A spatio-temporal transformer driven dynamic cooling resource scheduling system

Abstract

This invention discloses a spatiotemporal Transformer-driven dynamic cooling resource scheduling system, relating to the field of smart grid operation control technology. It includes a prediction module that inputs a spatiotemporal state sequence into a spatiotemporal Transformer prediction model to obtain a charging pile prediction tensor and a spatiotemporal attention weight matrix; a verification module that calculates theoretical values ​​for charging pile cooling resource scheduling based on physical laws, and verifies the results by comparing the theoretical values ​​with the charging pile prediction tensor, generating a verification report; and an instruction module that, based on the spatiotemporal impact report and the verification report, calculates theoretical values ​​based on physical laws and performs deviation analysis with the prediction results, combining statistical process control to generate a verification report, effectively ensuring the safety and reliability of control instructions under physical constraints; and achieving efficient, safe, and adaptive dynamic cooling management.

Description

Technical Field

[0001] This invention relates to the field of smart grid operation control technology, and in particular to a spatiotemporal Transformer-driven dynamic cooling resource scheduling system. Background Technology

[0002] With the rapid development of the electric vehicle industry, the deployment density of high-power DC charging piles in urban charging networks continues to increase, leading to increasingly prominent thermal management issues. Charging piles generate a large amount of heat during high-load operation. If cooling resources are not scheduled in a timely or balanced manner, it can easily lead to local overheating of equipment, accelerated aging of components, or even the risk of thermal runaway, seriously affecting charging safety and equipment lifespan. In recent years, data-driven intelligent scheduling methods have gradually emerged. Some studies have attempted to introduce recurrent neural networks (RNNs) or graph convolutional networks (GCNs) to model the thermal state of charging pile groups, learn temperature change trends through historical data, and pre-allocate cooling resources accordingly. Such methods have improved the foresight of scheduling to a certain extent, especially in dealing with scenarios of periodic load changes.

[0003] Because the heat load variations in charging pile clusters exhibit both temporal non-stationarity and spatial heterogeneity, traditional models struggle to effectively capture the complex spatiotemporal coupling relationships among multiple devices. For example, while LSTM models can characterize time dependencies, their attention mechanism is limited to within the sequence and cannot explicitly model spatial correlations across devices. Furthermore, while clustering strategies consider spatial proximity, they neglect potential thermal interaction effects (such as airflow transfer and accumulated heat radiation) between non-adjacent nodes. When dealing with sudden heat load peaks or spatial heat accumulation effects, problems such as insufficient local cooling or redundant global resource allocation can easily arise, thus affecting system energy efficiency and equipment lifespan. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a spatiotemporal Transformer-driven dynamic cooling resource scheduling system to solve the problem of accurate decision-making and energy efficiency optimization for collaborative cooling of multiple charging piles in highly dynamic scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a spatiotemporal Transformer-driven dynamic cooling resource scheduling system, comprising: a data acquisition module for acquiring device node data from multiple charging piles and preprocessing the data to form a spatiotemporal state sequence; a prediction module for inputting the spatiotemporal state sequence into a spatiotemporal Transformer prediction model to obtain a charging pile prediction tensor and a spatiotemporal attention weight matrix; a spatiotemporal impact module for identifying the Top-K device nodes and their impact periods with the greatest impact in the current decision using the spatiotemporal attention weight matrix, and obtaining a spatiotemporal impact report; a verification module for calculating the theoretical value of charging pile cooling resource scheduling based on physical laws and verifying it by comparing the theoretical value with the charging pile prediction tensor, and obtaining a verification report; an instruction module for generating a control instruction set based on the spatiotemporal impact report and the verification report using a reinforcement learning-optimized decision algorithm; and an execution module for distributing the control instruction set to the execution mechanism for dynamic scheduling of charging pile cooling resources.

[0008] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the device node data includes coolant temperature, pump power, grid load data, temperature, pressure, flow rate, power consumption, operating mode, and device status.

[0009] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the specific steps for forming the spatiotemporal state sequence are as follows:

[0010] The Z-Score statistical method was used to identify and remove pulse-type outliers in the device node data, and the remaining missing data was repaired using linear interpolation of time series data to obtain a cleaned time series dataset.

[0011] The cleaned time series dataset is spatiotemporally aligned, and all time series are resampled at fixed time intervals to generate a regularized time series data matrix.

[0012] Based on the coolant temperature, pump power, grid load data, temperature, pressure, flow rate and power consumption in the regularized time series data matrix, a standardized numerical matrix is ​​obtained by Min-Max transformation;

[0013] The equipment status and operating mode are converted into high-dimensional sparse binary vectors using one-hot encoding, and then concatenated and fused with the standardized numerical matrix to form an enhanced spatiotemporal feature table;

[0014] Device node data within a fixed-time sliding window is extracted from the enhanced spatiotemporal feature table, and the device node data within the window is reconstructed into a three-dimensional tensor structure to form a spatiotemporal state sequence.

[0015] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the specific steps for obtaining the charging pile prediction tensor and the spatiotemporal attention weight matrix are as follows:

[0016] The spatiotemporal state sequence is input into the spatiotemporal Transformer prediction model. The feature vectors of the device nodes are mapped to the high-dimensional embedding space through the linear projection layer, and learnable position encoding and node encoding are added to form the embedding representation.

[0017] The encoder obtains the dependencies between device nodes and time steps through a spatiotemporal self-attention mechanism, and obtains the hidden state representation through scaled dot product attention calculation;

[0018] The decoder operates in an autoregressive manner, initializes a sequence of start symbols, interacts with the hidden state representation through a cross-attention mechanism, and outputs the predicted values ​​of each device node at multiple future time steps through a linear projection layer to form the charging pile prediction tensor.

[0019] The attention weight matrix is ​​extracted from the encoder and post-processed to obtain the spatiotemporal attention weight matrix.

[0020] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the specific steps for extracting the attention weight matrix from the encoder and post-processing the attention weight matrix to obtain the spatiotemporal attention weight matrix are as follows:

[0021] During the forward propagation of the spatiotemporal Transformer prediction model, when the spatiotemporal state sequence passes through the last layer of the encoder, an attention weight matrix is ​​generated through the spatiotemporal self-attention mechanism, and averaged and aggregated on the attention head dimension to obtain a comprehensive attention view.

[0022] Based on the comprehensive attention view, the attention weight of the current time step is focused on in the query time step dimension to obtain the spatiotemporal attention weight matrix.

[0023] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the specific steps for extracting the attention weight row vectors corresponding to the current decision moment from the spatiotemporal attention weight matrix and reorganizing them into a two-dimensional matrix to form a device-time attention scoring matrix are as follows.

[0024] From the spatiotemporal attention weight matrix, extract the row vectors of attention weights corresponding to the current decision moment, and reorganize them into a two-dimensional matrix to form a device-time attention scoring matrix;

[0025] The device-time attention score matrix is ​​aggregated by device dimension to obtain the device node importance score; the device-time attention score matrix is ​​aggregated by time dimension to obtain the time importance score.

[0026] The device nodes are sorted by importance score, and the top-K device nodes with the highest scores are selected as key device nodes to form a list of key device nodes;

[0027] The time importance score with the highest score is selected as the key impact period. The list of key equipment nodes is then linked and integrated with the key impact period to generate a spatiotemporal impact report.

[0028] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the following steps are taken: The theoretical value of the charging pile cooling resource scheduling is calculated based on physical laws, and then compared and verified by comparing the theoretical value of the charging pile cooling resource scheduling with the charging pile prediction tensor to obtain a verification report.

[0029] Based on the spatiotemporal state sequence, the theoretical cooling capacity and theoretical flow rate are calculated using the cooling capacity energy balance equation and the pump performance curve equation, respectively, to obtain the theoretical value of charging pile cooling resource scheduling.

[0030] Based on the theoretical values ​​of the charging pile prediction tensor and the charging pile cooling resource scheduling, the relative error formula is used to calculate the relative deviation of cooling capacity and the relative deviation of flow rate, and then the results are integrated to obtain the instantaneous deviation.

[0031] Collect the relative deviations of cooling capacity and flow rate under all successful operating conditions, and calculate the mean and standard deviation of the relative deviations of cooling capacity and flow rate to obtain the upper limit of dynamic control;

[0032] The instantaneous deviation is compared with the upper limit of dynamic control to generate a verification report.

[0033] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the steps for comparing the instantaneous deviation with the dynamic control upper limit and generating a verification report are as follows:

[0034] The instantaneous deviation is compared with the dynamic control upper limit. If the instantaneous deviation does not exceed the dynamic control upper limit, the verification is passed. If any instantaneous deviation exceeds the dynamic control upper limit, the charging pile prediction tensor is determined to have failed the verification. The verification results, instantaneous deviation, and dynamic control upper limit are integrated into a verification report.

[0035] As a preferred embodiment of the spatiotemporal Transformer-driven dynamic cooling resource scheduling system described in this invention, the specific steps for generating the control instruction set are as follows:

[0036] Based on the spatiotemporal impact report and verification report, generate a list of key equipment operation priorities;

[0037] Using the physical limits of the equipment and the safe operating boundary as constraints, a multi-objective optimization function is constructed.

[0038] A reinforcement learning-based proximal policy optimization algorithm is used to solve the multi-objective optimization function. Combined with a list of key equipment operation priorities, a candidate control instruction set is generated. The spatiotemporal Transformer model is used to quickly simulate and verify the candidate instruction set, and the control instruction set is output.

[0039] As a preferred embodiment of the dynamic cooling resource scheduling system driven by the spatiotemporal Transformer described in this invention, the dynamic scheduling of cooling resources for charging piles refers to converting the control instruction set into industrial standard control signals, using the industrial standard control signals to drive power devices to output three-phase AC power of the corresponding frequency, and driving the variable frequency pump, cooling tower fan and valves to perform actions.

[0040] The beneficial effects of this invention are as follows: by calculating theoretical values ​​based on physical laws and performing deviation analysis with the prediction results, and combining statistical process control to generate a verification report, scientific verification and dynamic credibility assessment of data-driven prediction are realized, effectively ensuring the safety and reliability of control commands under the constraints of physical laws; and efficient, safe, and adaptive dynamic cooling management is achieved. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of a dynamic cooling resource scheduling system driven by a spatiotemporal Transformer.

[0043] Figure 2 A flowchart for generating a spatiotemporal state sequence.

[0044] Figure 3 This is a flowchart for physical verification and deviation analysis.

[0045] Figure 4 A flowchart for generating and verifying control instruction sets. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a spatiotemporal Transformer-driven dynamic cooling resource scheduling system, comprising the following steps:

[0050] The data acquisition module collects device node data from multiple charging piles, preprocesses the device node data, and forms a spatiotemporal state sequence.

[0051] It should be noted that the equipment node data of the charging pile includes coolant temperature, pump power, grid load data, temperature, pressure, flow rate, power consumption, operating mode, and equipment status.

[0052] The Z-Score statistical method was used to identify and remove pulse-type outliers in the device node data, and the remaining missing data was repaired using linear interpolation of time series data to obtain a cleaned time series dataset.

[0053] Specifically, the Z-Score statistical method is used to process the runtime sequence data, and the Z-Score value of each data point in the runtime sequence data is obtained. The Z-Score value is obtained from the mean and standard deviation of the runtime sequence data. When the absolute value of the Z-Score exceeds the outlier threshold, it is identified as an impulse outlier and removed.

[0054] For missing points in the remaining runtime time series data after removing impulse outliers, linear interpolation of time series data is applied to fill the missing data by linearly fitting the data values ​​of the adjacent timestamps before and after the missing points; thus obtaining the cleaned time series dataset.

[0055] It should be noted that the outlier threshold is based on the 3σ principle in statistics, and its value usually falls within the range of the mean ± 3 times the standard deviation.

[0056] The cleaned time series dataset is spatiotemporally aligned, and all time series are resampled at fixed time intervals to generate a regularized time series data matrix.

[0057] Specifically, the timestamps of all sensor data in the cleaned time series dataset are uniformly converted to the time base synchronized by the global network time protocol, and the sensor data is mapped to the corresponding device node position according to the device topology. All time series data are resampled at fixed time intervals (e.g., the time interval is 1 minute), and linear interpolation is used to obtain the values ​​at uniform time points for non-uniform sampled data, generating a regularized time series data matrix.

[0058] It should be noted that the time base is uniformly set by the global network time protocol server address (e.g., 192.168.1.100).

[0059] Based on the coolant temperature, pump power, grid load data, temperature, pressure, flow rate, and power consumption in the regularized time-series data matrix, a standardized numerical matrix is ​​obtained by Min-Max transformation.

[0060] Specifically, coolant temperature, pump power, grid load data, temperature, pressure, flow rate, and power consumption are extracted from the regularized time-series data matrix; the minimum and maximum values ​​of each column are extracted; based on the minimum and maximum values ​​of each column, the data are normalized using Min-Max and then recombined in the original order to form a standardized numerical matrix.

[0061] The equipment status and operating mode are converted into high-dimensional sparse binary vectors using one-hot encoding, and then spliced ​​and fused with standardized numerical matrices to form an enhanced spatiotemporal feature table.

[0062] Specifically, the one-hot encoding method is used to transform each category value of equipment status and operating mode: all possible categories of equipment status and operating mode are determined: all possible categories of equipment status are "normal operation", "abnormal alarm" and "maintenance shutdown", and all possible categories of operating mode are "automatic mode", "manual mode" and "energy saving mode"; the length of the high-dimensional sparse binary vector is equal to the total number of categories;

[0063] The generated high-dimensional sparse binary vectors and the standardized numerical matrix are aligned row-wise and concatenated along the feature dimension: the numerical features at each time point are connected with the corresponding high-dimensional sparse binary vectors to form extended feature vectors; the extended feature vectors at all time points are arranged in chronological order to form an enhanced spatiotemporal feature table.

[0064] Device node data within a fixed-time sliding window is extracted from the enhanced spatiotemporal feature table, and the device node data within the window is reconstructed into a three-dimensional tensor structure to form a spatiotemporal state sequence.

[0065] Specifically, the starting time point and fixed time length of the sliding window are obtained from the enhanced spatiotemporal feature table;

[0066] Extract all feature data (including coolant temperature, pump power, grid load data, temperature, pressure, flow rate, power consumption, etc., and binary features generated by one-hot encoding of equipment status and operating mode) from all device nodes within the sliding window in chronological order. Combine all feature data of device nodes according to different dimensions through tensor reconstruction operation. The first dimension corresponds to the number of samples, the second dimension corresponds to the time step (e.g., 60 time points), and the third dimension corresponds to the total number of features of device nodes, and reconstruct it into a three-dimensional tensor structure to form a spatiotemporal state sequence.

[0067] The prediction module inputs the spatiotemporal state sequence into the spatiotemporal Transformer prediction model to obtain the charging pile prediction tensor and the spatiotemporal attention weight matrix.

[0068] The spatiotemporal state sequence is input into the spatiotemporal Transformer prediction model. The feature vectors of the device nodes are mapped to a high-dimensional embedding space through a linear projection layer, and learnable position and node codes are added to form an embedding representation.

[0069] Specifically, the spatiotemporal state sequence is input into the spatiotemporal Transformer prediction model, and the feature vector of each device node is mapped to a high-dimensional embedding space through a linear projection layer to obtain the mapped feature vector.

[0070] Add a learnable time location encoding vector to each time step, and add a learnable node encoding vector to each device node;

[0071] The mapped feature vector is then element-wise superimposed with the corresponding temporal location encoding vector and node encoding vector to form an embedded representation.

[0072] It should be noted that the training of the spatiotemporal Transformer prediction model consists of spatiotemporal state sequence samples obtained after preprocessing historical running data and corresponding real future state labels; all parameters of the spatiotemporal Transformer prediction model are initialized, including the weight matrix of the linear projection layer, the position encoding vector, the node encoding vector, and the attention layer and feedforward network parameters in the encoder and decoder.

[0073] During forward propagation, the spatiotemporal state sequence is input into the spatiotemporal Transformer prediction model. An embedded representation is formed through linear projection and encoding superposition. After passing through a multi-layer encoder-decoder structure, a tensor predicting future states in multiple steps is output. Based on the data fitting loss between the predicted tensor and the true label, physical rule constraints based on thermodynamics and fluid dynamics are introduced to form a hybrid loss function. The gradient of the loss function with respect to the parameters of the spatiotemporal Transformer prediction model is calculated using the backpropagation algorithm. The parameters of the spatiotemporal Transformer prediction model are updated using the Adam optimizer to minimize the loss function. The forward propagation, loss calculation, and parameter update steps are repeated until the spatiotemporal Transformer prediction model converges, resulting in the trained spatiotemporal Transformer prediction model.

[0074] The encoder obtains the dependencies between device nodes and time steps through a spatiotemporal self-attention mechanism, and obtains the hidden state representation through scaling dot product attention calculation.

[0075] Specifically, the encoder uses three independent fully connected layers to perform linear transformations on the embedding representation to generate a query matrix, a key matrix, and a value matrix: the first fully connected layer projects the embedding representation into a query matrix, the second fully connected layer projects the embedding representation into a key matrix, and the third fully connected layer projects the embedding representation into a value matrix, outputting query matrices, key matrices, and value matrices of the same dimension;

[0076] The similarity between the query matrix and the key matrix is ​​obtained by scaling the dot product attention, and then normalized using the Softmax function to generate the attention weight matrix.

[0077] The product of the attention weight matrix and the value matrix is ​​used as the weighted output. At the same time, the dependencies between device nodes and time steps are combined, and a multi-head attention mechanism is used to concatenate and linearly transform the outputs of multiple attention heads to obtain the hidden state representation.

[0078] It should be noted that the inter-device dependency relationship refers to the mutual influence relationship between different physical devices or monitoring points. For example, an increase in the pump speed will directly affect the pipeline flow and pressure (hydraulic coupling between nodes); a change in the chiller power will cause a change in the cooling tower outlet water temperature (thermal coupling between nodes).

[0079] Inter-step dependency: refers to the temporal correlation of the state values ​​of the same device or monitoring point at different historical moments. For example, the current speed of a water pump is related to the speed change trend in the past 5 minutes (equipment inertia), and the current power consumption of a chiller is related to the load state 10 minutes ago (thermal delay effect).

[0080] The decoder operates in an autoregressive manner, initializing a sequence of start symbols and interacting with the hidden state representation through a cross-attention mechanism. It outputs the predicted values ​​of each device node at multiple future time steps through a linear projection layer, forming the charging pile prediction tensor.

[0081] Specifically, the decoder initializes a sequence of all-zero start symbols as the input to the first time step, and processes the generated time steps through a masked self-attention mechanism to ensure autoregressive properties;

[0082] A cross-attention mechanism is used to interact the intermediate representation of the decoder with the hidden state representation of the encoder. The attention weights of the decoder query and encoder key-value pairs are calculated and weighted summed to generate a context-aware fusion feature representation. The context-aware fusion feature representation is then mapped to the prediction dimension through a linear projection layer to output the predicted values ​​of temperature, pressure, flow rate and power consumption of each device node at multiple future time steps, forming the charging pile prediction tensor.

[0083] The attention weight matrix is ​​extracted from the encoder and post-processed to obtain the spatiotemporal attention weight matrix.

[0084] During the forward propagation of the spatiotemporal Transformer prediction model, when the spatiotemporal state sequence passes through the last layer of the encoder, an attention weight matrix is ​​generated through the spatiotemporal self-attention mechanism, and then averaged and aggregated on the attention head dimension to obtain a comprehensive attention view.

[0085] Specifically, during the forward propagation of the spatiotemporal Transformer prediction model, when the spatiotemporal state sequence passes through the last layer of the encoder, based on the similarity between the query matrix and the key matrix and the attention weight matrix, the weight matrices generated by multiple attention heads in the multi-head attention mechanism are averaged and aggregated along the attention head dimension to obtain a comprehensive attention view.

[0086] Based on the comprehensive attention view, the attention weight of the current time step is focused on in the query time step dimension to obtain the spatiotemporal attention weight matrix.

[0087] Specifically, based on the comprehensive attention view, the attention weight row vector corresponding to the current decision moment is selected in the query time step dimension, and the number of device nodes corresponding to the row dimension and the number of historical time steps corresponding to the column dimension are used to form a spatiotemporal attention weight matrix.

[0088] The spatiotemporal impact module uses a spatiotemporal attention weight matrix to identify the Top-K device nodes and their impact periods that have the greatest impact on the current decision, and then generates a spatiotemporal impact report.

[0089] From the spatiotemporal attention weight matrix, extract the row vectors of attention weights corresponding to the current decision moment, and reorganize them into a two-dimensional matrix to form a device-time attention scoring matrix.

[0090] Specifically, the attention weight row vector corresponding to the current decision moment is extracted from the spatiotemporal attention weight matrix. The attention weight row vector contains the attention weight values ​​of the current decision moment for all historical time steps and each device node.

[0091] The matrix rows correspond to device node indices, the columns correspond to historical time step numbers, and the element values ​​represent the importance score of a specific device to the current decision at a specific historical moment, forming a device-time attention score matrix.

[0092] The device-time attention score matrix is ​​aggregated by device dimension to obtain the device node importance score; the device-time attention score matrix is ​​aggregated by time dimension to obtain the time importance score.

[0093] Specifically, the device-time attention score matrix is ​​summed row by row to obtain the sum of attention weights for each device node across all historical time steps, thus obtaining the device node importance score.

[0094] The device-time attention score matrix is ​​summed column-wise (time dimension) to obtain the sum of attention weights for each historical time step across all device nodes, thus yielding the time importance score.

[0095] The importance scores of the device nodes are sorted, and the top-K device nodes with the highest scores are selected as key device nodes to form a list of key device nodes.

[0096] Specifically, the importance scores of the device nodes are sorted in descending order. Based on the sorting results, the top-K device nodes with the highest scores (e.g., K is 5) are selected, and their device IDs and device type information are recorded. At the same time, the corresponding device node importance scores are also recorded, forming a list of key device nodes containing device IDs, device types, and importance scores.

[0097] The time importance score with the highest score is selected as the key impact period. The list of key equipment nodes is then linked and integrated with the key impact period to generate a spatiotemporal impact report.

[0098] Specifically, the time importance scores are sorted in descending order, and the time period corresponding to the highest time importance score is selected as the key period of influence.

[0099] Each device node in the list of key device nodes is associated with a key period of impact, and the trend of attention weight change for each key device node during the key period of impact is recorded. The list of key device nodes, information on key periods of impact, and weight change data of each device during the corresponding period are integrated to generate a spatiotemporal impact report.

[0100] It should be noted that, based on the spatiotemporal attention weight matrix, the attention weight corresponding to the current decision moment is extracted and reconstructed into a device-time attention scoring matrix. The device scores are sorted to select the Top-K key device nodes, and the time scores are sorted to determine the most important critical impact period. The list of key devices, critical periods, and their dynamic characteristics are integrated to generate a spatiotemporal impact report. In contrast, existing technologies mostly focus on overall time-series trends or global features (e.g., principal component analysis to extract global feature contribution or recurrent neural network hidden states to represent the overall operating trend), making it difficult to accurately locate the specific devices and critical time windows that play a key role in the current decision. This invention, by aggregating and sorting attention weights, can automatically identify the Top-K key device nodes and critical impact periods, enabling rapid location of anomaly sources or core influencing factors, and providing precise guidance for fault diagnosis, root cause analysis, and operation and maintenance response.

[0101] The verification module calculates the theoretical value of the charging pile cooling resource scheduling based on physical laws, and compares and verifies the theoretical value of the charging pile cooling resource scheduling with the charging pile prediction tensor to obtain a verification report.

[0102] Based on the spatiotemporal state sequence, the theoretical cooling capacity and theoretical flow rate are calculated using the cooling capacity energy balance equation and the pump performance curve equation, respectively, to obtain the theoretical value of charging pile cooling resource scheduling.

[0103] Specifically, based on the spatiotemporal state sequence, the current cooling water flow rate measurement, supply and return water temperature measurement, and water pump power consumption measurement are extracted through three-dimensional tensor indexing operations. The theoretical cooling capacity value is calculated using the cooling capacity energy balance equation. The theoretical flow rate value is calculated using the pump performance curve equation, and the theoretical value of charging pile cooling resource scheduling is output.

[0104] It should be noted that the expression for calculating the theoretical cooling capacity is as follows:

[0105] ;

[0106] in, This represents the theoretical cooling capacity. Indicates the mass flow rate of cooling water. This represents the specific heat capacity of water at constant pressure. Indicates the return water temperature. This indicates the water supply temperature.

[0107] The expression for calculating the theoretical flow rate is:

[0108] ;

[0109] in, This represents the theoretical flow rate. Indicates the input power of the water pump. Indicates pump efficiency. This indicates the density of water. Represents gravitational acceleration. This indicates the pump head.

[0110] It should be noted that the two expressions above were standardized using Min-Max before calculation to ensure dimensional uniformity.

[0111] Based on the theoretical values ​​of the charging pile prediction tensor and the charging pile cooling resource scheduling, the relative error formula is used to calculate the relative deviation of cooling capacity and the relative deviation of flow rate, and then integrate them to obtain the instantaneous deviation.

[0112] Specifically, the predicted cooling capacity and flow rate are extracted based on the charging pile prediction tensor at the current moment, and the theoretical cooling capacity and flow rate are obtained based on the theoretical value of the charging pile cooling resource scheduling. The relative error formula is used to calculate the relative deviation of cooling capacity and the relative deviation of flow rate. The relative deviation of cooling capacity and the relative deviation of flow rate are combined to form the instantaneous deviation.

[0113] It should be noted that the expression for calculating the relative deviation of cooling capacity is as follows:

[0114] ;

[0115] in, This indicates the relative deviation of the cooling capacity. This represents the predicted cooling capacity.

[0116] The expression for calculating the relative deviation of flow rate is:

[0117] ;

[0118] in, This indicates a relative deviation in flow rate. This represents the predicted flow rate. This represents the theoretical flow rate.

[0119] Collect the relative deviations of cooling capacity and flow rate under all successful operating conditions, and calculate the mean and standard deviation of the relative deviations of cooling capacity and flow rate to obtain the upper limit of dynamic control.

[0120] Specifically, the relative deviations of cooling capacity and flow rate recorded under all successful operating conditions are collected to form a relative deviation sequence of cooling capacity and a relative deviation sequence of flow rate, respectively.

[0121] The arithmetic mean and standard deviation of the relative deviation sequence of cooling capacity and the relative deviation sequence of flow rate are calculated respectively. Based on the 3σ principle, the dynamic control upper limit of the relative deviation of cooling capacity and the dynamic control upper limit of the relative deviation of flow rate are obtained respectively, and then integrated into a dynamic control upper limit.

[0122] The instantaneous deviation is compared with the upper limit of dynamic control to generate a verification report.

[0123] The instantaneous deviation is compared with the dynamic control upper limit. If the instantaneous deviation does not exceed the dynamic control upper limit, the verification is passed. If any instantaneous deviation exceeds the dynamic control upper limit, the charging pile prediction tensor is determined to have failed the verification. The verification results, instantaneous deviation, and dynamic control upper limit are integrated into a verification report.

[0124] Specifically, the relative deviation of cooling capacity in the instantaneous deviation vector is compared with the upper limit of dynamic control, and the relative deviation of flow rate in the instantaneous deviation vector is compared with the upper limit of dynamic control.

[0125] The verification is deemed successful when the relative deviation of cooling capacity is not greater than the upper limit of dynamic control of relative deviation of cooling capacity and the relative deviation of flow rate is not greater than the upper limit of dynamic control of relative deviation of flow rate. The verification is deemed unsuccessful when the relative deviation of cooling capacity is greater than the upper limit of dynamic control of relative deviation of cooling capacity or the relative deviation of flow rate is greater than the upper limit of dynamic control of relative deviation of flow rate.

[0126] The verification conclusions, instantaneous deviations, and dynamic control limits are integrated to form a verification report that includes the verification status, deviation data, and control limits.

[0127] It should be noted that by collecting actual operating data, the theoretical cooling capacity and theoretical flow rate are calculated using the cooling capacity energy balance equation and pump performance curve equation to form the theoretical value of cooling resource scheduling. Combined with the charging pile prediction tensor predicted by the spatiotemporal Transformer prediction model, the instantaneous deviation is calculated, and a dynamic control upper limit is generated based on historical successful operating conditions. By comparing the instantaneous deviation with the control upper limit, the rationality of the prediction results is judged, and a verification report is generated. In contrast, the verification of traditional models (such as LSTM models and ARIMA time series models) relies heavily on historical data statistics, making it difficult to guarantee the physical rationality of the prediction results. This invention calculates the theoretical value of cooling resource scheduling based on physical laws such as energy balance and pump performance, performs deviation analysis with the charging pile prediction tensor, and constructs a dynamic control upper limit based on the 3σ principle, thus achieving dual verification of the physical consistency and operational reliability of the prediction results.

[0128] The instruction module generates a set of control instructions based on spatiotemporal impact reports and verification reports, using reinforcement learning to optimize decision-making algorithms.

[0129] Based on the spatiotemporal impact report and verification report, a list of key equipment operation priorities is generated.

[0130] Specifically, based on the list of key equipment nodes and the importance score of the equipment nodes in the spatiotemporal impact report, and combined with the verification conclusions in the verification report, the importance score of the equipment nodes is adjusted by weighting: when the verification report determines that the verification has passed, the importance score of the equipment nodes remains unchanged; when the verification report determines that the verification has failed, the product of the importance score of the equipment nodes and a reduction coefficient (e.g., the reduction coefficient is 0.5) is used as the adjusted importance score of the equipment nodes; the equipment nodes are re-sorted according to the adjusted importance score, and the top-K equipment nodes with the highest scores are selected to generate a list of key equipment operation priorities.

[0131] A multi-objective optimization function is constructed by using the physical limits of the equipment and the safe operating boundary as constraints.

[0132] Specifically, the maximum pump speed constraint and maximum valve opening constraint in the physical limits of the equipment, and the pipeline pressure limit and flow lower limit in the safe operation boundary are used as inequality constraints. A multi-objective optimization function containing energy cost objective, performance risk objective and equipment risk objective is constructed by weighted summation.

[0133] By dynamically adjusting the weight coefficients of each objective function based on the conclusions of the verification report, a multi-objective optimization function is formed.

[0134] It should be noted that the physical limits of the equipment (maximum pump speed, maximum valve opening) are derived from the equipment design parameters provided by the equipment manufacturer, while the safe operating boundaries (such as pipeline pressure limits and lower flow limits) are the minimum / maximum operating thresholds for ensuring stability and safety summarized from historical operating data.

[0135] A reinforcement learning-based proximal policy optimization algorithm is used to solve the multi-objective optimization function. Combined with a list of key equipment operation priorities, a candidate control instruction set is generated. The spatiotemporal Transformer model is used to quickly simulate and verify the candidate instruction set, and the control instruction set is output.

[0136] Specifically, a reinforcement learning proximal policy optimization algorithm is used to iteratively solve the multi-objective optimization function: the parameters of the deep learning policy network, which include the setpoints of the pump speed and the valve opening, are initialized, and a candidate control instruction set is generated based on the current deep learning policy network parameters in each iteration.

[0137] The control instruction set is filtered by the priority list of key equipment operation, and the operation instructions for key equipment are retained. The filtered candidate control instruction set is then input into the spatiotemporal Transformer model for fast closed-loop simulation to predict future states and obtain the optimized objective function value.

[0138] The policy gradient is calculated based on the objective function value. The parameters of the deep learning policy network are updated by combining the trust domain constraint of the near-end policy optimization. The process is repeated iteratively until convergence, and then the control instruction set is output.

[0139] It should be noted that the reinforcement learning proximal policy optimization algorithm (PPO) is an advanced model-independent policy gradient algorithm that significantly improves the stability and sampling efficiency of the training process by limiting the step size of policy updates to ensure that the new policy does not differ too much from the old policy.

[0140] The execution module sends control command sets to the execution mechanism to dynamically schedule the cooling resources of the charging pile.

[0141] Dynamic scheduling of cooling resources for charging piles refers to converting control command sets into industrial standard control signals, using these signals to drive power devices to output three-phase AC power at the corresponding frequency, thereby driving variable frequency pumps, cooling tower fans, and valves to operate.

[0142] Specifically, the pump speed setpoint in the control command set is converted into a frequency command acceptable to the frequency converter by proportional scaling (e.g., 75% speed corresponds to 37.5Hz), and the valve opening setpoint is converted into a 4-20mA analog signal;

[0143] The frequency command is encapsulated into a write-hold register command via the ModbusTCP protocol and sent to the frequency converter, which outputs a current signal to the valve positioner. After receiving the command, the frequency converter drives the IGBT power device to generate three-phase AC power of the corresponding frequency to control the motor speed. After receiving the current signal, the valve positioner drives the motor to adjust the valve core position. At the same time, the actual output frequency of the frequency converter and the actual opening degree of the valve are collected in real time to complete the dynamic scheduling of the charging pile cooling resources.

[0144] In summary, this invention achieves scientific verification and dynamic credibility assessment of data-driven predictions by calculating theoretical values ​​based on physical laws and performing deviation analysis with the predicted results, combined with statistical process control to generate verification reports. This effectively ensures the safety and reliability of control commands under the constraints of physical laws and realizes efficient, safe, and adaptive dynamic cooling management.

[0145] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A spatiotemporal Transformer-driven dynamic cooling resource scheduling system, characterized in that: include, The data acquisition module collects device node data from multiple charging piles, preprocesses the device node data, and forms a spatiotemporal state sequence. The prediction module inputs the spatiotemporal state sequence into the spatiotemporal Transformer prediction model to obtain the charging pile prediction tensor and the spatiotemporal attention weight matrix. The spatiotemporal impact module uses a spatiotemporal attention weight matrix to identify the Top-K device nodes and their impact periods in the current decision-making process, and then generates a spatiotemporal impact report. The verification module calculates the theoretical value of the charging pile cooling resource scheduling based on physical laws, and compares and verifies the theoretical value of the charging pile cooling resource scheduling with the charging pile prediction tensor to obtain a verification report. The instruction module generates a set of control instructions based on spatiotemporal impact reports and verification reports, using reinforcement learning to optimize decision-making algorithms. The execution module sends control command sets to the execution mechanism to dynamically schedule the cooling resources of the charging pile.

2. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The device node data includes coolant temperature, pump power, grid load data, temperature, pressure, flow rate, power consumption, operating mode, and device status.

3. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The specific steps for forming the spatiotemporal state sequence are as follows: The Z-Score statistical method was used to identify and remove pulse-type outliers in the device node data, and the remaining missing data was repaired using linear interpolation of time series data to obtain a cleaned time series dataset. The cleaned time series dataset is spatiotemporally aligned, and all time series are resampled at fixed time intervals to generate a regularized time series data matrix. Based on the coolant temperature, pump power, grid load data, temperature, pressure, flow rate and power consumption in the regularized time series data matrix, a standardized numerical matrix is ​​obtained by Min-Max transformation; The equipment status and operating mode are converted into high-dimensional sparse binary vectors using one-hot encoding, and then concatenated and fused with the standardized numerical matrix to form an enhanced spatiotemporal feature table; Device node data within a fixed-time sliding window is extracted from the enhanced spatiotemporal feature table, and the device node data within the window is reconstructed into a three-dimensional tensor structure to form a spatiotemporal state sequence.

4. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The specific steps for obtaining the charging pile prediction tensor and the spatiotemporal attention weight matrix are as follows: The spatiotemporal state sequence is input into the spatiotemporal Transformer prediction model. The feature vectors of the device nodes are mapped to the high-dimensional embedding space through the linear projection layer, and learnable positional encoding node encoding is added to form the embedding representation. The encoder obtains the dependencies between device nodes and time steps through a spatiotemporal self-attention mechanism, and obtains the hidden state representation through scaled dot product attention calculation; The decoder operates in an autoregressive manner, initializes a sequence of start symbols, interacts with the hidden state representation through a cross-attention mechanism, and outputs the predicted values ​​of each device node at multiple future time steps through a linear projection layer to form the charging pile prediction tensor. The attention weight matrix is ​​extracted from the encoder and post-processed to obtain the spatiotemporal attention weight matrix.

5. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 4, characterized in that: The steps for extracting the attention weight matrix from the encoder and post-processing it to obtain the spatiotemporal attention weight matrix are as follows: During the forward propagation of the spatiotemporal Transformer prediction model, when the spatiotemporal state sequence passes through the last layer of the encoder, an attention weight matrix is ​​generated through the spatiotemporal self-attention mechanism, and averaged and aggregated on the attention head dimension to obtain a comprehensive attention view. Based on the comprehensive attention view, the attention weight of the current time step is focused on in the query time step dimension to obtain the spatiotemporal attention weight matrix.

6. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 5, characterized in that: The process of using a spatiotemporal attention weight matrix to identify the Top-K device nodes and their respective time periods with the greatest impact on the current decision, and obtaining a spatiotemporal impact report, is detailed below. From the spatiotemporal attention weight matrix, extract the row vectors of attention weights corresponding to the current decision moment, and reorganize them into a two-dimensional matrix to form a device-time attention scoring matrix; The device-time attention score matrix is ​​aggregated by device dimension to obtain the device node importance score; The device-time attention score matrix is ​​aggregated along the time dimension to obtain the time importance score; The device nodes are sorted by importance score, and the top-K device nodes with the highest scores are selected as key device nodes to form a list of key device nodes; The time importance score with the highest score is selected as the key impact period. The list of key equipment nodes is then linked and integrated with the key impact period to generate a spatiotemporal impact report.

7. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The theoretical values ​​for charging pile cooling resource scheduling are calculated based on physical laws. These theoretical values ​​are then compared and verified with the predicted tensor of the charging piles to obtain a verification report. The specific steps are as follows: Based on the spatiotemporal state sequence, the theoretical cooling capacity and theoretical flow rate are calculated using the cooling capacity energy balance equation and the pump performance curve equation, respectively, to obtain the theoretical value of charging pile cooling resource scheduling. Based on the theoretical values ​​of the charging pile prediction tensor and the charging pile cooling resource scheduling, the relative error formula is used to calculate the relative deviation of cooling capacity and the relative deviation of flow rate, and then the results are integrated to obtain the instantaneous deviation. Collect the relative deviations of cooling capacity and flow rate under all successful operating conditions, and calculate the mean and standard deviation of the relative deviations of cooling capacity and flow rate to obtain the upper limit of dynamic control; The instantaneous deviation is compared with the upper limit of dynamic control to generate a verification report.

8. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 7, characterized in that: The process of comparing the instantaneous deviation with the dynamic control upper limit and generating a verification report involves the following steps: The instantaneous deviation is compared with the dynamic control upper limit. If the instantaneous deviation does not exceed the dynamic control upper limit, the verification is passed. If any instantaneous deviation exceeds the dynamic control upper limit, the charging pile prediction tensor is determined to have failed the verification. The verification results, instantaneous deviations, and dynamic control limits are integrated into a verification report.

9. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The specific steps for generating the control instruction set are as follows: Based on the spatiotemporal impact report and verification report, generate a list of key equipment operation priorities; Using the physical limits of the equipment and the safe operating boundary as constraints, a multi-objective optimization function is constructed. A reinforcement learning-based proximal policy optimization algorithm is used to solve the multi-objective optimization function. Combined with a list of key equipment operation priorities, a candidate control instruction set is generated. The spatiotemporal Transformer model is used to quickly simulate and verify the candidate instruction set, and the control instruction set is output.

10. The spatiotemporal Transformer-driven dynamic cooling resource scheduling system as described in claim 1, characterized in that: The dynamic scheduling of charging pile cooling resources refers to converting the control command set into industrial standard control signals, using the industrial standard control signals to drive power devices to output three-phase AC power of the corresponding frequency, and driving the variable frequency pump, cooling tower fan and valve to perform actions.

Images