Intelligent campus comprehensive energy-saving regulation and control method and system based on AI decision

Abstract

The invention relates to a smart campus comprehensive energy-saving regulation and control method and system based on AI decision making. The method comprises the following steps: constructing a spatial-temporal feature tensor based on campus multi-source heterogeneous data; according to the spatio-temporal feature tensor, performing prediction processing through a dynamic spatio-temporal diagram neural network, and generating a global state tensor including an energy consumption prediction baseline and an environment change gradient; inputting the global state tensor into a hierarchical reinforcement learning model for processing, and generating an equipment-level control instruction set; and inputting the equipment-level control instruction set into the digital twin environment model for simulation safety verification to obtain a verification result, and generating an equipment control instruction and a strategy interpretable report which pass verification according to the verification result. According to the method, by fusing multi-source heterogeneous data and combining a dynamic space-time diagram neural network and a hierarchical reinforcement learning model, the energy consumption prediction precision and the scientificity of regulation and control decision are improved, and by means of simulation verification of a digital twin environment model, the safety of an equipment control instruction and the interpretability of a strategy are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a smart campus integrated energy-saving control method and system based on AI decision-making. Background Technology

[0002] With the deepening of educational informatization in my country, smart campuses have become an important direction for campus development. Energy consumption in campus buildings, such as HVAC, lighting, and teaching equipment, constitutes the core energy consumption of campus operations, and its energy-saving control effect directly relates to the control of campus operating costs and the achievement of green and low-carbon development goals. Currently, data-driven AI technology has been gradually introduced into the field of energy-saving control in smart campuses, but many shortcomings still need to be addressed. On the one hand, campus energy consumption is influenced by complex factors, involving multi-source heterogeneous data such as energy consumption data of campus buildings and indoor and outdoor temperature and humidity data. Existing energy management technologies typically only process single types of data, failing to integrate real-time data, static data, and event data, resulting in the neglect of spatiotemporal correlation information contained in the data, thus limiting the accuracy of subsequent predictions. On the other hand, existing energy consumption prediction models mostly use traditional machine learning algorithms such as support vector machines, failing to combine the inherent electrical and thermal conduction correlations between campus buildings to construct dynamic spatiotemporal graph neural networks. This makes it difficult to accurately output a global state tensor containing the energy consumption prediction baseline and environmental change gradient, thus failing to provide comprehensive state support for subsequent control decisions. Furthermore, the decision-making process often employs a single control strategy, failing to achieve coordination between high-level global constraints and low-level equipment control through a hierarchical reinforcement learning model. Consequently, the generated equipment-level control instruction sets often suffer from prioritizing energy conservation over comfort or focusing on local aspects over global ones. Summary of the Invention

[0003] Therefore, it is necessary to provide a smart campus integrated energy-saving control method and system based on AI decision-making to address the above-mentioned technical problems, aiming to improve the accuracy of campus energy consumption prediction and the scientific nature of control decisions.

[0004] Firstly, this application provides a comprehensive energy-saving control method for smart campuses based on AI decision-making, including:

[0005] A spatiotemporal feature tensor is constructed based on multi-source heterogeneous data from the campus. Based on the spatiotemporal feature tensor, a global state tensor containing the energy consumption prediction baseline and the environmental change gradient is generated through prediction processing using a dynamic spatiotemporal graph neural network.

[0006] The global state tensor is input into a hierarchical reinforcement learning model for processing to generate a device-level control instruction set.

[0007] The device-level control instruction set is input into the digital twin environment model for simulation safety verification to obtain the verification results. Based on the verification results, an interpretable report of the verified device control instructions and strategies is generated.

[0008] In one embodiment, a spatiotemporal feature tensor is constructed based on multi-source heterogeneous data from the campus, including:

[0009] Acquire multi-source heterogeneous data from the campus, including real-time data, static data, and event data. Real-time data includes energy consumption data of campus buildings, indoor and outdoor temperature and humidity data, indoor lighting data, and equipment status data. Static data includes a database of thermal parameters of campus buildings, a topology diagram of the campus power distribution network, and building location information. Event data includes course schedules, access control card swipe statistics, and weather forecast data.

[0010] Real-time data and event data are unified to a preset time granularity, and real-time data, static data and event data are spatially aligned to obtain spatiotemporally aligned data.

[0011] Based on the campus power distribution network topology map in the spatiotemporally aligned data, an electrical correlation matrix between buildings is constructed; based on the campus building thermal parameter library in the spatiotemporally aligned data, a heat conduction weight matrix between buildings is constructed.

[0012] The course schedule is parsed and processed to generate a time series matrix of building reservation occupancy rates for future preset time periods; combined with access control card swipe statistics and the time series matrix of building reservation occupancy rates, a real-time pedestrian flow correction factor is generated, and a solar radiation heat gain coefficient is generated based on weather forecast data and building location information.

[0013] The energy consumption sequence of historical buildings is obtained, and tensor splicing is performed on the electrical correlation matrix between buildings, the heat conduction weight matrix between buildings, the time series matrix of building reservation occupancy rate, the real-time pedestrian flow correction factor, the solar radiation heat gain coefficient, and the energy consumption sequence of historical buildings to construct a spatiotemporal feature tensor.

[0014] In one embodiment, based on the spatiotemporal feature tensor, a dynamic spatiotemporal graph neural network is used for prediction processing to generate a global state tensor containing an energy consumption prediction baseline and environmental change gradients, including:

[0015] Using building units as nodes, the building reservation occupancy rate time series matrix, real-time pedestrian flow correction factor, solar radiation heat gain coefficient and historical building sub-energy consumption sequence in the spatiotemporal feature tensor are extracted as the initial features of the nodes to obtain the node feature set.

[0016] Based on preset weighting coefficients, the electrical correlation matrix between buildings and the heat conduction weighting matrix between buildings in the spatiotemporal feature tensor are weighted and summed to generate static adjacency relationships.

[0017] The correlation coefficient between the real-time pedestrian flow correction factor in the spatiotemporal feature tensor and the time series matrix of building reservation occupancy rate is calculated to generate dynamic edge weights. The dynamic edge weights are used to characterize the dynamic interaction intensity of different building units at different time periods.

[0018] Static adjacency relationships and dynamic edge weights are fused together to generate a spatiotemporal graph adjacency matrix.

[0019] The node feature set and the spatiotemporal graph adjacency matrix are input into the spatial convolution module of the dynamic spatiotemporal graph neural network for convolution processing to obtain spatial features; the spatial features and the historical building energy consumption sequence in the spatiotemporal feature tensor are input into the temporal convolution module of the dynamic spatiotemporal graph neural network for gated recurrent unit operation processing to generate node state features.

[0020] The node state features are respectively input into two parallel fully connected prediction layers of the dynamic spatiotemporal graph neural network for prediction processing. The energy consumption prediction baseline for the future preset period is obtained through the first fully connected prediction layer, and the environmental change gradient for the future preset period is obtained through the second fully connected prediction layer.

[0021] Tensor integration processing is performed on the energy consumption prediction baseline and environmental change gradient to generate a global state tensor.

[0022] In one embodiment, the global state tensor is input into a hierarchical reinforcement learning model and processed to generate a device-level control instruction set, including:

[0023] The total energy consumption prediction and the maximum HVAC energy consumption prediction are extracted from the global state tensor, and the high-rise strategy input state is constructed by combining the preset electricity price stage information and the preset global comfort index.

[0024] The high-level policy input state is input into the high-level policy network of the hierarchical reinforcement learning model for inference, generating a regional constraint vector. The regional constraint vector includes the preset maximum power deviation and adjustable range of temperature setpoint for each building unit.

[0025] Extract global state tensor slices corresponding to each building unit from the global state tensor, and concatenate the constraint information of the corresponding building unit in the regional constraint vector with the global state tensor slices to generate the underlying policy input state.

[0026] The underlying policy input state is processed by the underlying policy network of the hierarchical reinforcement learning model to generate initial device control parameters, including temperature setpoint adjustment value, lighting brightness adjustment coefficient and socket power supply duration.

[0027] A simplified simulation model is constructed based on the campus building thermal parameter library. The initial equipment control parameters are input into the simplified simulation model for simulation. The simulated energy consumption data, simulated environmental temperature and humidity data, and simulated comfort data are obtained and integrated to obtain the execution effect data of the initial equipment control parameters.

[0028] A collaborative optimization reward function is adopted, and the reward value is calculated based on the execution effect data. The gradient descent method is used to update the parameters of the high-level policy network and the low-level policy network using the reward value, so as to obtain the optimized hierarchical reinforcement learning model.

[0029] The high-level policy input state is used as input to optimize the hierarchical reinforcement learning model, which is then reprocessed for inference to output optimized device control parameters and generate a device-level control instruction set.

[0030] In one embodiment, the device-level control instruction set is input into a digital twin environment model for simulation security verification to obtain verification results. Based on the verification results, an interpretable report of the verified device control instructions and strategies is generated, including:

[0031] Based on the campus building thermal parameter library, a simplified building thermodynamic model is constructed. The simplified building thermodynamic model and the inter-building electrical correlation matrix in the spatiotemporal feature tensor are embedded into the digital twin environment model for initialization, resulting in the initialized digital twin environment model.

[0032] The device-level control instruction set is input into the initialized digital twin environment model for simulation to obtain state evolution prediction results. The state evolution prediction results include simulation energy consumption data, simulation environment temperature and humidity data, and simulation environment change gradient.

[0033] Set a safety constraint threshold, compare the state evolution prediction result with the safety constraint threshold to obtain the verification result. If the verification result is that the safety constraint is met, the corresponding device control command is used as the verified device control command; if the verification result is that the safety constraint is violated, the safety circuit breaker mechanism is triggered and the preset backup control command is called as the verified device control command.

[0034] Using the energy consumption prediction baseline as the target energy consumption prediction result, the influence weight of each input feature in the spatiotemporal feature tensor on the target energy consumption prediction result is calculated; based on the influence weight, input features with influence weights greater than a preset threshold are selected, and a list of key influence factors is generated.

[0035] The element values ​​in the adjacency matrix of the spatiotemporal graph are extracted as cross-building association strength data, which is used to characterize the degree of association between different building units.

[0036] The list of key influencing factors is matched with cross-building correlation strength data to generate a strategy-interpretable report.

[0037] In one embodiment, the method further includes:

[0038] Based on the actual energy consumption data and actual environmental monitoring data after the actual execution of the verified device control commands, the actual energy consumption data, actual environmental monitoring data and the energy consumption prediction baseline and environmental change gradient in the global state tensor are compared to generate model error samples.

[0039] Gradient attribution results are extracted from the policy interpretability report. Based on the influence weights of each feature in the gradient attribution results, a feature importance weight matrix is ​​constructed. The model error samples and the feature importance weight matrix are weighted and fused. The parameters of the dynamic spatiotemporal graph neural network are updated using the backpropagation algorithm to obtain the initial updated dynamic spatiotemporal graph neural network.

[0040] Extract the maximum power deviation threshold from the region-level constraint vector, and identify the samples whose error values ​​exceed the maximum power deviation threshold in the model error samples as boundary samples; inject the boundary samples into the experience pool of the optimized hierarchical reinforcement learning model for update processing to obtain the updated experience pool.

[0041] Extract the heat conduction weight matrix between buildings from the spatiotemporal feature tensor, calculate the cosine similarity of the heat conduction weight matrix between different building units as the building topological similarity; set a similarity threshold, and retrieve building units with a building topological similarity greater than the similarity threshold as topologically similar building units;

[0042] Extract the node parameters of the initial update dynamic spatiotemporal graph neural network corresponding to the topologically similar building unit, and transfer them to the nodes of the initial update dynamic spatiotemporal graph neural network of the target building unit to obtain the initialized target building node parameters; adjust the temporal convolution module parameters of the initial update dynamic spatiotemporal graph neural network corresponding to the initialized target building node parameters to obtain the optimized dynamic spatiotemporal graph neural network.

[0043] In one embodiment, the dynamic edge weights are calculated using the following formula:

[0044]

[0045] in, For the first Time step building unit With building unit Dynamic edge weights between them, range of values , The time-series correlation weighting coefficient. The Pearson correlation coefficient function is used to calculate building units. With building unit In the timing window Correlation of pedestrian flow sequences within the region The length of the time series sequence. and Building units and building units In the timing window The time-series subsequence of pedestrian traffic within the area, This represents the total number of building units on campus. For the traversal index of building units, For building units With building unit physical center distance, The spatial distance attenuation threshold, For time series smoothing coefficients, , This represents the rate of change in time-series correlation.

[0046] Secondly, this application also provides a smart campus integrated energy-saving control system based on AI decision-making, including:

[0047] The spatiotemporal feature tensor construction module is used to construct spatiotemporal feature tensors based on multi-source heterogeneous data from the campus; based on the spatiotemporal feature tensors, a dynamic spatiotemporal graph neural network is used for prediction processing to generate a global state tensor containing the energy consumption prediction baseline and the environmental change gradient.

[0048] The global state tensor prediction module is used to input the global state tensor into the hierarchical reinforcement learning model for processing and to generate a set of device-level control instructions.

[0049] The instruction generation and security verification module is used to input the device-level control instruction set into the digital twin environment model for simulation security verification to obtain the verification results. Based on the verification results, it generates an interpretable report of the verified device control instructions and strategies.

[0050] 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.

[0051] 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.

[0052] The aforementioned building energy management and optimization design methods, systems, equipment, and media first construct a spatiotemporal feature tensor based on multi-source heterogeneous campus data, improving the completeness of data fusion and the foundation of feature representation. Secondly, a global state tensor is generated from this spatiotemporal feature tensor using a dynamic spatiotemporal graph neural network, addressing the problems of traditional prediction models neglecting the electrical and thermal conduction correlations between buildings and having a single prediction dimension, thus improving the accuracy of energy consumption prediction baselines and environmental change gradients. Furthermore, the global state tensor is input into a hierarchical reinforcement learning model to generate a set of device-level control instructions, enhancing the synergy and rationality of control decisions. Finally, the results are verified and generated through a digital twin environment model, solving the problems of insufficient security and lack of strategy interpretability in traditional simple simulation verification, thus improving the reliability of control instructions and the traceability of model iterations. Attached Figure Description

[0053] 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.

[0054] Figure 1 A flowchart of an AI-based decision-making-based smart campus integrated energy-saving control method is provided as an exemplary embodiment of the present invention.

[0055] Figure 2 A flowchart illustrating a method for constructing a spatiotemporal feature tensor, as provided in an exemplary embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the structure of an AI-based smart campus integrated energy-saving control system, which is provided as an exemplary embodiment of the present invention. Detailed Implementation

[0057] 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.

[0058] In one embodiment, such as Figure 1 As shown, a smart campus integrated energy-saving control method based on AI decision-making is provided. This embodiment illustrates the method applied to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0059] S101: Construct a spatiotemporal feature tensor based on multi-source heterogeneous data from the campus; based on the spatiotemporal feature tensor, perform prediction processing through a dynamic spatiotemporal graph neural network to generate a global state tensor containing the energy consumption prediction baseline and the environmental change gradient.

[0060] Specifically, multi-source data can be collected through sensor networks and management systems deployed on campus. Then, real-time data and event data from different collection frequencies are unified to the same time using timestamp alignment. Static and dynamic data can be spatially matched based on campus building coordinates. Finally, multi-dimensional data is integrated into a unified spatiotemporal feature tensor through tensor concatenation. This tensor can completely preserve the spatial-temporal-feature three-dimensional relationships of the data. After the spatiotemporal feature tensor is constructed, prediction processing can be performed using a Dynamic Spatio-Temporal Graph Neural Network (DST-GNN). DST-GNN is a deep learning model designed for spatiotemporal data, effectively handling spatiotemporal dependencies in the data. This network constructs a dynamic graph structure, representing the electrical and thermal conduction relationships between campus buildings as edges of the graph, while the energy consumption data and related environmental parameters of each building serve as node features of the graph. During the prediction process, DST-GNN utilizes graph convolution operations and time series modeling techniques to process the spatiotemporal feature tensor layer by layer, generating a global state tensor that includes the energy consumption prediction baseline and the environmental change gradient. The energy consumption prediction baseline reflects the expected energy consumption level of each building on campus under current environmental conditions, while the environmental change gradient describes the sensitivity of energy consumption to changes in environmental parameters. This global state tensor provides comprehensive energy consumption status information for subsequent control decisions, enabling subsequent control strategies to more accurately adapt to the actual operating conditions of the campus.

[0061] S102: Input the global state tensor into the hierarchical reinforcement learning model for processing to generate a device-level control instruction set.

[0062] Specifically, the hierarchical reinforcement learning model is an intelligent algorithm capable of effectively handling complex decision-making problems. It decomposes the decision-making process into multiple levels, thereby achieving collaborative optimization between high-level global constraints and low-level device control. Illustratively, the high-level strategy of the hierarchical reinforcement learning model is responsible for formulating global energy-saving goals and constraints, such as the overall energy consumption reduction ratio and equipment operating time constraints. Its low-level strategies, guided by the high-level strategy, generate specific device-level control instruction sets, including equipment on / off states and operating power adjustments. This hierarchical reinforcement learning model can continuously optimize its strategies through interactive learning with the environment, ensuring that the generated device-level control instruction sets can meet global energy-saving goals while also considering equipment operational stability and user comfort.

[0063] S103: Input the device-level control instruction set into the digital twin environment model for simulation safety verification to obtain the verification results. Based on the verification results, generate an interpretable report of the verified device control instructions and strategies.

[0064] Specifically, a digital twin environment model is a technology based on virtual modeling and simulation that can simulate the actual operating environment of a campus with high precision. By applying device-level control command sets to virtual devices in the digital twin model, the safety and effectiveness of control commands in actual operation can be verified in advance. The simulation safety verification process can include monitoring device operating status, assessing energy consumption changes, and providing early warnings of potential safety risks. If the verification results indicate that the control commands have safety hazards or may cause device failure, the control strategy can be automatically adjusted, and the device control commands can be regenerated. Only after successful verification can the final device control commands be issued to the actual devices. Furthermore, by extracting key intermediate results from the prediction and decision-making models, such as the main characteristics affecting energy consumption and the specific impact weights of high-level constraints on low-level decisions, an interpretable strategy report can be obtained. This report can then present the logic of command generation in a structured form, providing transparency support for management decisions.

[0065] The above method first overcomes the shortcomings of traditional methods that only handle single data types and thus lose spatiotemporal correlation information by constructing a spatiotemporal feature tensor, achieving effective fusion of multi-source heterogeneous data. Second, it utilizes a dynamic spatiotemporal graph neural network to generate a global state tensor, overcoming the deficiency of traditional models that do not incorporate the correlation between electrical and thermal conduction, thereby improving the accuracy of energy consumption prediction. Next, the global state tensor is input into a hierarchical reinforcement learning model to generate a device-level control instruction set, solving the problems of energy saving versus comfort and local versus global imbalances caused by single control strategies, and enhancing the scientific nature of control decisions. Finally, simulation verification is performed using a digital twin environment model, generating an interpretable report, avoiding security risks to device control instructions, and improving the interpretability and reliability of the strategy.

[0066] In one embodiment, such as Figure 2 As shown, a spatiotemporal feature tensor is constructed based on multi-source heterogeneous campus data, including:

[0067] S201: Acquire multi-source heterogeneous data on campus. Multi-source heterogeneous data on campus includes real-time data, static data, and event data. Real-time data includes campus building energy consumption data, indoor and outdoor temperature and humidity data, indoor lighting data, and equipment status data. Static data includes campus building thermal parameter database, campus power distribution network topology diagram, and building location information. Event data includes course schedules, access control card swipe statistics, and weather forecast data.

[0068] S202: Unify real-time data and event data to a preset time granularity, and spatially align real-time data, static data, and event data to obtain spatiotemporally aligned data; construct an electrical correlation matrix between buildings based on the campus power distribution network topology map in the spatiotemporally aligned data; construct a heat conduction weight matrix between buildings based on the campus building thermal parameter library in the spatiotemporally aligned data.

[0069] S203: Parse and process the course schedule to generate a time series matrix of building reservation occupancy rate for future preset time periods; combine access control card swipe statistics with the time series matrix of building reservation occupancy rate to generate a real-time pedestrian flow correction factor, and generate a solar radiation heat gain coefficient based on weather forecast data and building location information.

[0070] S204: Obtain the historical building energy consumption sequence, and perform tensor splicing processing on the electrical correlation matrix between buildings, the heat conduction weight matrix between buildings, the time series matrix of building reservation occupancy rate, the real-time pedestrian flow correction factor, the solar radiation heat gain coefficient, and the historical building energy consumption sequence to construct a spatiotemporal feature tensor.

[0071] Specifically, real-time data can be collected through a distributed sensor network and equipment monitoring system deployed on campus. For example, indoor and outdoor temperature and humidity data are collected by temperature and humidity sensors deployed inside and outside buildings, while equipment status data, such as air conditioning start / stop and lighting switch status, are obtained through the building automation system. Static data can be extracted from the campus infrastructure management system and equipment archive. The campus building thermal parameter database can contain parameters such as the heat transfer coefficient of the building envelope, window-to-wall ratio, and insulation layer thickness for each building. The campus power distribution network topology map can be exported from the power management system, thus clarifying the connection relationships and impedance parameters of power lines between buildings. Building location information can be obtained through the campus GIS system. Event data can be obtained through the campus management system. For example, course schedules are exported from the academic affairs management system, containing information such as course name, classroom, start and end time, etc. Access control card swipe statistics are extracted from the access control management system, recording the building and time of the card swipe. Weather forecast data can include hourly solar radiation intensity and outdoor temperature and humidity data for the next 24 hours.

[0072] Furthermore, after data collection, the temporal and spatial misalignment of multi-source data can directly lead to the failure of feature association. Therefore, a benchmark timestamp alignment method can be used for alignment processing. For example, with a preset time granularity of 15 minutes, real-time data with a higher granularity is aggregated to a 15-minute granularity through statistical averaging, while event data with a lower granularity, such as hourly timetables, is extended to a 15-minute granularity through time interval mapping. A unified format timestamp is added to all data. Additionally, a unique ID can be assigned to each campus building, binding static data such as building thermal parameters to this ID. Real-time data and event data collection terminals, such as sensors and access control machines, can be associated with the corresponding building ID, thereby enabling spatial association of the three types of data. The final spatiotemporally aligned data output can accurately reflect the multi-dimensional state information of a building at a specific time granularity. Moreover, since campus buildings are not isolated, the electrical load conduction and heat energy exchange between adjacent buildings directly affect their respective energy consumption levels. Therefore, an electrical association matrix between buildings can be constructed based on the campus power distribution network topology. Each building is represented as a row and column of the matrix (the matrix dimension is N×N, where N is the total number of buildings). The electrical association weights between buildings are calculated using the line parameters of the power distribution network. If building i and building j are directly connected by a power line, the weight is set to the reciprocal of the line impedance, where a smaller impedance indicates a stronger association. If they are indirectly connected, the weight is the reciprocal of the minimum impedance along the path; otherwise, it is set to 0. Furthermore, a heat conduction weight matrix between buildings can be constructed using a database of campus building thermal parameters. The elements of this matrix can be calculated based on Fourier's law of heat conduction using w=k / d, where w is the weight value, k is the equivalent heat transfer coefficient between building i and building j (calculated from the heat transfer coefficients of the building envelopes and the contact area), and d is the center-to-center distance between the two buildings. This matrix can accurately quantify the differences in heat conduction intensity between different buildings, providing crucial support for subsequent models to capture spatial thermal coupling effects.

[0073] Specifically, the course schedule can be used to extract the classroom-start and end time information, mapping the time intervals to preset time granularities. If a building has a usage plan within a certain time granularity, the occupancy rate is set to 1.0; otherwise, it is set to 0.0. For multi-functional buildings such as auditoriums, data from the event reservation system can be used to supplement this, ultimately forming an N×T dimension time-series matrix of building reservation occupancy rates for future preset time periods (N is the total number of buildings, and T is the number of time steps). This matrix reflects the distribution of people reservations for each building within the future preset time period. Furthermore, a real-time traffic correction factor can be calculated to correct the deviation between the reserved and actual values. This factor is obtained by calculating the actual number of people entering each building within a certain time granularity using access control card swipe statistics and then comparing it with the theoretical number of people corresponding to the same period's reservation occupancy rate. It is also possible to obtain the hourly solar radiation intensity for a future preset period, and calculate the radiation receiving coefficient by combining the building's latitude, longitude, orientation and tilt angle (the receiving coefficient for south-facing buildings is set to 0.9, and for north-facing buildings it is set to 0.3). The solar radiation heat gain coefficient is calculated by multiplying the radiation intensity by the receiving coefficient by the building's daylighting area. This coefficient can accurately quantify the differences in the solar radiation impact of different buildings at different times, providing key input for environmental factor modeling in subsequent energy consumption prediction. Subsequently, the energy consumption sequence of historical buildings can be obtained first. For example, the energy consumption sequence of historical buildings over the past 30 days can be selected and resampled to a T×N dimension according to a preset time granularity. All features are then unified to a three-dimensional dimension of building-time-feature. For instance, the electrical correlation matrix between buildings (N×N) and the heat conduction weight matrix (N×N) can be expanded to N×N×1, the building reservation occupancy rate time series matrix (N×T) can be expanded to N×T×1, the real-time pedestrian flow correction factor (N×T) can be expanded to N×T×1, the solar radiation heat gain coefficient (N×T) can be expanded to N×T×1, and the energy consumption sequence of historical buildings (N×T) can be expanded to N×T×1. After unifying the dimensions, all features can be integrated into a spatiotemporal feature tensor of N×T×M dimensions through tensor splicing operations, where M is the total number of features.

[0074] In one embodiment, based on the spatiotemporal feature tensor, a dynamic spatiotemporal graph neural network is used for prediction processing to generate a global state tensor containing an energy consumption prediction baseline and environmental change gradients, including:

[0075] Using building units as nodes, the building reservation occupancy rate time series matrix, real-time pedestrian flow correction factor, solar radiation heat gain coefficient and historical building sub-energy consumption sequence in the spatiotemporal feature tensor are extracted as the initial features of the nodes to obtain the node feature set.

[0076] Based on preset weighting coefficients, the electrical correlation matrix between buildings and the heat conduction weighting matrix between buildings in the spatiotemporal feature tensor are weighted and summed to generate static adjacency relationships.

[0077] The correlation coefficient between the real-time pedestrian flow correction factor in the spatiotemporal feature tensor and the time series matrix of building reservation occupancy rate is calculated to generate dynamic edge weights. The dynamic edge weights are used to characterize the dynamic interaction intensity of different building units at different time periods. The static adjacency relationship and the dynamic edge weights are fused to generate a spatiotemporal graph adjacency matrix.

[0078] The node feature set and the spatiotemporal graph adjacency matrix are input into the spatial convolution module of the dynamic spatiotemporal graph neural network for convolution processing to obtain spatial features; the spatial features and the historical building energy consumption sequence in the spatiotemporal feature tensor are input into the temporal convolution module of the dynamic spatiotemporal graph neural network for gated recurrent unit operation processing to generate node state features.

[0079] The node state features are input into two parallel fully connected prediction layers of the dynamic spatiotemporal graph neural network for prediction processing. The energy consumption prediction baseline for the future preset period is obtained through the first fully connected prediction layer, and the environmental change gradient for the future preset period is obtained through the second fully connected prediction layer. The energy consumption prediction baseline and the environmental change gradient are integrated by tensor processing to generate a global state tensor.

[0080] Specifically, the node features of the dynamic spatiotemporal graph neural network can be used to carry the core state information of building units. Since the energy consumption and environmental changes of campus buildings are mainly influenced by the synergistic effects of human activities, environmental impacts, and historical states, the corresponding dimensions in the spatiotemporal feature tensor can be selected as the initial node features. For example, from the building-time dimension of the spatiotemporal feature tensor, the following can be extracted: the building reservation occupancy rate time series matrix (reflecting the basic plan of human activities), the real-time pedestrian flow correction factor (correcting the deviation between the reserved value and the actual pedestrian flow), the solar radiation heat gain coefficient (quantifying environmental heat input), and the historical building sub-item energy consumption sequence (reflecting the temporal evolution of energy consumption). Then, the dimensions of each feature are unified, that is, the building reservation occupancy rate time series matrix (N×T), the real-time pedestrian flow correction factor (N×T), the solar radiation heat gain coefficient (N×T), and the historical building sub-item energy consumption sequence (N×T) are all expanded into N×T×1 three-dimensional features (N is the total number of building units, T is the number of time steps), and finally concatenated into an N×T×4 node feature set. This feature covers three core influencing factors: personnel, environment, and history, providing comprehensive information support for the node state representation of the graph neural network. Subsequently, the electrical association matrix (N×N) and heat conduction weight matrix (N×N) between buildings can be extracted from the spatiotemporal feature tensor. Preset weight coefficients are selected, such as 0.6 for electrical association and 0.4 for heat conduction. The corresponding elements of the two matrices are weighted and summed to obtain the element value in the i-th row and j-th column of the static adjacency relation matrix, thus obtaining the corresponding static adjacency relation. This static adjacency relation represents the fixed spatial association between building units and can serve as the basic edge structure of the dynamic spatiotemporal graph, ensuring that the model captures the inherent physical interactions of the buildings.

[0081] Specifically, the interaction intensity of campus buildings changes dynamically over time. For example, during breaks, the interaction of people flowing between adjacent classrooms increases. Static adjacency cannot reflect this type of change, so dynamic edge weights can be calculated based on the characteristics of people's activities. For example, firstly, the real-time pedestrian flow correction factor and the time series matrix of building reservation occupancy rate are extracted from the spatiotemporal feature tensor. Based on these two sequences, the dynamic edge weights are calculated using the following formula:

[0082]

[0083] in, For the first Time step building unit With building unit Dynamic edge weights between them The time-series correlation weighting coefficient. The Pearson correlation coefficient function is used to calculate building units. With building unit In the timing window Correlation of pedestrian flow sequences within the region The length of the time series sequence. and Building units and building units In the timing window The time-series subsequence of pedestrian traffic within the area, This represents the total number of building units on campus. For the traversal index of building units, For building units With building unit physical center distance, The spatial distance attenuation threshold, For time series smoothing coefficients, , This represents the rate of change in time-series correlation.

[0084] Calculated using the above formula The range of values ​​is This method can accurately characterize the dynamic interaction intensity between buildings i and j at time step t. Furthermore, it can fuse the static adjacency matrix and the dynamic edge weight matrix through element-wise addition to generate a spatiotemporal graph adjacency matrix. This matrix is ​​dynamically updated with each time step, preserving the inherent physical connections between buildings while incorporating dynamic interactions that change with human activity, thus overcoming the limitation of traditional static graph structures in adapting to dynamic campus operating scenarios.

[0085] Specifically, the node feature set and the spatiotemporal graph adjacency matrix are input into the spatial convolution module (using a Graph Convolutional Network (GCN) layer) of the dynamic spatiotemporal graph neural network. The spatiotemporal graph adjacency matrix is ​​first normalized, then the normalized adjacency matrix is ​​multiplied by the node feature set, and the spatial features are obtained after passing through the ReLU activation function. This process enables the transmission and fusion of spatial association information between building units, such as transmitting the energy consumption characteristics of adjacent buildings to the target unit, capturing the spatial coupling effect. Furthermore, the historical building energy consumption sequence can be extracted from the spatiotemporal feature tensor, concatenated with the spatial features, and input into the temporal convolution module (using a gated recurrent unit (GRU) layer with a hidden layer dimension of 64). The GRU gating mechanism captures the temporal evolution of the features, ultimately generating node state features that integrate spatiotemporal information.

[0086] Specifically, the node state features are input into two parallel fully connected prediction layers. The first fully connected prediction layer (input dimension 64, output dimension 1) uses linear transformation and the Sigmoid activation function to predict the energy consumption baseline for each building unit over the next 24 hours (corresponding to 96 15-minute time steps). The second fully connected prediction layer (input dimension 64, output dimension 2) predicts the temperature and humidity gradients during the same period (such as temperature and humidity changes every 15 minutes). Then, the energy consumption prediction baseline (N×96×1) and the environmental gradient (N×96×2) are dimension-matched and integrated into a global state tensor of N×96×3 through tensor concatenation. This tensor provides accurate prediction information for both energy consumption and the environment. Compared to traditional single energy consumption prediction models, it provides more comprehensive state support for subsequent control decisions, enabling decisions to simultaneously consider energy consumption control and environmental comfort.

[0087] In one embodiment, the global state tensor is input into a hierarchical reinforcement learning model for processing to generate a device-level control instruction set, including:

[0088] The total energy consumption prediction and the maximum HVAC energy consumption prediction are extracted from the global state tensor, and the high-rise strategy input state is constructed by combining the preset electricity price stage information and the preset global comfort index.

[0089] The high-level policy input state is input into the high-level policy network of the hierarchical reinforcement learning model for inference, generating a regional constraint vector. The regional constraint vector includes the preset maximum power deviation and adjustable range of temperature setpoint for each building unit.

[0090] Extract global state tensor slices corresponding to each building unit from the global state tensor, and concatenate the constraint information of the corresponding building unit in the regional constraint vector with the global state tensor slices to generate the underlying policy input state.

[0091] The underlying policy input state is processed by the underlying policy network of the hierarchical reinforcement learning model to generate initial device control parameters, including temperature setpoint adjustment value, lighting brightness adjustment coefficient and socket power supply duration.

[0092] A simplified simulation model is constructed based on the campus building thermal parameter library. The initial equipment control parameters are input into the simplified simulation model for simulation. The simulated energy consumption data, simulated environmental temperature and humidity data, and simulated comfort data are obtained and integrated to obtain the execution effect data of the initial equipment control parameters.

[0093] A collaborative optimization reward function is adopted, and the reward value is calculated based on the execution effect data. The gradient descent method is used to update the parameters of the high-level policy network and the low-level policy network using the reward value, so as to obtain the optimized hierarchical reinforcement learning model.

[0094] The high-level policy input state is used as input to optimize the hierarchical reinforcement learning model, which is then reprocessed for inference to output optimized device control parameters and generate a device-level control instruction set.

[0095] Specifically, high-level strategies can control energy-saving goals and operational constraints from a global perspective. Therefore, key features that reflect the overall state, core energy consumption, cost factors, and comfort requirements can be selected. For example, the total energy consumption prediction value (i.e., the sum of the energy consumption prediction baselines of each building unit, used to characterize the basic level of the global energy-saving goal) and the maximum HVAC energy consumption prediction value can be extracted from the global state tensor. The total energy consumption prediction value is the sum of the energy consumption prediction baselines of each building unit, used to characterize the basic level of the global energy-saving goal, and the maximum HVAC energy consumption prediction value is the maximum value of HVAC energy consumption in each building unit. Preset electricity price stage information can be obtained through the campus power grid interface, and preset global comfort indicators can be called, such as those obtained using PMV values, with a preset reasonable range of -0.5 to 0.5, characterizing the constraint boundary of human comfort. By concatenating the above features in the order of total energy consumption prediction value - maximum HVAC energy consumption prediction value - electricity price stage - global comfort indicator, a high-level strategy input state vector with a dimension of 4 can be formed. This vector not only covers energy-saving goals and energy-consuming objects, but also incorporates cost and comfort constraints, providing a comprehensive global basis for high-level decision-making.

[0096] Specifically, high-level strategies need to set operational boundaries for low-level controls to prevent low-level decisions from deviating from the global objective. The high-level strategy input states are then fed into the high-level strategy network for inference, generating regional constraint vectors. Illustratively, the high-level strategy network uses a 3-layer fully connected neural network (with a hidden layer dimension of 32 and ReLU activation function to enhance nonlinear fitting capabilities). Taking the high-level strategy input states as input, it outputs regional constraint vectors corresponding to each building unit. These vectors contain two types of constraints: first, the maximum power deviation of each building unit, such as ±10%, which represents the allowable deviation range between the building's actual energy consumption and the predicted baseline; negative values ​​correspond to tighter energy-saving requirements. Second, the adjustable range of the temperature setpoint to match the human comfort zone and the operating range of HVAC equipment. Low-level strategies can combine the local states of building units with high-level constraints to achieve refined control. For example, tensor slices corresponding to each building unit can be extracted from the global state tensor, representing local state information such as the building unit's energy consumption prediction baseline and environmental change gradient, with a dimension of 96×3, corresponding to 15-minute granular data for the next 24 hours. The constraint information (maximum power deviation, adjustable range of temperature setpoint) of the corresponding building unit in the regional constraint vector is concatenated with the tensor slice to form a bottom-level strategy input state with a dimension of 96×5. This state contains both the local operating characteristics of the building unit and the constraint boundaries set by the higher level, so that the bottom-level decision can adapt to the differences of the local scene within the global framework.

[0097] The underlying policy input state is then fed into the underlying policy network, which employs a Deep Q-Network (DQN). Using the underlying policy input state as input, the network outputs initial equipment control parameters for each building unit. These parameters may include temperature setpoint adjustment values ​​(within the temperature range constrained by high-rise buildings), lighting brightness adjustment coefficients (ranging from 0 to 1, corresponding to 0% to 100% brightness of lighting equipment), and socket power supply duration (based on the building's predetermined occupancy timeline matrix, such as setting a 1-hour delay after a class). This local perception capability of the underlying network allows for precise adaptation of equipment parameter adjustments to the real-time operating scenarios of each building.

[0098] Specifically, directly verifying control parameters on actual equipment carries safety risks and cost losses. Therefore, simplified simulations can quickly assess the effectiveness of parameter execution. For example, a simplified thermodynamic model in the form of a resistive-capacitive network can be constructed based on a campus building thermal parameter library (including parameters such as building envelope heat transfer coefficient and window-to-wall ratio), while embedding energy consumption characteristic curves for lighting and sockets. Inputting the initial equipment control parameters into this model can simulate changes in energy consumption during equipment operation (such as energy fluctuations after air conditioning temperature adjustment), the evolution of indoor environmental temperature and humidity, and changes in occupant comfort (such as real-time calculation of PMV values). By integrating the above simulation results into the execution effect data of the initial equipment control parameters, this data can intuitively reflect whether the parameters meet energy-saving targets and comfort constraints, providing a basis for subsequent network optimization. Subsequently, a reward value can be calculated and the network parameters updated using a collaborative optimization reward function. This collaborative optimization reward function can include three parts: energy-saving term, cost term, and comfort term. The energy-saving item awards points when the actual simulated energy consumption is lower than the predicted baseline; the larger the difference, the more points are awarded. In the cost item, a higher proportion of energy consumption during off-peak hours awards more points, while a high proportion of energy consumption during peak hours deducts points. The comfort item awards points when the PMV value is within a preset reasonable range, and deducts points when it exceeds the range. After substituting the execution effect data into this function to calculate the reward value, the parameters of the high-level and low-level policy networks can be updated using gradient descent to obtain an optimized hierarchical reinforcement learning model. This process allows the model to learn better decision logic through trial and error, improving the accuracy of subsequent control. Finally, the initially constructed high-level policy input state is re-input into the optimized hierarchical reinforcement learning model. Through collaborative reasoning between the high-level and low-level policy networks, optimized equipment control parameters for each building unit can be output. Integrating the parameters of all building units according to the structure of building ID-equipment type-control parameter generates a structured equipment-level control instruction set. This instruction set meets both global energy-saving and comfort requirements and is adapted to the local operating scenarios of each building, and can be directly issued to the campus building automation system for execution.

[0099] In one embodiment, the device-level control instruction set is input into a digital twin environment model for simulation security verification to obtain verification results. Based on the verification results, an interpretable report of the verified device control instructions and policies is generated, including:

[0100] Based on the campus building thermal parameter library, a simplified building thermodynamic model is constructed. The simplified building thermodynamic model and the inter-building electrical correlation matrix in the spatiotemporal feature tensor are embedded into the digital twin environment model for initialization, resulting in the initialized digital twin environment model.

[0101] The device-level control instruction set is input into the initialized digital twin environment model for simulation to obtain state evolution prediction results. The state evolution prediction results include simulation energy consumption data, simulation environment temperature and humidity data, and simulation environment change gradient.

[0102] Set a safety constraint threshold, compare the state evolution prediction result with the safety constraint threshold to obtain the verification result. If the verification result is that the safety constraint is met, the corresponding device control command is used as the verified device control command; if the verification result is that the safety constraint is violated, the safety circuit breaker mechanism is triggered and the preset backup control command is called as the verified device control command.

[0103] Using the energy consumption prediction baseline as the target energy consumption prediction result, the influence weight of each input feature in the spatiotemporal feature tensor on the target energy consumption prediction result is calculated; based on the influence weight, input features with influence weights greater than a preset threshold are selected, and a list of key influence factors is generated.

[0104] The element values ​​in the adjacency matrix of the spatiotemporal graph are extracted as cross-building association strength data, which is used to characterize the degree of association between different building units.

[0105] The list of key influencing factors is matched with cross-building correlation strength data to generate a strategy-interpretable report.

[0106] Specifically, the energy consumption and environmental evolution of campus buildings are constrained by both thermodynamic laws and electrical connections. Modeling with only a single physics field can lead to excessive simulation bias. Therefore, a multi-physics twin model can be constructed by integrating thermodynamic and electrical connection characteristics. For example, a simplified thermodynamic model of the building, namely a resistance-capacitance (RC) network model, can be built based on a library of thermal parameters of campus buildings. This model can then characterize the heat transfer characteristics of the building envelope. The walls, roof, windows, and other envelope components correspond to different resistance (heat transfer impedance) and capacitance (heat capacity) parameters, and the parameter values ​​can be directly extracted from the thermal parameter library. Furthermore, by extracting the electrical connection matrix between buildings from the spatiotemporal feature tensor and embedding it into the power distribution network module of the twin model, the connection relationships of power distribution lines, impedance parameters, and load conduction paths between each building unit can be clearly defined. In addition, during the model initialization process, a three-dimensional geometric model of the campus buildings (exported from the campus BIM system) can be imported. By aligning the thermodynamic model, electrical network module, and three-dimensional geometric model with coordinates, an initial digital twin environment model that accurately reproduces the physical structure, heat transfer characteristics, and power distribution connections of the buildings can be obtained. Subsequently, the device-level control instruction set can be parsed into an instruction format recognizable by the twin model, based on the structure of building ID-equipment type-control parameters-execution sequence. Simulation parameters can be set, with the simulation time step consistent with the previously mentioned time granularity (15 minutes / step), and the simulation duration covering the next 24 hours to ensure complete coverage of the entire instruction execution cycle. After starting the simulation, the model can calculate the heat exchange processes between the interior and exterior of each building, such as solar radiation heat gain and heat loss due to heat transfer from the building envelope, based on the thermodynamic module. It can also calculate the energy consumption conduction and load distribution of each device during operation through the electrical network module, ultimately outputting the state evolution prediction results. These results include simulation energy consumption data for each time step (statistically categorized by building), simulation environment temperature and humidity data, and simulation environment change gradients (temperature and humidity change rates per time step).

[0107] Specifically, when control commands directly act on actual equipment, parameters exceeding safety thresholds may lead to equipment overload and damage or a significant decrease in personnel comfort. Therefore, risks can be predicted and backup plans activated through threshold comparisons. For example, energy consumption safety thresholds can refer to the maximum carrying capacity of the campus power distribution system, and environmental temperature and humidity safety thresholds can be preset according to specifications, such as indoor temperatures not lower than 18℃ in winter and not higher than 28℃ in summer, with relative humidity between 40% and 60%. The environmental change gradient threshold can refer to the human body's thermal comfort adaptability. By comparing the state evolution prediction results with the above thresholds step by step and index by index: if the simulation data at all time steps does not exceed the threshold, the verification result can be considered to meet the safety constraints, and the original equipment-level control command can be directly used as the verification passed command. If any index exceeds the threshold, such as the energy consumption exceeding the power distribution threshold due to the air conditioning load of a building in the simulation, the safety circuit breaker mechanism can be triggered immediately, that is, the backup control command pre-stored in the twin model database can be called. This backup command is based on the equipment's safe operating parameters, such as adjusting the air conditioning temperature setpoint to 26°C to reduce the load and lowering the lighting brightness adjustment coefficient to 0.8. It has been verified through historical simulations to ensure safety, and can be used as a verified equipment control command after replacement.

[0108] Specifically, after security verification, impact weight calculations and a list of key impact factors can be constructed to clarify the contribution of each input feature to the energy consumption prediction results. For example, using the energy consumption prediction baseline output by the dynamic spatiotemporal graph neural network as the target energy consumption prediction result, the gradient value of the target prediction result with respect to each input feature is calculated through backpropagation. The absolute value of the gradient is standardized and used as the impact weight. The larger the weight value, the more significant the impact of the feature on the prediction result. The input features to be analyzed include key dimensions such as building reservation occupancy rate, real-time pedestrian flow correction factor, solar radiation heat gain coefficient, and historical energy consumption sequence. For example, if the calculated weight value of the solar radiation heat gain coefficient is 0.32 and the weight value of the real-time pedestrian flow correction factor is 0.28, it can be said that these two types of features are the core factors affecting energy consumption prediction. Subsequently, a preset threshold can be set to filter out input features with weight values ​​greater than the threshold, and the key impact factor list is generated by sorting them from largest to smallest weight. This list can simultaneously annotate the specific impact scenarios of each factor. For example, the solar radiation heat gain coefficient has the most significant impact between 10:00 and 14:00 in summer, thus providing a core basis for subsequent explanations of decision-making logic and making previously invisible characteristic influence relationships explicit. Furthermore, energy consumption decisions for campus buildings are not isolated; a spatiotemporal graph adjacency matrix can be constructed to quantify the spatial synergy of decisions by combining the strength of inter-building connections. For instance, the spatiotemporal graph adjacency matrix can be extracted from the modeling results of a dynamic spatiotemporal graph neural network. The element value in the i-th row and j-th column of the matrix represents the cross-building connection strength data between building i and building j. The larger this value, the closer the electrical conduction, heat exchange, or personnel flow connection between the two buildings. Subsequently, the list of key influencing factors can be matched with the correlation strength data. That is, for each key factor in the list, the building unit most significantly affected by that factor is retrieved, and then the correlation strength data of that building unit with other buildings is matched to form a three-level correlation relationship of key factor-core affected building-related building. For example, the solar radiation heat gain coefficient-Teaching-01-Teaching-02 (correlation strength 0.72) can indicate that the air conditioning control command of Teaching-01 is significantly affected by solar radiation, and the execution of this command will affect the closely related Teaching-02 through heat conduction.

[0109] Specifically, based on the above analysis, a strategy-interpretable report can be generated to transform the quantitative analysis results into decision-making basis that managers can understand. This report can be written in a structured format, including a description of validation results, analysis of key influencing factors, and explanations of cross-building correlations. The description of validation results can clearly indicate the details of validated instructions, the specific exceedances of failed instructions, and the backup instructions to replace them after a circuit breaker. The analysis of key influencing factors can elaborate on the impact weights of each core factor, the time periods of impact, and the specific control logic for equipment parameters. The explanation of cross-building correlations can visualize the strength of correlations between buildings using heatmaps, explaining how the instructions of core buildings affect surrounding buildings through correlations, and how the status of related buildings should be considered during decision-making.

[0110] In one embodiment, the method further includes:

[0111] Based on the actual energy consumption data and actual environmental monitoring data after the actual execution of the verified device control commands, the actual energy consumption data, actual environmental monitoring data and the energy consumption prediction baseline and environmental change gradient in the global state tensor are compared to generate model error samples.

[0112] Gradient attribution results are extracted from the policy interpretability report. Based on the influence weights of each feature in the gradient attribution results, a feature importance weight matrix is ​​constructed. The model error samples and the feature importance weight matrix are weighted and fused. The parameters of the dynamic spatiotemporal graph neural network are updated using the backpropagation algorithm to obtain the initial updated dynamic spatiotemporal graph neural network.

[0113] Extract the maximum power deviation threshold from the region-level constraint vector, and identify the samples whose error values ​​exceed the maximum power deviation threshold in the model error samples as boundary samples; inject the boundary samples into the experience pool of the optimized hierarchical reinforcement learning model for update processing to obtain the updated experience pool.

[0114] Extract the heat conduction weight matrix between buildings from the spatiotemporal feature tensor, calculate the cosine similarity of the heat conduction weight matrix between different building units as the building topological similarity; set a similarity threshold, and retrieve building units with a building topological similarity greater than the similarity threshold as topologically similar building units;

[0115] Extract the node parameters of the initial update dynamic spatiotemporal graph neural network corresponding to the topologically similar building unit, and transfer them to the nodes of the initial update dynamic spatiotemporal graph neural network of the target building unit to obtain the initialized target building node parameters; adjust the temporal convolution module parameters of the initial update dynamic spatiotemporal graph neural network corresponding to the initialized target building node parameters to obtain the optimized dynamic spatiotemporal graph neural network.

[0116] Specifically, data from the actual execution of verified equipment control commands can be collected through the campus sensor network and equipment monitoring system. Then, the energy consumption prediction baseline and environmental change gradient for the corresponding time period are extracted from the global state tensor generated in the above embodiment, and the error is calculated using a time-step, index-by-index difference comparison method. The energy consumption error is the absolute difference between the actual energy consumption data and the energy consumption prediction baseline, and the environmental error is the absolute difference between the actual environmental temperature and humidity and the integral result of the predicted environmental change gradient. By integrating the energy consumption error and environmental error for all time steps according to the structure of building ID-time step-error type-error value, model error samples can be generated. After the error samples are generated, the parameters of the dynamic spatiotemporal graph neural network can be updated. For example, gradient attribution results can be extracted from the policy interpretability report, which quantifies the influence weights of each input feature in the spatiotemporal feature tensor, such as the solar radiation heat gain coefficient and the real-time pedestrian flow correction factor, on the energy consumption prediction baseline. Based on these weights, a feature importance weight matrix can be constructed. The matrix dimension is consistent with the model input feature dimension, and the matrix element values ​​directly correspond to the influence weights of each feature; the higher the weight, the larger the element value. The model error samples are then weighted and fused with the weight matrix. This means that for each error sample, the error is amplified or reduced according to the weight value of the corresponding feature, ensuring that the update process prioritizes correcting model biases related to key features. Finally, the parameters of the dynamic spatiotemporal graph neural network are updated using the backpropagation algorithm, with a learning rate of 0.001 and 50 iterations, resulting in the initial updated dynamic spatiotemporal graph neural network.

[0117] Furthermore, since the robustness of reinforcement learning models depends on learning from extreme scenarios, and boundary samples correspond to "critical scenarios approaching the safety constraint threshold," they can be injected into the experience pool to improve the model's adaptability to complex scenarios. For example, the maximum power deviation threshold can be extracted from the region-level constraint vector; this threshold is the core safety boundary set by the higher-level policy. By traversing the model error samples, samples with error values ​​exceeding this threshold can be marked as boundary samples. For instance, if the actual energy consumption of a teaching building at a certain time step is 12% higher than the predicted baseline, then this sample is a boundary sample. Subsequently, the selected boundary samples are organized according to the standard format of the reinforcement learning experience pool (state-action-reward-next state) and injected into the experience pool of the optimized hierarchical reinforcement learning model. This experience pool can adopt a first-in, first-out update rule. When the pool capacity reaches 10,000 samples, the earliest stored sample is deleted to ensure the timeliness of the samples in the pool. At the same time, based on the newly injected boundary samples, the policy network of the reinforcement learning model can be lightly fine-tuned (iterated 10 times) using temporal difference learning to obtain the updated experience pool. This update process can further improve the robustness of the model in extreme scenarios through reinforcement learning of boundary samples.

[0118] Furthermore, since there are many buildings with similar structures and uses on campus, directly modeling each building would lead to repetitive work. Therefore, parameter transfer through topological similarity can significantly improve modeling efficiency and generalization. For example, the heat conduction weight matrix between buildings can be extracted from the spatiotemporal feature tensor. This matrix quantifies the heat conduction correlation between buildings, which is the core manifestation of building topological characteristics. Therefore, the cosine similarity of this matrix between different building units can be calculated as the building topological similarity. The cosine similarity value ranges from [0,1]. The closer the value is to 1, the more similar the heat conduction topology, envelope characteristics, and spatial layout of the two buildings are.

[0119] Specifically, a similarity threshold can be set based on the statistical analysis of campus building types, such as 0.8. Building units with a topological similarity greater than this threshold can then be identified as topologically similar building units. By extracting the node parameters of the initial updated dynamic spatiotemporal graph neural network corresponding to the topologically similar building unit (e.g., Teaching Building-01), including the weight matrix of the spatial convolution module and the embedding parameters of node features, these parameters can be directly transferred to the initial updated dynamic spatiotemporal graph neural network nodes of the target building unit, such as the newly built Teaching Building-04, resulting in initialized target building node parameters. Furthermore, local fine-tuning can be performed based on the characteristics of the target building. For example, since the temporal convolution module of the initial updated dynamic spatiotemporal graph neural network is more sensitive to the temporal operational characteristics of the building, the historical operational data (energy consumption and environmental data) of the target building over the past 7 days can be used as fine-tuning data. After 20 iterations, parameter adjustment is completed, resulting in the final optimized dynamic spatiotemporal graph neural network. This process, through topological transfer, enables efficient modeling of similar buildings, continuously improving the adaptability and accuracy of the entire energy-saving control system over time, solving the problem that traditional fixed models are difficult to adapt to dynamic campus operation scenarios.

[0120] Based on the same inventive concept, this application also provides an AI-based smart campus integrated energy-saving control system for implementing the AI-based smart campus integrated energy-saving control method described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more AI-based smart campus integrated energy-saving control system embodiments provided below can be found in the limitations of the AI-based smart campus integrated energy-saving control method described above, and will not be repeated here.

[0121] In one exemplary embodiment, such as Figure 3 As shown, a smart campus integrated energy-saving control system 300 based on AI decision-making is provided, including:

[0122] The spatiotemporal feature tensor construction module 301 is used to construct a spatiotemporal feature tensor based on multi-source heterogeneous data from the campus; based on the spatiotemporal feature tensor, a dynamic spatiotemporal graph neural network is used for prediction processing to generate a global state tensor containing the energy consumption prediction baseline and the environmental change gradient.

[0123] The global state tensor prediction module 302 is used to input the global state tensor into the hierarchical reinforcement learning model for processing and to generate a device-level control instruction set.

[0124] The instruction generation and security verification module 303 is used to input the device-level control instruction set into the digital twin environment model for simulation security verification to obtain the verification results. Based on the verification results, it generates a report of the verified device control instructions and strategies that can be interpreted.

[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 AI-based decision-making-based smart campus integrated energy-saving control 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 AI-based decision-making-based smart campus integrated energy-saving control method of this 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 smart campus integrated energy-saving control method based on AI decision-making, characterized in that, The method includes: A spatiotemporal feature tensor is constructed based on multi-source heterogeneous data from the campus; based on the spatiotemporal feature tensor, a global state tensor containing the energy consumption prediction baseline and the environmental change gradient is generated through a dynamic spatiotemporal graph neural network. The global state tensor is input into a hierarchical reinforcement learning model for processing to generate a device-level control instruction set. The device-level control instruction set is input into the digital twin environment model for simulation security verification to obtain the verification results. Based on the verification results, an interpretable report of the verified device control instructions and strategies is generated.

2. The method according to claim 1, characterized in that, The construction of the spatiotemporal feature tensor based on multi-source heterogeneous campus data includes: The campus multi-source heterogeneous data is acquired, which includes real-time data, static data, and event data. The real-time data includes campus building energy consumption data, indoor and outdoor temperature and humidity data, indoor lighting data, and equipment status data. The static data includes a campus building thermal parameter library, a campus power distribution network topology map, and building location information. The event data includes course schedules, access control card swipe statistics, and weather forecast data. The real-time data and the event data are unified to a preset time granularity, and the real-time data, the static data, and the event data are spatially aligned to obtain spatiotemporally aligned data; Based on the campus power distribution network topology map in the spatiotemporal aligned data, an electrical correlation matrix between buildings is constructed; based on the campus building thermal parameter library in the spatiotemporal aligned data, a heat conduction weight matrix between buildings is constructed. The course schedule is parsed to generate a time series matrix of building reservation occupancy rates for future preset time periods; combined with the access control card swipe statistics and the time series matrix of building reservation occupancy rates, a real-time pedestrian flow correction factor is generated, and a solar radiation heat gain coefficient is generated based on the weather forecast data and the building location information; The historical building energy consumption sequence is obtained, and the electrical correlation matrix between buildings, the heat conduction weight matrix between buildings, the time series matrix of the building's scheduled occupancy rate, the real-time pedestrian flow correction factor, the solar radiation heat gain coefficient, and the historical building energy consumption sequence are subjected to tensor spatiotemporal feature tensor splicing processing to construct the spatiotemporal feature tensor.

3. The method according to claim 2, characterized in that, The step of generating a global state tensor containing an energy consumption prediction baseline and environmental change gradients by performing prediction processing on the spatiotemporal feature tensor through a dynamic spatiotemporal graph neural network includes: Using building units as nodes, extract the building reservation occupancy rate time series matrix, real-time pedestrian flow correction factor, solar radiation heat gain coefficient and historical building sub-energy consumption sequence from the spatiotemporal feature tensor as the initial features of the nodes to obtain the node feature set; Based on preset weighting coefficients, the inter-building electrical correlation matrix and inter-building heat conduction weighting matrix in the spatiotemporal feature tensor are weighted and summed to generate static adjacency relationships. The correlation coefficient between the real-time pedestrian flow correction factor in the spatiotemporal feature tensor and the time series matrix of the building's pre-occupancy rate is calculated to generate dynamic edge weights, which are used to characterize the dynamic interaction intensity of different building units at different time periods. The static adjacency relationship and the dynamic edge weight are fused together to generate a spatiotemporal graph adjacency matrix; The node feature set and the spatiotemporal graph adjacency matrix are input into the spatial convolution module of the dynamic spatiotemporal graph neural network for convolution processing to obtain spatial features; the spatial features and the historical building energy consumption sequence in the spatiotemporal feature tensor are input into the temporal convolution module of the dynamic spatiotemporal graph neural network for gated cyclic unit operation processing to generate node state features. The node state features are respectively input into two parallel fully connected prediction layers of the dynamic spatiotemporal graph neural network for prediction processing. The energy consumption prediction baseline for the future preset period is obtained through the first fully connected prediction layer, and the environmental change gradient for the future preset period is obtained through the second fully connected prediction layer. The energy consumption prediction baseline and environmental change gradient are subjected to tensor integration processing to generate the global state tensor.

4. The method according to claim 2, characterized in that, The step of inputting the global state tensor into a hierarchical reinforcement learning model for processing to generate a device-level control instruction set includes: The total energy consumption prediction value and the maximum HVAC energy consumption prediction value are extracted from the global state tensor, and the high-rise strategy input state is constructed by combining the preset electricity price stage information and the preset global comfort index. The high-level policy input state is input into the high-level policy network of the hierarchical reinforcement learning model for inference, generating a regional constraint vector. The regional constraint vector includes the preset maximum power deviation and adjustable range of temperature setpoint for each building unit. Extract the global state tensor slices corresponding to each building unit from the global state tensor, and concatenate the constraint information of the building unit in the region-level constraint vector with the global state tensor slices to generate the underlying policy input state; The underlying policy input state is input into the underlying policy network of the hierarchical reinforcement learning model for processing to generate initial device control parameters, which include temperature setpoint adjustment value, lighting brightness adjustment coefficient and socket power supply duration. A simplified simulation model is constructed based on the campus building thermal parameter library. The initial equipment control parameters are input into the simplified simulation model for simulation. Simulated energy consumption data, simulated environmental temperature and humidity data, and simulated comfort data are obtained and integrated to obtain the execution effect data of the initial equipment control parameters. A collaborative optimization reward function is adopted, and a reward value is calculated based on the execution effect data. The reward value is then used to update the parameters of the high-level policy network and the low-level policy network using the gradient descent method to obtain an optimized hierarchical reinforcement learning model. The high-level policy input state is input into the optimized hierarchical reinforcement learning model for inference processing again, and the optimized device control parameters are output to generate the device-level control instruction set.

5. The method according to claim 3, characterized in that, The process involves inputting the device-level control instruction set into a digital twin environment model for simulation security verification to obtain verification results. Based on these results, an interpretable report of the verified device control instructions and strategies is generated, including: Based on the campus building thermal parameter library, a simplified building thermodynamic model is constructed. The simplified building thermodynamic model and the inter-building electrical correlation matrix in the spatiotemporal feature tensor are embedded into the digital twin environment model for initialization to obtain the initialized digital twin environment model. The device-level control instruction set is input into the initialized digital twin environment model for simulation to obtain state evolution prediction results. The state evolution prediction results include simulation energy consumption data, simulation environment temperature and humidity data, and simulation environment change gradient. A safety constraint threshold is set, and the state evolution prediction result is compared with the safety constraint threshold to obtain a verification result. If the verification result satisfies the safety constraint, the corresponding device control command is used as the verified device control command; if the verification result violates the safety constraint, a safety circuit breaker mechanism is triggered and a preset backup control command is called as the verified device control command. Using the energy consumption prediction baseline as the target energy consumption prediction result, the influence weight of each input feature in the spatiotemporal feature tensor on the target energy consumption prediction result is calculated; based on the influence weight, the input features with influence weights greater than a preset threshold are selected to generate a list of key influence factors. The element values ​​in the adjacency matrix of the spatiotemporal graph are extracted as cross-building association strength data, which is used to characterize the degree of association between different building units; The list of key influencing factors is matched with the cross-building correlation strength data to generate an interpretable report of the strategy.

6. The method according to claim 4, characterized in that, The method further includes: Based on the actual energy consumption data and actual environmental monitoring data obtained after the actual execution of the verified device control command, the actual energy consumption data and actual environmental monitoring data are compared with the energy consumption prediction baseline and environmental change gradient in the global state tensor to generate model error samples. The gradient attribution results are extracted from the strategy-interpretable report. Based on the influence weights of each feature in the gradient attribution results, a feature importance weight matrix is ​​constructed. The model error samples are weighted and fused with the feature importance weight matrix. The parameters of the dynamic spatiotemporal graph neural network are updated using the backpropagation algorithm to obtain the initial updated dynamic spatiotemporal graph neural network. Extract the maximum power deviation threshold from the region-level constraint vector, and determine the samples whose error values ​​exceed the maximum power deviation threshold in the model error samples as boundary samples; inject the boundary samples into the experience pool of the optimized hierarchical reinforcement learning model for update processing to obtain the updated experience pool. Extract the inter-building heat conduction weight matrix from the spatiotemporal feature tensor, calculate the cosine similarity of the heat conduction weight matrix between different building units as the building topological similarity; set a similarity threshold, and retrieve building units whose building topological similarity is greater than the similarity threshold as topologically similar building units; Extract and migrate the node parameters of the initial updated dynamic spatiotemporal graph neural network corresponding to the topologically similar building unit to the nodes of the initial updated dynamic spatiotemporal graph neural network of the target building unit to obtain the initialized target building node parameters; adjust the temporal convolution module parameters of the initial updated dynamic spatiotemporal graph neural network corresponding to the initialized target building node parameters to obtain the optimized dynamic spatiotemporal graph neural network.

7. The method according to claim 3, characterized in that, The dynamic edge weights are calculated using the following formula: in, For the first Time step building unit With building unit The dynamic edge weights between them have a range of values. , The time-series correlation weighting coefficient. The Pearson correlation coefficient function is used to calculate building units. With building unit In the timing window Correlation of pedestrian flow sequences within the region The length of the time series. and Building units and building units In the timing window The time-series subsequence of pedestrian traffic within the area, This represents the total number of building units within the campus. The traversal index for the building unit. For building units With building unit physical center distance, The spatial distance attenuation threshold, For time series smoothing coefficients, , This represents the rate of change in time-series correlation.

8. A smart campus integrated energy-saving control system based on AI decision-making, characterized in that, The system includes: The spatiotemporal feature tensor construction module is used to construct a spatiotemporal feature tensor based on multi-source heterogeneous data from the campus; based on the spatiotemporal feature tensor, a dynamic spatiotemporal graph neural network is used for prediction processing to generate a global state tensor containing the energy consumption prediction baseline and the environmental change gradient. The global state tensor prediction module is used to input the global state tensor into the hierarchical reinforcement learning model for processing and to generate a device-level control instruction set. The instruction generation and security verification module is used to input the device-level control instruction set into the digital twin environment model for simulation security verification to obtain the verification result, and generate a report of the verified device control instructions and strategies that can be interpreted based on the verification result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images