dimensional distribution of comfort states. For example, the value in the g-th row and g'-th column of the matrix can directly reflect the comfort correlation (or its own comfort) between the g-th partition and the g'-th partition. This can provide partition-level comfort constraints for subsequent calculation of dynamic environmental protection weight factors and lay the foundation for the policy network to output precise equipment control commands for each partition.

[0087] In one embodiment, feature extraction is performed on the device operating parameters, environmental sensor readings, and user comfort feedback data in the original dataset to obtain corresponding features. These features are then fused with a dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector, including:

[0088] The time-series data corresponding to the device operating parameters in the original dataset are processed by a one-dimensional convolutional network to extract features and obtain the device operating feature vector.

[0089] The environmental feature vector is obtained by extracting features from the environmental sensor readings in the original dataset through an encoder network.

[0090] A graph neural network is used to extract features from the user comfort feedback data in the original dataset to obtain a user comfort feature vector.

[0091] Using the device operation feature vector as the query basis, and the environmental feature vector and the user comfort feature vector as key-value pairs, a cross-modal attention mechanism is used to perform fusion processing to generate fused features;

[0092] Construct a trainable weight matrix, and perform a dot product operation on the dynamic environmental protection weight factor and the trainable weight matrix to obtain the dot product result.

[0093] The dot product operation result and the fused features are processed by element-wise multiplication to generate a unified state vector.

[0094] Specifically, the temporal fluctuation patterns of equipment operating parameters such as air conditioning power, water pump speed, and lighting load, including air conditioning start-stop cycles, peak power periods, and load response characteristics to environmental changes, can directly reflect the equipment's energy consumption level and operating status. One-dimensional convolutional networks (1D-CNNs) can capture local correlation features of temporal data through sliding local convolutional kernels, avoiding the problem of ignoring short-term fluctuation details in traditional time series analysis. Therefore, this network can be used to extract features from the temporal data corresponding to equipment operating parameters. For example, the equipment operating parameters can first undergo temporal preprocessing, that is, continuous temporal data can be divided into fixed-length time segments according to a preset time window, such as 5 minutes / window, and Min-Max normalization can be used to map the data to the [0,1] interval, eliminating the dimensional differences between different equipment parameters. Subsequently, a 1D-CNN network structure was constructed. The input layer receives preprocessed data with dimensions of [temporal window length, number of device parameters]. This data is then passed through three convolutional layers, two max-pooling layers, and finally a global average pooling layer and a fully connected layer to flatten the three-dimensional feature tensor into a 512-dimensional device operation feature vector. This vector accurately characterizes the dynamic features of device operation, providing a quantitative representation of the device state for subsequent fusion processing. The convolutional kernel sizes were set to 5, 3, and 3, with the number of kernels being 64, 128, and 256 respectively. The ReLU activation function was used to enhance nonlinear fitting capabilities. The pooling window size was 2, and the stride was 2, achieving feature dimensionality reduction while preserving key fluctuation features.

[0095] Specifically, environmental sensor readings, such as indoor temperature and humidity, light intensity, CO2 concentration, and outdoor meteorological parameters, exhibit significant spatial distribution characteristics and dynamic correlations. For example, the temperature and humidity in south-facing zones are more affected by outdoor sunlight, and there is a causal relationship between CO2 concentration and the operating status of ventilation equipment. Data from a single sensor cannot reflect the overall distribution and interaction patterns of the building environment.

[0096] In this embodiment, a Transformer encoder based on a self-attention mechanism can be used to construct the encoder network. This allows for the capture of spatial correlations and temporal dynamics of multi-sensor data through a multi-head attention mechanism, avoiding the limitations of traditional single-sensor independent analysis. For example, missing sensor data is first filled in using inverse distance weighted interpolation. The sensor data is then categorized by building zone, constructing a three-dimensional data tensor of zone-sensor type-time series. This tensor is then converted into an encoder input format of sequence length (number of zones × time step) × feature dimension (number of sensor types). The input is a Transformer structure containing six encoder layers, each layer containing an eight-head self-attention mechanism and a feedforward neural network. The self-attention mechanism automatically assigns attention weights by calculating the similarity between different zones and different sensor data, capturing the spatial correlations and dynamic changes of environmental parameters. Finally, pooling layers and fully connected layers map the encoded features into a 512-dimensional environmental feature vector. This vector integrates the spatial distribution and dynamic changes of the built environment, providing a comprehensive representation of environmental conditions for decision-making.

[0097] Specifically, the same user exhibits consistent behavior across different zones, and comfort needs in adjacent zones can influence each other; for example, low comfort in a meeting room may spread to adjacent office areas. Traditional vectorization processes can disrupt this spatial relationship. Therefore, this embodiment utilizes a graph neural network, specifically a graph attention network (GAT), to capture spatial relationship features through node association modeling, achieving a structured representation of comfort needs. For instance, each zone of a building can be used as a graph node, with undirected edges (representing spatial connectivity) constructed between adjacent zones. The initial feature of each node is the temporal comfort value of the corresponding zone in the spatialized comfort matrix. Subsequently, a network structure containing three layers of GAT is constructed. Each layer calculates the attention weights between adjacent nodes through a multi-head attention mechanism. For example, adjacent nodes with smaller comfort value differences have higher weights, reflecting the spatial synergy of comfort needs. Furthermore, the current node's features can be updated by aggregating the features of adjacent nodes, achieving feature transfer from local comfort needs to global comfort distribution. Finally, the final features of all nodes are averaged and mapped to a 512-dimensional user comfort feature vector through a fully connected layer. This vector not only contains the comfort level of each zone but also integrates the spatial relationship patterns of comfort needs, providing a user demand representation that aligns with actual usage scenarios for decision-making.

[0098] Furthermore, the device operation feature vector can be used as the query vector (Q), and the environmental feature vector and user comfort feature vector can be concatenated to form the key vector (K) and value vector (V), ensuring that the dimensions of the three are consistent, such as 512. Then, by multiplying the transpose of Q and K, scaling (dividing by the square root of the feature dimension to avoid gradient vanishing), and performing Softmax normalization, an attention weight matrix (512×1024 dimensions) can be obtained. Higher values ​​in this weight matrix indicate a stronger correlation between the corresponding key feature and the query feature. Finally, by multiplying this weight matrix with the V matrix, a fusion feature of dimension 512 can be obtained. This fusion process can dynamically strengthen key related features through attention weights. For example, when the device operation feature indicates that the air conditioner is running at high power, the weights of "the deviation of temperature and humidity from the set value" in the environmental feature and "the proportion of discomfort feedback" in the user feature will automatically increase, allowing the fusion feature to focus on the core correlation between device operation, environmental conditions, and user needs, providing an integrated state basis for decision-making. Furthermore, since the dynamic environmental protection weight factor is a scalar value (range [0,1]) representing the priority of the current environmental protection goal, while the fusion feature is a high-dimensional vector, direct multiplication would result in the environmental protection orientation having a consistent impact on all feature dimensions, failing to adapt to the different correlations between different features and environmental protection goals. Therefore, a trainable weight matrix of dimension 512×512 can be constructed, with its initial value initialized using Xavier to ensure gradient stability during training. Each element of this matrix represents the environmental protection correlation weight of the corresponding feature dimension. By performing a dot product operation between the dynamic environmental protection weight factor and this matrix, a dot product operation result of dimension 512 can be obtained, thereby enabling the differentiated impact of environmental protection priority on different feature dimensions. For example, when the environmental protection priority is high, the weights of feature dimensions related to equipment power and carbon emissions in the matrix are amplified. By multiplying the dot product result (dimension 512) with the fusion feature (dimension 512) one by one according to the corresponding dimensions, a unified state vector of dimension 512 can be obtained. This vector provides the policy network with precise state input integrating four dimensions: device, environment, user, and environmental protection. It solves the technical problems of one-dimensional data and chaotic multi-dimensional data fusion in traditional decision-making, laying the foundation for the accuracy of subsequent intelligent decision-making.

[0099] In one embodiment, the preset policy network is a near-end policy optimization network, which is constructed based on the policy gradient algorithm and includes a policy network module and a value network module.

[0100] The unified state vector is input into a preset policy network for decision-making and reasoning processing, and the output is a set of device control actions, including:

[0101] Cost accounting is performed based on the time-of-use electricity price and equipment operating parameters in the original dataset. The current energy cost is obtained by multiplying the real-time power in the equipment operating parameters with the time-of-use electricity price for the corresponding time period and then summing the results.

[0102] Emissions are calculated based on the real-time carbon intensity of the power grid and equipment operating parameters in the original dataset. The current carbon emissions are obtained by multiplying the power consumption in the equipment operating parameters by the real-time carbon intensity of the power grid and then summing the results.

[0103] The overall average comfort level of the building was calculated based on user comfort feedback data in the original dataset.

[0104] Obtain the energy cost weight and construct an economic incentive item based on the energy cost weight and the current energy cost; obtain the carbon emission weight and construct an environmental incentive item based on the dynamic environmental protection weight factor, the carbon emission weight, and the negative current carbon emissions; obtain the comfort weight and multiply the comfort weight by the overall average comfort level of the building to obtain a comfort incentive item.

[0105] The economic incentives, environmental incentives, and comfort incentives are added together to generate the incentive value.

[0106] The unified state vector is input into the proximal policy optimization network, and processed by the policy network module and the value network module respectively, outputting the corresponding initial action distribution and state value.

[0107] By combining reward value and state value, the initial action distribution is optimized to obtain an optimized action distribution, and a set of device control actions is output based on the optimized action distribution.

[0108] Specifically, the real-time power time-series data of each device in the original dataset (such as the power per second of air conditioners and lighting equipment) and the corresponding time-of-use electricity price data (matched to the peak-valley-flat electricity price table published by the power company) can be extracted first. Then, the time periods are divided according to a preset time window, such as 15 minutes, and the real-time power of the devices in each time period is integrated to obtain the power consumption of that time period. The power consumption of each time period is multiplied by the time-of-use electricity price of the corresponding time period, and finally the product results of all time periods are summed to obtain the current energy cost. This calculation process accurately quantifies the economic cost of building energy use, provides a direct basis for the economic orientation of the reward function, and enables decisions to avoid unnecessary high-load operation during high-electricity-price periods, thereby improving the economic efficiency of energy utilization.

[0109] Furthermore, the carbon intensity of the power grid directly determines the carbon emission intensity of building energy consumption, while the power consumption of equipment is the carrier of carbon emissions. Therefore, based on the real-time power of the equipment operating parameters, the power consumption of each equipment during a given time period can be obtained through integral calculation (power × time interval). The real-time carbon intensity data of the power grid for the corresponding time period can then be extracted. The power consumption during each time period is multiplied by the carbon intensity for each time period and then summed to obtain the current carbon emissions. This process accurately characterizes the scale of carbon emissions from building energy consumption, providing a quantitative basis for the environmental orientation of the reward function. This enables decisions to proactively reduce emissions when carbon intensity is high and rationally release load when carbon intensity is low, achieving dynamic adaptation between low carbon and energy demand. In addition, user comfort is one of the core constraints of energy management; if decisions sacrifice comfort, they lose their application value. Therefore, based on the generated spatial comfort matrix, the arithmetic mean of all elements in the matrix can be taken to obtain the overall average comfort of the building. This mean integrates the comfort status of each zone and time period of the building, providing a quantitative basis for the comfort orientation of the reward function, ensuring that user comfort is always the bottom line when optimizing economic and environmental goals.

[0110] Specifically, the economic reward can be obtained by multiplying the energy cost weight by the negative current energy cost, indicating that the lower the cost, the higher the reward. The environmental reward can be obtained by multiplying the dynamic environmental weight factor, the carbon emission weight, and the negative current carbon emissions, indicating that the less emissions, the higher the reward. The dynamic environmental weight factor allows environmental priority to be dynamically adjusted according to factors such as grid carbon intensity. The comfort reward can be obtained by multiplying the comfort weight by the overall average comfort level of the building, indicating that the higher the comfort level, the higher the reward. Adding these three reward items generates a reward value. This reward value transforms the three heterogeneous objectives of economy, environmental protection, and comfort into a unified quantitative indicator, enabling the strategy network to dynamically balance these three aspects and avoid the unbalanced results caused by optimizing a single objective. The unified state vector is input into the proximal strategy optimization network. When input into the strategy network module, it can be mapped through a multi-layer neural network to output the initial action distribution, such as the probability distribution of control actions for each device, including the adjustment range of air conditioning temperature and the probability of device start / stop commands. When inputting into the value network module, the value evaluation network can output the state value, which can realize a quantitative assessment of the quality of the current state. The higher the value, the higher the long-term benefit of taking the optimal action in that state.

[0111] Specifically, by combining reward values ​​and state values, the stable optimization of action distribution can be achieved by pruning the objective function. For example, firstly, the ratio of action probabilities between the old and new strategies is calculated. By limiting this ratio to a preset range, such as [0.8, 1.25], excessive policy updates can lead to training instability. Then, a loss function is constructed using reward values ​​and state values ​​to guide the strategy to iterate towards an action distribution with high rewards and high state values. The optimized action distribution integrates feedback from multiple objectives—rewards and state values. By sampling or selecting the action with the highest probability, the set of equipment control actions can be obtained. This process enables decision-making to respond to real-time economic, environmental, and comfort needs while ensuring the long-term stability and convergence of the strategy, thereby achieving intelligent and dynamic optimized operation of the building energy system under multi-objective constraints.

[0112] In one embodiment, the method further includes:

[0113] Execution instructions are generated based on the set of equipment control actions and sent to the equipment corresponding to the target building. The execution instructions are used to instruct the equipment to perform the corresponding operation according to the set of equipment control actions.

[0114] After collecting a set of equipment control actions from multiple sources, a new original dataset is obtained. The new original dataset includes real-time carbon intensity of the new power grid, new time-of-use electricity price, new actual operating parameters of the new equipment, new environmental sensor readings, and new user comfort feedback data.

[0115] Feature extraction is performed on the new original dataset to obtain and construct a new unified state vector based on the corresponding new features;

[0116] Based on the new original dataset, the real-time carbon intensity gradient of the new power grid, the real-time power purchase of new buildings, the output of new local renewable energy, and the overall average comfort of new buildings were calculated.

[0117] The parameters of the pre-defined gated cyclic unit network are fine-tuned by taking the real-time carbon intensity gradient of the new power grid, the real-time power purchase of new buildings, the output of new local renewable energy, and the overall average comfort of new buildings as inputs, and using the historical best scheduling results as labels, to obtain the updated gated cyclic unit network parameters.

[0118] Specifically, building energy systems involve diverse equipment types (such as air conditioners, lighting, and water pumps) and differ in their control protocols (such as Modbus, BACnet, and TCP / IP). Therefore, based on the equipment's communication interfaces and control logic, the set of equipment control actions can be converted into executable instructions that the equipment can recognize. For example, air conditioner temperature adjustment instructions need to include parameters such as target temperature and adjustment rate, while lighting control instructions need to specify the on / off status and brightness level. Subsequently, multi-source data can be collected after the equipment executes the set of equipment control actions to obtain a new raw dataset. Among these, the new real-time carbon intensity of the power grid reflects the change in carbon emission attributes on the grid side after the actions are executed. The new time-of-use electricity pricing can provide a new benchmark for subsequent economic cost accounting. The new actual operating parameters of the equipment, such as the actual power of the air conditioner and the equipment's start / stop status, can verify the execution accuracy of the control actions, such as whether there is a delay in command response or a deviation in power adjustment. New environmental sensor readings, such as indoor temperature and humidity and CO2 concentration, can reflect the effect of the control actions on the building environment. New user comfort feedback data can directly assess the impact of the control actions on the comfort experience.

[0119] Specifically, based on the technical logic of constructing a unified state vector in the above embodiments, feature extraction processing is performed on the new original dataset to obtain and construct a new unified state vector based on the corresponding new features. Specifically, new features of device operation are extracted from the time-series data of new equipment operating parameters using a one-dimensional convolutional network; new environmental features are extracted from the readings of new environmental sensors using an encoder network; and new user comfort features are extracted from the new user comfort feedback data using a graph neural network. These new features are then fused with the updated dynamic environmental protection weight factors (values ​​updated based on the new original dataset through a gated recurrent unit network) through a cross-modal attention mechanism to generate the new unified state vector. This vector serves as a new input for the policy network's decision-making, ensuring that the policy network's decisions are based on the latest system state after device execution, thus improving the timeliness and accuracy of the decisions. Similarly, based on the new original dataset, the real-time carbon intensity change gradient of the new power grid, the real-time electricity purchase power of the new building, the output of new local renewable energy, and the overall average comfort level of the new building can be calculated using the same technical logic as in the above embodiments. The real-time carbon intensity change gradient of the new power grid reflects the trend of change in the grid's carbon intensity after the execution of control actions. The real-time electricity purchase power of the new building reflects the actual energy consumption scale of the building after the execution of the actions. New local renewable energy output can reflect changes in clean energy utilization after implementation. Meanwhile, the overall average comfort level of new buildings can quantify changes in comfort experience after implementation.

[0120] Furthermore, historical optimal scheduling results can be obtained. These historical optimal scheduling results refer to scheduling schemes that have been verified to have the best overall energy management effect (economic, environmental, and comfort) under similar or identical past scenarios. For example, the optimal air conditioning start-stop combination and power settings when carbon intensity is high during a certain period. Subsequently, using the real-time carbon intensity gradient of the new power grid, the real-time electricity purchase power of new buildings, the output of new local renewable energy, and the overall average comfort level of new buildings as inputs, and the historical optimal scheduling results as labels, the parameters of the gated cyclic unit network can be fine-tuned using a mini-batch gradient descent method. This involves adjusting the weights, biases, and other parameters in the network based on new scenario data, while retaining the network's generalization ability across different scenarios and quickly adapting to new control actions and indicator changes. The updated gated cyclic unit network parameters allow for more accurate calculation of subsequent dynamic environmental protection weight factors that match current power grid carbon intensity fluctuations, building energy load, renewable energy output, and user comfort needs, providing a more reliable environmental guidance for multi-objective optimization.

[0121] Based on the same inventive concept, such as Figure 3 As shown in the embodiments of this application, a building energy management and optimization design system 300 is also provided for implementing the building energy management and optimization design method involved above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more building energy management and optimization design system embodiments provided below can be found in the limitations of the method embodiments above, and will not be repeated here. The system includes:

[0122] The raw data processing module 301 is used to acquire the raw dataset of the target building. Based on the raw dataset, it performs calculations through a preset gated cyclic unit network to obtain dynamic environmental protection weight factors. The raw dataset includes real-time carbon intensity of the power grid, time-of-use electricity price, equipment operating parameters, environmental sensor readings, and user comfort feedback data.

[0123] The feature fusion driving module 302 is used to extract features from the equipment operating parameters, environmental sensor readings and user comfort feedback data in the original dataset to obtain the corresponding features. The features are then fused with the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector.

[0124] The strategy decision execution module 303 is used to input a unified state vector into a preset strategy network for decision reasoning processing and output a set of device control actions.

[0125] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the building energy management and optimization design method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0126] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the building energy management and optimization design method of the present application. The computer-readable storage medium may include: a read-only memory, a random access memory (RAM), a solid-state drive (SSD), or an optical disc, etc.

[0127] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A building energy management and optimized design method, characterized in that, The method comprises: obtaining an original data set of a target building, based on the original data set, performing calculation processing through a preset gated recurrent unit network to obtain a dynamic environmental protection weight factor, the original data set including real-time carbon intensity of a power grid, time-of-use electricity price, equipment operation parameters, environmental sensor readings and user comfort feedback data; performing feature extraction processing on the equipment operation parameters, environmental sensor readings and user comfort feedback data in the original data set respectively to obtain corresponding features, fusing the features and the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector; inputting the unified state vector into a preset strategy network for decision reasoning processing to output a device control action set.

2. The method of claim 1, wherein, Based on the original data set, the dynamic environmental protection weight factor is obtained by performing calculation processing through a preset gated recurrent unit network, which comprises: performing gradient calculation processing based on the time series data of the real-time carbon intensity of the power grid to obtain the historical change gradient of the real-time carbon intensity of the power grid; performing power calculation processing based on the equipment operation parameters in the original data set to obtain real-time power purchase power of the building; collecting the local renewable energy power generation equipment operation parameters of the target building, and performing output calculation processing based on the local renewable energy power generation equipment operation parameters to obtain local renewable energy output; performing analysis processing on the user comfort feedback data in the original data set to obtain a spatialized comfort degree matrix, and performing mean calculation processing based on the spatialized comfort degree matrix to obtain an overall average comfort degree of the building; inputting the real-time carbon intensity of the power grid, the historical change gradient of the real-time carbon intensity of the power grid, the real-time power purchase power of the building, the local renewable energy output and the overall average comfort degree of the building into the preset gated recurrent unit network, and performing operation processing through an activation function to output the dynamic environmental protection weight factor.

3. The method of claim 1, wherein, The feature extraction processing on the equipment operation parameters, environmental sensor readings and user comfort feedback data in the original data set respectively to obtain corresponding features, fusing the features and the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector, comprises: performing feature extraction processing on the time series data corresponding to the equipment operation parameters in the original data set through a one-dimensional convolution network to obtain an equipment operation feature vector; performing feature extraction processing on the environmental sensor readings in the original data set through an encoder network to obtain an environmental feature vector; performing feature extraction processing on the user comfort feedback data in the original data set using a graph neural network to obtain a user comfort feature vector; taking the equipment operation feature vector as a query reference and taking the environmental feature vector and the user comfort feature vector as a key-value pair, fusing through the cross-modal attention mechanism to generate a fused feature; constructing a trainable weight matrix, and performing dot multiplication operation processing on the dynamic environmental protection weight factor and the trainable weight matrix to obtain a dot multiplication operation result; performing element-wise product operation processing on the dot multiplication operation result and the fused feature to generate the unified state vector.

4. The method of claim 1, wherein, The preset policy network is a proximal policy optimization network, which is constructed based on a policy gradient algorithm and includes a policy network module and a value network module; The decision-making inference processing of the unified state vector in the preset policy network includes: Based on the time-of-use electricity price and the device operating parameters in the original data set, the cost accounting processing is performed, and the current energy cost is obtained by multiplying the real-time power in the device operating parameters by the time-of-use electricity price of the corresponding period and then summing up; According to the real-time carbon intensity of the power grid and the device operating parameters in the original data set, the emission accounting processing is performed, and the current carbon emission is obtained by multiplying the power consumption in the device operating parameters by the real-time carbon intensity of the power grid and then summing up; Based on the user comfort feedback data in the original data set, the overall average comfort of the building is calculated; An energy cost weight is obtained, and an economic reward term is constructed according to the energy cost weight and the current energy cost; a carbon emission weight is obtained, and an environmental protection reward term is constructed according to the dynamic environmental protection weight factor, the carbon emission weight, and the negative current carbon emission; a comfort weight is obtained, and a comfort reward term is obtained by multiplying the comfort weight and the overall average comfort of the building; The economic reward term, the environmental protection reward term, and the comfort reward term are added to generate a reward value; The unified state vector is input into the proximal policy optimization network, and the policy network module and the value network module are processed respectively to output the corresponding initial action distribution and state value; The initial action distribution is optimized by combining the reward value and the state value to obtain an optimized action distribution, and the device control action set is output according to the optimized action distribution.

5. The method of claim 2, wherein, The analysis processing of the user comfort feedback data in the original data set obtains a spatialized comfort matrix, which includes: The feedback text in the user comfort feedback data is analyzed by a natural language processing model to obtain a sentiment polarity score; The physiological indicators in the user comfort feedback data are mapped to comfort indexes by a machine learning model; the physiological indicators include body temperature and heart rate; The indoor positioning data of the user is obtained, and the sentiment polarity score and the comfort index are associated to the building partition of the target building based on the indoor positioning data to obtain an association result; The spatialized comfort matrix is constructed based on the association result.

6. The method of claim 1, wherein, The method further includes: According to the device control action set, an execution instruction is generated and sent to the corresponding device of the target building, and the execution instruction is used to instruct the device to perform corresponding operations according to the device control action set; After the device executes the device control action set, multi-source data is collected to obtain a new original data set, which includes new real-time carbon intensity of the power grid, new time-of-use electricity price, new device actual operating parameters, new environmental sensor readings, and new user comfort feedback data; According to the new original data set, feature extraction processing is performed to obtain a new unified state vector according to the corresponding new features; Based on the new original data set, the change gradient of the new power grid real-time carbon intensity, the new building real-time power purchase power, the new local renewable energy output and the new building overall average comfort level are calculated; The historical optimal scheduling result is obtained, the change gradient of the new power grid real-time carbon intensity, the new building real-time power purchase power, the new local renewable energy output and the new building overall average comfort level are taken as inputs, and the historical optimal scheduling result is taken as a label. The parameters of the preset Gated Recurrent Unit network are fine-tuned to obtain updated Gated Recurrent Unit network parameters.

7. The method of claim 5, wherein, The spatialized comfort matrix is constructed by the following formula: wherein, is a spatialized comfort matrix at time t, with dimension , is the number of building zones of the target building, is the total number of users participating in comfort feedback at time t, is the positioning weight of the i-th user, used to represent the importance of the building zone where the user is located, is the sentiment polarity score of the i-th user at time t, is the comfort index of the i-th user at time t.

8. A building energy management and optimized design system, characterized by, The system comprises: An original data processing module is configured to obtain an original data set of a target building, and perform calculation and processing on the original data set based on a preset Gated Recurrent Unit network to obtain a dynamic environmental protection weight factor. The original data set includes power grid real-time carbon intensity, time-of-use electricity price, equipment operation parameters, environmental sensor readings and user comfort feedback data. A feature fusion driving module is configured to perform feature extraction processing on the equipment operation parameters, environmental sensor readings and user comfort feedback data in the original data set respectively to obtain corresponding features, and perform fusion processing on the features and the dynamic environmental protection weight factor through a cross-modal attention mechanism to generate a unified state vector. A strategy decision execution module is configured to input the unified state vector into a preset strategy network for decision reasoning processing, and output a device control action set. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1 to 7.

Images