Intelligent control method and storage medium for power storage battery of transformer area

Abstract

The application provides a kind of intelligent control method and storage medium of transformer area energy storage battery. The method includes: using SOH prediction model to predict the remaining life based on the real-time operation data of transformer area energy storage battery, to obtain real-time multi-dimensional health index;The historical multi-dimensional health index sequence and real-time multi-dimensional health index are fused under the self-attention mechanism to obtain the charge-discharge constraint parameter of energy storage battery;Based on the charge-discharge constraint parameter, the energy storage battery executes the charge-discharge action.The scheme of the application realizes the fusion of historical multi-dimensional health index sequence and real-time multi-dimensional health index, so that real-time health index can adaptively focus on the time step related to the historical sequence, and the historical attenuation trend information is aggregated into the feature representation of the current time. By controlling the energy storage battery to execute the charge-discharge action based on the charge-discharge constraint parameter, a complete closed loop from health status evaluation to charge-discharge execution control is realized.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to an intelligent control method and storage medium for a transformer substation energy storage battery. Background Technology

[0002] Energy storage batteries are widely used in power quality regulation scenarios in power distribution areas. Accurate assessment of their health status and the rational formulation of charging and discharging strategies have a significant impact on the safe and stable operation of the power system. With the accumulation of battery cycles, a series of irreversible degradation processes continuously occur inside the battery, including electrode material aging, loss of active materials, increased internal resistance, and lithium plating side reactions. These degradation processes collectively lead to a decrease in usable battery capacity and power output capability, while simultaneously exacerbating safety hazards such as abnormal temperature rises. Summary of the Invention

[0003] One objective of this invention is to improve the multi-dimensional comprehensive characterization capability of energy storage battery health status assessment.

[0004] A further objective of this invention is to improve the adaptability of charge and discharge constraint parameters to real-time dynamic changes in battery health status.

[0005] Specifically, the present invention provides an intelligent control method for a power grid energy storage battery, comprising: The remaining life of energy storage batteries in a distribution area is predicted using a SOH prediction model based on real-time operating data. This yields real-time multidimensional health indicators, including at least one of the following: battery health status estimate, internal resistance change rate, lithium plating risk parameters, and abnormal temperature rise parameters. The SOH prediction model is trained using federated learning on historical operating data labeled with battery health status tags. During training, a loss function is used to optimize the model parameters. This loss function is a weighted sum of at least one of the following: battery health status estimation loss, internal resistance evolution loss, lithium plating risk discrimination loss, and thermal behavior analysis loss. The battery health status estimation loss is determined using a mean square error loss function or a mean absolute error loss function. The internal resistance evolution loss is determined using a logarithmic mean square error loss function or a Huber loss function. The lithium plating risk discrimination loss is determined using a weighted cross-entropy loss function or a Focal loss function. The thermal behavior analysis loss is determined using a composite loss function combining the mean square error loss function and a temperature compensation correction term. The weighting coefficients of the weighted sum are preset according to the actual needs of the distribution area energy storage battery remaining life prediction task. The historical multidimensional health indicator sequence and the real-time multidimensional health indicator are fused under a self-attention mechanism to obtain the charge and discharge constraint parameters of the energy storage battery; the historical multidimensional health indicator sequence is a time series data formed by arranging the multidimensional health indicators recorded and stored by the energy storage battery in the past operating cycle in chronological order; the charge and discharge constraint parameters include at least one of the following: upper limit of charge and discharge rate, upper limit of charge and discharge depth, and temperature operating range. The energy storage battery is controlled to perform charging and discharging actions based on the charging and discharging constraint parameters.

[0006] Optionally, the remaining lifespan can be predicted using a SOH prediction model based on real-time operating data of the energy storage batteries in the distribution area, resulting in real-time multi-dimensional health indicators, including: The SOH prediction model includes a battery health state estimation branch, an internal resistance evolution branch, a lithium plating risk discrimination branch, and a thermal behavior analysis branch. Based on the battery health state estimation branch, the voltage, current and temperature time-series features in the real-time operating data are extracted and decoded to obtain the battery health state estimate. And / or, based on the internal resistance evolution branch, feature extraction and decoding are performed on the internal resistance time-series evolution component in the real-time running data to obtain the internal resistance change rate; And / or, based on the lithium plating risk discrimination branch, feature extraction and decoding are performed on the negative electrode potential time series component in the real-time operating data to obtain the lithium plating risk parameter; And / or, based on the thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain the temperature rise anomaly parameters.

[0007] Optionally, based on the battery health state estimation branch, feature extraction and decoding are performed on the voltage, current, and temperature time-series features in the real-time operating data to obtain the battery health state estimate, including: The battery health state estimation branch extracts and decodes the time-series features of voltage, current and temperature in the real-time operating data to obtain the first health feature map. The first health feature map is compressed into a one-dimensional feature vector by global average pooling. The one-dimensional feature vector is input into multiple fully connected layers for nonlinear transformation, and the result of the nonlinear transformation is mapped to the battery health state estimate through an activation function.

[0008] Optionally, based on the internal resistance evolution branch, feature extraction and decoding are performed on the internal resistance time-series evolution component in the real-time operating data to obtain the internal resistance change rate, including: The internal resistance time-series evolution component in the real-time running data is extracted and decoded by the internal resistance evolution branch to obtain the second health feature map. The internal resistance change trend is predicted by using a temporal convolutional layer for each time step feature in the second health feature map, and the internal resistance change trend is decoded by inverse normalization to obtain the internal resistance change rate.

[0009] Optionally, based on the lithium plating risk discrimination branch, feature extraction and decoding are performed on the negative electrode potential time-series component in the real-time operating data to obtain the lithium plating risk parameters, including: The lithium plating risk discrimination branch is used to extract and decode the negative electrode potential time series component in the real-time operation data to obtain the third health feature map. The lithium plating-sensitive features in the third health feature map are enhanced by an attention-weighted layer, and the enhanced features are mapped to lithium plating risk level or lithium plating risk probability value by a classification output layer to obtain the lithium plating risk parameter.

[0010] Optionally, based on the thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain the temperature rise anomaly parameters, including: The thermal behavior analysis branch is used to extract and decode the features of the multi-point temperature spatiotemporal matrix in the real-time operation data to obtain the fourth health feature map. The temperature rise offset is predicted for each spatial location in the fourth health feature map by a regression convolutional layer, and the temperature rise offset is decoded and corrected by temperature compensation to obtain the temperature rise anomaly parameters.

[0011] Optionally, the historical multidimensional health indicator sequence and the real-time multidimensional health indicator are fused using a self-attention mechanism to obtain the charge and discharge constraint parameters of the energy storage battery, including: The real-time multidimensional health indicator is appended to the last time step of the historical multidimensional health indicator sequence as the current time query vector to construct a complete input sequence containing historical and current time information. Each time step in the complete input sequence is position-encoded to obtain an encoded sequence carrying the temporal position. The feature vector at each time step in the encoded sequence is mapped to a query vector, a key vector, and a value vector through a query projection layer, a key projection layer, and a value projection layer, respectively. A query matrix, a key matrix, and a value matrix are obtained based on the query vector, the key vector, and the value vector. The query projection layer, the key projection layer, and the value projection layer are each independent linear mapping networks. Multi-head self-attention computation is performed on the query matrix, key matrix, and value matrix to obtain a self-attention enhanced output sequence; Extract the output vector corresponding to the last time step of the real-time multidimensional health indicator from the self-attention enhanced output sequence, and use it as a composite health feature vector that integrates historical decay trend information and current health status information. The composite health feature vector is input into a fully connected mapping network. The output layer of the fully connected mapping network contains the same number of neurons as the charge-discharge constraint parameters, and each neuron outputs a charge-discharge constraint parameter.

[0012] Optionally, controlling the energy storage battery to perform charging and discharging actions based on the charging and discharging constraint parameters includes: According to the preset charging and discharging constraint parameter-execution command mapping relationship, the upper limit of the charging and discharging rate, the upper limit of the charging and discharging depth, and the temperature operating range are converted into control parameters for the energy storage battery. The control parameters include the upper limit of the current amplitude, the voltage protection threshold, and the heat dissipation start / stop threshold.

[0013] According to another aspect of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the intelligent control method for the substation energy storage battery described above.

[0014] According to another aspect of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the steps of the intelligent control method for the substation energy storage battery described above.

[0015] According to another aspect of the present invention, a computer device is also provided, which includes a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any intelligent control method for a substation energy storage battery.

[0016] The intelligent control method for energy storage batteries in this invention outputs real-time multi-dimensional health indicators, including battery health state estimates, internal resistance change rate, lithium plating risk parameters, and temperature rise anomaly parameters, through a State of Health (SOH) prediction model. This overcomes the limitation of a single SOH value in distinguishing different degradation modes, achieving a comprehensive multi-dimensional characterization of the energy storage battery's health state. A self-attention mechanism is used to fuse historical multi-dimensional health indicator sequences with real-time multi-dimensional health indicators, enabling the real-time health indicators to adaptively focus on relevant time steps in the historical sequence. Historical degradation trend information is aggregated into the current moment's feature representation, ensuring that the generated charge / discharge constraint parameters not only reflect the battery's current health state but also incorporate the future safety boundary change direction predicted by historical aging trajectories, improving the foresight of control decisions. By controlling the energy storage battery to perform charge / discharge actions based on these constraint parameters, a complete closed loop from health state assessment to charge / discharge execution control is achieved.

[0017] Furthermore, the intelligent control method for the energy storage battery in the distribution area of ​​the present invention appends real-time multidimensional health indicators as query vectors to the end of the historical sequence and performs self-attention calculation on the complete sequence, enabling the current health state to establish a direct attentional association with the degradation information of all time steps in the historical sequence. Compared to relying on data from the most recent few time steps for prediction, the present invention can adaptively capture the historical degradation stage most similar to the current state, thereby improving the accuracy and adaptability of the prediction of charge and discharge constraint parameters.

[0018] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0019] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a schematic flowchart of an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 2 This is a schematic flowchart illustrating the process of obtaining a battery health status estimate in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 3 This is a schematic flowchart illustrating the method for obtaining the rate of change of internal resistance in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 4 This is a schematic flowchart illustrating the process of obtaining lithium plating risk parameters in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 5 This is a schematic flowchart illustrating the method for obtaining abnormal temperature rise parameters in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 6 This is a schematic flowchart illustrating the process of obtaining the charging and discharging constraint parameters of an energy storage battery in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 7 This is a schematic diagram of a computer program product according to an embodiment of the present invention; Figure 8 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention; Figure 9 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0020] Energy storage batteries are widely used in power quality regulation in distribution substations, undertaking tasks such as smoothing out fluctuations in distributed photovoltaic output, compensating for reactive power in the grid, and addressing three-phase load imbalances. However, the load and photovoltaic output in distribution substations exhibit significant temporal, random, and uncertain characteristics, causing energy storage batteries to operate under complex and variable conditions for extended periods. This necessitates frequent deep charge / discharge state switching and continuous, rapid response to dynamic grid dispatch commands, resulting in frequent and drastic fluctuations in charge / discharge power and operating current. These operating conditions place higher demands on the real-time accuracy of battery health status sensing and the dynamic adaptability of control strategies.

[0021] Accurate assessment of the health status of energy storage batteries and the rational formulation of charging and discharging strategies have a significant impact on the safe and stable operation of the power system. Inaccurate battery health status assessments will lead to deviations in the judgment of the remaining battery life by maintenance personnel, resulting in scheduling tasks being arranged according to factory parameters even when the battery is severely aged, increasing the risk of system operation. Inappropriate charging and discharging strategies may lead to the issuance of high-rate charging and discharging commands when the battery's actual capacity is insufficient, accelerating battery performance degradation and even causing safety accidents.

[0022] As the number of cycles in an energy storage battery accumulates, a series of irreversible degradation processes continuously occur inside the battery. On the positive electrode side, lattice distortion and transition metal dissolution occur in the active material, leading to a reduction in lithium-ion insertion and extraction sites and a continuous decrease in the specific capacity of the positive electrode. On the negative electrode side, the solid electrolyte interface film on the graphite surface continuously thickens and fractures, increasing interfacial impedance while consuming active lithium ions. In some areas, lithium dendrites also precipitate and grow, increasing the risk of internal short circuits. On the electrolyte side, lithium salt decomposition and additive consumption lead to a decrease in conductivity, and decomposition products may clog the membrane channels, increasing ion transport resistance. These internal degradations manifest at the power output end as a weakened battery voltage support capability and an increased power tracking error. Consequently, on the AC side, this results in a decreased ability of the energy storage system to respond to grid dispatch commands and a reduction in regulation accuracy. In severe cases, it may cause the energy storage system to shut down due to protection actions, affecting the effectiveness of power quality management in the distribution area.

[0023] Figure 1 This is a schematic flowchart of an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 1 As shown, the intelligent control method for the energy storage battery in this distribution area generally includes: Step S101: The remaining life of the energy storage batteries in the distribution area is predicted using the SOH prediction model based on real-time operating data. Real-time multi-dimensional health indicators are obtained, including at least one of the following: battery health status estimate, internal resistance change rate, lithium plating risk parameter, and temperature rise anomaly parameter. The SOH prediction model is trained through federated learning on historical operating data labeled with battery health status tags. During training, a loss function is used to optimize the model parameters of the SOH prediction model. The loss function is a weighted sum of at least one of the following: battery health status estimation loss, internal resistance evolution loss, lithium plating risk discrimination loss, and thermal behavior analysis loss. The battery health status estimation loss is determined using a mean square error loss function or a mean absolute error loss function; the internal resistance evolution loss is determined using a logarithmic mean square error loss function or a Huber loss function; the lithium plating risk discrimination loss is determined using a weighted cross-entropy loss function or a Focal loss function; and the thermal behavior analysis loss is determined using a composite loss function combining the mean square error loss function and a temperature compensation correction term. The weighting coefficients of the weighted sum are preset according to the actual needs of the energy storage battery remaining life prediction task in the distribution area. Step S102: The historical multidimensional health index sequence and the real-time multidimensional health index are fused under the self-attention mechanism to obtain the charge and discharge constraint parameters of the energy storage battery; the historical multidimensional health index sequence is the time series data formed by arranging the multidimensional health indicators recorded and stored in the past operating cycles of the energy storage battery in the distribution area in chronological order; the charge and discharge constraint parameters include at least one of the following: upper limit of charge and discharge rate, upper limit of charge and discharge depth, and temperature operating range. Step S103: Control the energy storage battery to perform charging and discharging actions based on charging and discharging constraint parameters.

[0024] In this embodiment, a State of Health (SOH) prediction model is deployed at the edge node of the distribution area. This SOH prediction model adopts a temporal neural network architecture and is trained by federated learning on historical operating data labeled with battery health status. Real-time operating data of the energy storage battery is continuously collected at a preset sampling period, with a sampling frequency range of 1Hz to 10Hz. The collected data includes at least the current number of charge-discharge cycles, cumulative throughput, historical average operating temperature, historical maximum charge-discharge rate, terminal voltage response curve, and current excitation signal. The above real-time operating data is input into the SOH prediction model, which performs forward inference calculations and outputs real-time multidimensional health indicators. These real-time multidimensional health indicators include at least one of the following: battery health status estimate, internal resistance change rate, lithium plating risk parameters, and abnormal temperature rise parameters.

[0025] The battery health status estimate is the percentage of the battery's full charge capacity to its factory-status capacity at the current moment, ranging from 0% to 100%, used to quantify the overall aging degree of the battery. The internal resistance change rate is used to characterize the rate of battery polarization growth and active material degradation. The lithium plating risk parameter is determined based on charging current, charging end voltage, and battery temperature, with a value range of 0 to 1; a higher value indicates a greater risk of lithium plating. The temperature rise anomaly parameter is determined based on battery surface temperature, ambient temperature, and current charge / discharge power, and outputs the temperature rise anomaly level after comparison with a preset safety threshold.

[0026] In this embodiment, the SOH prediction model is trained based on a decentralized federated learning framework. The specific training process uses historical operating data labeled with the true values ​​of battery health status in multiple areas as the training sample set. A training mode of distributed local training and global parameter aggregation and update is adopted. After each node completes the joint training of the multi-branch model locally, it encrypts the model parameters and broadcasts them to neighboring nodes. Each node receives the model parameters from multiple neighboring nodes and performs local aggregation, updating the local model parameters by replacement or weighted average. The process is iteratively executed until the global model converges.

[0027] The battery health status estimation loss is a regression task loss, which uses the mean squared error loss function or the mean absolute error loss function to quantify the numerical deviation between the battery health status estimate and the actual battery health status. The internal resistance evolution loss adopts the logarithmic mean square error loss function or Huber loss function. When the residual is small, it is expressed as mean square error, and when the residual is large, it is expressed as mean absolute error, so as to suppress the interference of sudden changes in charging and discharging conditions and sampling anomalies on the fitting of time series trends. The lithium plating risk discrimination loss adopts the weighted cross-entropy loss function or the Focal loss function. By reducing the loss weight of simple negative samples and focusing on difficult-to-classify lithium plating abnormal samples, it alleviates the data imbalance problem of sparse battery lithium plating samples and high proportion of normal samples. The mean square error loss function is used in the thermal behavior analysis to quantify the error between the predicted and actual values ​​of the temperature rise offset at each measuring point.

[0028] The weighted coefficients for each branch loss are pre-configured according to the operation and maintenance requirements of the energy storage batteries in the distribution area: if the focus is on accurate assessment of remaining life, the weight coefficients of battery health status estimation loss and internal resistance evolution loss are increased; if the focus is on operational safety early warning, the weight coefficients of lithium plating risk discrimination loss and thermal behavior analysis loss are increased.

[0029] The supervised training method based on multi-loss joint backpropagation drives the collaborative iterative optimization of the parameters of each branch network of the model, and establishes a coupled correlation representation between various aging phenomena such as battery capacity decay, internal resistance increase, lithium plating growth and abnormal temperature rise. This enables the model to have the ability to perceive across degradation and improves the accuracy of charge and discharge constraint parameters.

[0030] In this embodiment, the multidimensional health indicators recorded and stored by the energy storage batteries in the distribution area within a preset historical period are arranged chronologically to form a standardized historical health indicator sequence. The historical multidimensional health indicators and the real-time multidimensional health indicators have the same feature dimensions and parameter types. The real-time multidimensional health indicators are appended to the end of the historical health indicator sequence as the feature vector of the current moment, constructing a complete input sequence containing historical time-series information and current moment information. The complete input sequence is input into a Transformer-based temporal fusion network. The temporal fusion network performs temporal feature fusion on the complete input sequence based on a self-attention mechanism. It mines the correlation weights between each historical time step and the final time step through parallel attention calculation of a query-key-value structure, obtaining a self-attention-enhanced output sequence. The output vector corresponding to the final time step is extracted from the self-attention-enhanced output sequence as a composite health feature vector that deeply couples the historical degradation trend with the current health state. The composite health feature vector is input into a fully connected mapping network. After feature mapping and parameter regression, charge and discharge constraint parameters adapted to the real-time health conditions of the battery are output. The charge and discharge constraint parameters include at least one of the following: upper limit of charge and discharge rate, upper limit of charge and discharge depth, and temperature operating range.

[0031] In this embodiment, adaptively generated charge and discharge constraint parameters are used as boundary control conditions to construct a closed-loop control logic for the charge and discharge operation of the energy storage battery. This closed-loop control logic is used in scenarios such as power quality regulation in distribution areas, photovoltaic power output stabilization, and reactive power compensation scheduling. It uses the charge and discharge constraint parameters as real-time control boundaries to constrain the battery's charge and discharge rate, charge and discharge depth, and operating temperature range. This allows the battery's actual operating parameters to dynamically match the safe operating boundary corresponding to its current health state, which is defined by the charge and discharge constraint parameters. Within each control cycle, it is determined whether the power command issued by the grid dispatch exceeds the allowable range defined by the charge and discharge constraint parameters. If it does, the dispatched power is limited according to a preset stepped derating strategy, causing the actual output power to gradually decrease until it falls within the allowable range. This achieves dynamic adaptation between the energy storage system control strategy and the battery aging state.

[0032] Based on the above steps, a real-time multi-dimensional health index, including battery health state estimates, internal resistance change rate, lithium plating risk parameters, and temperature rise anomaly parameters, is output through a SOH prediction model. This overcomes the limitation of a single SOH value in distinguishing different degradation modes, achieving a comprehensive multi-dimensional characterization of the energy storage battery's health state. A self-attention mechanism is used to fuse historical multi-dimensional health index sequences with real-time multi-dimensional health indices. This allows the real-time health index to adaptively focus on relevant time steps in the historical sequence, aggregating historical degradation trend information into the current moment's feature representation. The generated charge / discharge constraint parameters not only reflect the battery's current health state but also incorporate the future safety boundary change direction predicted by historical aging trajectories, improving the foresight of control decisions. By controlling the energy storage battery's charge / discharge actions based on these constraint parameters, a complete closed loop from health state assessment to charge / discharge execution control is achieved.

[0033] In the implementation of step S101, the remaining lifespan is predicted using the SOH prediction model based on the real-time operating data of the energy storage batteries in the distribution area, resulting in real-time multidimensional health indicators, including: The SOH prediction model includes a battery health state estimation branch, an internal resistance evolution branch, a lithium plating risk discrimination branch, and a thermal behavior analysis branch. Based on the battery health state estimation branch, feature extraction and decoding are performed on the time-series features of voltage, current and temperature in real-time operating data to obtain the battery health state estimate. And / or, based on the internal resistance evolution branch, feature extraction and decoding of the internal resistance time-series evolution component in the real-time running data are performed to obtain the internal resistance change rate; And / or, based on the lithium plating risk discrimination branch, feature extraction and decoding of the negative electrode potential time series component in the real-time operation data are performed to obtain the lithium plating risk parameters; And / or, based on the thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain temperature rise anomaly parameters.

[0034] In this embodiment, the battery health status estimation branch takes voltage time-series signal, current time-series signal and temperature time-series signal from real-time operating data as input, and captures multi-dimensional degradation information such as voltage change, current response characteristics and temperature drift trend through the time-series feature extraction network. It is then mapped to scalar value in the range of 0 to 1 through the decoding network, which serves as the battery health status estimate and is used to characterize the retention rate of the battery's current available capacity relative to the factory nominal capacity.

[0035] The internal resistance evolution branch takes the internal resistance time-series evolution component from real-time operating data as input. This component is obtained by identifying the response relationship between terminal voltage and current using a recursive least squares method. After feature extraction and decoding, the internal resistance change rate is output, which characterizes the ohmic internal resistance drift rate caused by battery polarization growth and active material dissolution.

[0036] The lithium plating risk assessment branch takes the negative electrode potential time-series component from real-time operating data as input. The negative electrode potential time-series component is estimated by using a three-electrode cell or equivalent circuit model. After feature extraction and decoding, a lithium plating risk parameter is output. This parameter is a continuous risk index in the range of 0 to 1. The higher the value, the closer the negative electrode potential is to the lithium plating critical potential, and the greater the probability of lithium plating.

[0037] The thermal behavior analysis branch takes the multi-point temperature spatiotemporal matrix from real-time operating data as input. This matrix is ​​formed by temperature sensors synchronously collected from multiple points located on the battery tabs, casing, and inside the module. The spatiotemporal feature extraction network captures the spatial heterogeneity and temporal evolution trend of the temperature field, outputting temperature rise anomaly parameters to characterize the risk level of local overheating or abnormal heat dissipation in the battery.

[0038] Based on the above implementation methods, a State of Health (SOH) prediction model is constructed, comprising four parallel branches: battery health state estimation, internal resistance evolution, lithium plating risk assessment, and thermal behavior analysis. This model generates real-time health indicators from four dimensions: capacity retention, impedance growth rate, lithium plating risk level, and abnormal temperature rise amplitude. This multi-branch architecture enables the decoupled extraction and simultaneous quantification of multi-mode battery degradation characteristics, overcoming the technical limitation that a single health indicator cannot comprehensively characterize the multi-field coupled aging effects of electrochemistry, heat, and safety.

[0039] Figure 2 This is a schematic flowchart illustrating the process of obtaining a battery health status estimate in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 2 As shown, based on the battery health state estimation branch, feature extraction and decoding are performed on the time-series features of voltage, current, and temperature in real-time operating data to obtain the battery health state estimate, including: Step S201: Extract and decode the time-series features of voltage, current and temperature in real-time operating data through the battery health state estimation branch to obtain the first health feature map; Step S202: Compress the first health feature map into a one-dimensional feature vector by global average pooling; Step S203: Input the one-dimensional feature vector into multiple fully connected layers for nonlinear transformation, and use an activation function to map the result of the nonlinear transformation to a battery health state estimate.

[0040] In this embodiment, the voltage, current, and temperature time-series signals from the real-time operating data are denoised by median filtering and then organized into a multi-channel time-series input tensor using a preset sliding window. This tensor is then input into a time-series feature extraction network, which consists of alternating stacked one-dimensional convolutional layers and max-pooling layers. The one-dimensional convolutional layers gradually abstract local time-series patterns by sliding the convolutional kernel along the time dimension, while the max-pooling layers gradually reduce the temporal resolution of the feature maps through downsampling operations with a preset step size. This allows the time-series feature extraction network to simultaneously acquire short-term transient fluctuation features and long-term aging trajectory features, achieving a comprehensive representation of degradation information across time scales and outputting a first health feature map.

[0041] In this embodiment, the number of channels in the first healthy feature map is a fixed hyperparameter preset by the network. A global average pooling operation is performed on the first healthy feature map to calculate the arithmetic mean of the temporal activation values ​​of each channel at all time steps. Each channel independently obtains a scalar, thereby compressing the feature map into a one-dimensional feature vector. This operation makes the input dimension of the subsequent fully connected layer a fixed value and is independent of the number of time steps of the feature map.

[0042] In this embodiment, a one-dimensional feature vector is sequentially input into multiple fully connected layers. Each fully connected layer is followed by a batch normalization layer, a ReLU activation function layer, and a Dropout layer. The batch normalization layer is used to accelerate convergence and stabilize the training process; the ReLU activation function layer introduces non-linear transformation capabilities into the network, enabling it to fit the complex mapping relationship between temporal features and battery health status; the Dropout layer randomly discards some neurons during the training phase to prevent overfitting. The final fully connected layer outputs a one-dimensional scalar, which is then mapped to the [0,1] interval using the Sigmoid function to obtain the battery health status estimate.

[0043] Based on the above steps, the battery health status estimation branch performs end-to-end feature extraction and health status mapping on the voltage, current and temperature time-series signals in the real-time operating data. This enables the battery health status estimate to integrate the co-evolution information of multiple physical quantities such as voltage decay, current response characteristics and temperature drift trend, thereby improving the accuracy and robustness of the battery health status estimate under complex operating conditions.

[0044] Figure 3 This is a schematic flowchart illustrating the process of obtaining the rate of change of internal resistance in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 3 As shown, based on the internal resistance evolution branch, feature extraction and decoding are performed on the internal resistance time-series evolution components in the real-time operating data to obtain the internal resistance change rate, including: Step S301: Extract and decode the internal resistance time-series evolution component in the real-time running data through the internal resistance evolution branch to obtain the second health feature map. Step S302: The internal resistance change trend is predicted for each time step feature in the second health feature map by using a temporal convolutional layer, and the internal resistance change trend is decoded by inverse normalization to obtain the internal resistance change rate.

[0045] In this embodiment, the real-time operating data of the energy storage battery in the distribution area is decomposed into components by an internal resistance evolution branch, separating the internal resistance time-series evolution component to characterize the degradation characteristics of the battery's internal impedance. The internal resistance time-series evolution component is characterized by slow degradation, strong temporal continuity, and susceptibility to interference from operating condition coupling. The internal resistance evolution branch extracts long-term decay trends, short-term fluctuations, and abrupt changes from the internal resistance time-series evolution component through a multi-layer feature extraction and decoding structure, and filters out interference information caused by charge / discharge rate fluctuations, ambient temperature disturbances, and acquisition noise, thereby eliminating feature aliasing caused by multi-parameter coupling. After layer-by-layer feature transformation and dimensionality regularization, a second health feature map characterizing the battery's internal resistance aging behavior is output. This second health feature map retains the temporal evolution details and degradation characteristics of the internal resistance at different time scales.

[0046] In this embodiment, the second health feature map is input into a preset temporal convolutional layer. The temporal convolutional layer performs convolution operations and temporal correlation modeling on the internal resistance features at each time step, mines the evolution correlation of internal resistance features between adjacent time steps, and predicts the changing trend of battery internal resistance. Based on the normalization parameters recorded in the normalization preprocessing stage, inverse normalization decoding is performed on the predicted internal resistance changing trend results, mapping the standardized values ​​to the real physical value range, obtaining the internal resistance change rate characterizing the rate of battery internal resistance degradation, and completing the real-time quantitative identification of the battery internal resistance degradation state.

[0047] Based on the above steps, feature extraction and decoding of the internal resistance time-series evolution component are performed through the internal resistance evolution branch. This separates the dominant degradation feature characterizing battery impedance growth from the original internal resistance signal, while suppressing coupling interference caused by fluctuations in parameters such as voltage, current, and temperature. This allows the extracted second health feature map to reflect the long-term decay trend, short-term fluctuation characteristics, and abnormal abrupt changes in internal resistance, improving the effectiveness of internal resistance features in representing the aging state. A temporal convolutional layer performs convolution operations on the internal resistance features at each time step, uncovering the evolutionary correlation between adjacent time steps and outputting the predicted internal resistance change trend. The inverse normalization decoding operation, based on the normalization parameters recorded during the training phase, maps the predicted results from a standardized dimensionless numerical range to the real physical numerical range, giving the output internal resistance change rate a clear physical meaning and dimension for subsequent safety boundary determination.

[0048] Figure 4 This is a schematic flowchart illustrating the process of obtaining lithium plating risk parameters in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 4As shown, based on the lithium plating risk discrimination branch, feature extraction and decoding are performed on the negative electrode potential time series components in the real-time operating data to obtain lithium plating risk parameters, including: Step S401: The negative electrode potential time series component in the real-time operation data is extracted and decoded through the lithium plating risk discrimination branch to obtain the third health feature map; Step S402: The lithium plating sensitive features in the third health feature map are enhanced by an attention weighting layer, and the enhanced features are mapped to lithium plating risk level or lithium plating risk probability value by a classification output layer to obtain lithium plating risk parameters.

[0049] In this embodiment, a lithium plating risk discrimination branch is used to analyze the time-series component of the negative electrode potential in the real-time battery operating data. The time-series evolution characteristics of the negative electrode potential are used to characterize the stability of the battery's negative electrode interface and serve as an electrochemical criterion for judging lithium dendrite deposition behavior. The lithium plating risk discrimination branch performs feature decoding on the negative electrode potential time sequence through a multi-layer time-series feature extraction structure, extracting precursor features of lithium plating such as a negative electrode potential that is continuously below the lithium plating threshold, a sudden drop in potential at the end of charge and discharge, and time-series drift distortion, while filtering out interference components composed of grid operating condition fluctuations, sampling errors, and conventional polarization fluctuations. After feature recombination and dimensionality normalization, a third health feature map is output to characterize the battery's negative electrode interface degradation state and potential lithium plating risk.

[0050] In this embodiment, the third health feature map is input into the attention-weighted layer. The attention-weighted layer calculates the weight distribution of the feature map in the time dimension through a self-attention mechanism, assigning high weights to regions with abnormal negative electrode potentials and low weights to steady-state regions, so that the model focuses on the time regions where lithium plating precursor features are significant. The weight distribution is multiplied element-wise with the third health feature map to obtain the attention-enhanced feature. The attention-enhanced feature is input into the fully connected mapping layer, and after nonlinear transformation and dimensionality compression, a quantified lithium plating risk parameter is output. The lithium plating risk parameter is a continuous probability value ranging from 0 to 1, representing the degree of risk of lithium dendrite precipitation at the current moment. The higher the value, the greater the risk of lithium plating.

[0051] Based on the above steps, the lithium plating risk discrimination branch uses the negative electrode potential time-series component as input. This component reflects the electrochemical state at the battery's negative electrode interface, and its value is related to the thermodynamic conditions for lithium ion reduction and deposition on the negative electrode surface. Specifically, when the negative electrode potential is below the critical lithium plating potential, the thermodynamic prerequisite for lithium plating is met. A multi-layer feature extraction structure is used to decode the negative electrode potential time-series component, extracting the trend of a persistently low negative electrode potential on a long-term timescale, the instantaneous potential drop under high-rate conditions on a short-term timescale, and the potential recovery lag after the drop during the time-series recovery phase. This establishes a correlation between these features and the lithium dendrite deposition mechanism at the electrochemical mechanism level, weakening feature aliasing caused by multi-parameter coupling, reducing the interference of normal operating condition fluctuations on lithium plating discrimination, and improving the accuracy of lithium plating risk feature extraction.

[0052] Figure 5 This is a schematic flowchart illustrating the process of obtaining abnormal temperature rise parameters in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 5 As shown, based on the thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain the following temperature rise anomaly parameters: Step S501: The feature extraction and decoding of the multi-point temperature spatiotemporal matrix in the real-time operation data is performed through the thermal behavior analysis branch to obtain the fourth health feature map. Step S502: The temperature rise offset is predicted for each spatial location in the fourth health feature map by a regression convolutional layer. The temperature rise offset is then decoded and corrected by temperature compensation to obtain the abnormal temperature rise parameters.

[0053] In this embodiment, continuous temperature data from multiple measurement points in the energy storage battery module, including the battery casing, cell surface, and busbar location, are collected to construct a multi-point temperature spatiotemporal matrix containing both spatial location information and temporal evolution information. This temperature spatiotemporal matrix is ​​input into a thermal behavior analysis branch network, where a spatiotemporal feature extraction structure performs feature decoding operations. This process extracts thermal degradation features such as localized overheating, excessive cell temperature difference, abnormal temperature rise rate, and long-term thermal drift in both spatial and temporal dimensions, while filtering out interference components from periodic fluctuations in ambient temperature and acquisition noise. After multi-layer feature fusion and dimensional normalization, a fourth health feature map characterizing the overall thermal behavior state of the battery is output.

[0054] In this embodiment, the fourth health feature map is input into a regression convolutional layer. The regression convolutional layer performs temperature rise offset mapping on the spatial measurement point features in the fourth health feature map through convolution operations, predicting the real-time temperature rise offset value at each measurement point. Dimensional inverse mapping is performed on the temperature rise offset value to restore the normalized predicted value to the physical quantity of temperature, obtaining the estimated temperature rise offset value for each measurement point. Based on the estimated temperature rise offset value for each measurement point, the maximum temperature rise offset, the average temperature rise offset, and the local extreme temperature difference are extracted. These three statistical quantities are combined into a temperature rise anomaly parameter, which is used to characterize the current overall temperature rise anomaly degree and local thermal imbalance risk of the energy storage battery.

[0055] Based on the above steps, a multi-point temperature spatiotemporal matrix containing spatial location information and temporal evolution information is constructed to achieve full-domain perception of the temperature field of the energy storage battery module, overcoming the limitation of single-point temperature measurement in identifying local thermal defects. Through the spatiotemporal feature extraction structure of the thermal behavior analysis branch, thermal degradation features such as local overheating, excessive cell temperature difference, abnormal temperature rise rate, and long-term thermal drift are extracted in both spatial and temporal dimensions. Simultaneously, interference components such as periodic fluctuations in ambient temperature and acquisition noise are filtered out, achieving a multimodal comprehensive characterization of battery thermal behavior. By using regression convolutional layers to predict the temperature rise offset at each spatial location in the fourth health feature map, differentiated quantitative assessment of the thermal state of different cells and measurement points is achieved.

[0056] Figure 6 This is a schematic flowchart illustrating the process of obtaining the charging and discharging constraint parameters of an energy storage battery in an intelligent control method for a power grid energy storage battery according to an embodiment of the present invention. Figure 6 As shown, by fusing historical multidimensional health indicator sequences and real-time multidimensional health indicators using a self-attention mechanism, the charge and discharge constraint parameters of the energy storage battery are obtained, including: Step S601: The real-time multidimensional health indicator is appended to the last time step of the historical multidimensional health indicator sequence as the query vector of the current moment, to construct a complete input sequence containing information of the historical moment and the current moment. The position is encoded for each time step in the complete input sequence to obtain an encoded sequence carrying the temporal position. Step S602: The feature vector of each time step in the encoded sequence is mapped to a query vector, a key vector, and a value vector through a query projection layer, a key projection layer, and a value projection layer, respectively. The query matrix, the key matrix, and the value matrix are obtained based on the query vector, the key vector, and the value vector. The query projection layer, the key projection layer, and the value projection layer are each independent linear mapping networks. Step S603: Perform multi-head self-attention calculation on the query matrix, key matrix, and value matrix to obtain the self-attention enhanced output sequence; Step S604: Extract the output vector corresponding to the last time step of the real-time multidimensional health indicator from the self-attention enhanced output sequence, and use it as a composite health feature vector that integrates historical decay trend information and current health status information. Step S605: Input the composite health feature vector into the fully connected mapping network. The output layer of the fully connected mapping network contains the same number of neurons as the charge-discharge constraint parameters, and each neuron outputs a charge-discharge constraint parameter.

[0057] In this embodiment, the features of each time step in the complete input sequence are positionally encoded, and each time step is assigned a temporal position identifier to obtain an encoded sequence carrying complete temporal position information. By using multiple independent attention heads to perform attention score calculation and feature weighted fusion in parallel, the long-distance temporal dependencies and long-term aging decay patterns of battery health indicator sequences are mined from different feature subspaces. The attention weights of each historical time-series node are adaptively allocated to weaken the interference of invalid historical information and strengthen key degradation temporal features, resulting in an attention-enhanced output sequence.

[0058] In this embodiment, the output vector at the final time step integrates historical health decay trends and current real-time health status information through global temporal attention weighted fusion, achieving deep coupling between historical temporal trend features and current instantaneous state features, and defining it as a composite health feature vector. The number of neurons in the output layer of the fully connected mapping network is the same as the number of types of charge / discharge constraint parameters to be output, with each output neuron corresponding to one type of charge / discharge constraint parameter. Through nonlinear feature transformation and dimensional mapping of the multi-layer fully connected network, the upper limit of charge / discharge rate, the upper limit of charge / discharge depth, and the temperature operating range are output simultaneously, realizing adaptive constraint control of the energy storage battery based on temporal health features.

[0059] Based on the above steps, by appending real-time multidimensional health indicators as query vectors to the end of the historical sequence and performing self-attention calculation on the complete sequence, the current health status can establish a direct attentional association with the degradation information of all time steps in the historical sequence. Compared to relying on data from the most recent few time steps for prediction, this invention can adaptively capture the historical degradation stage most similar to the current state, thereby improving the accuracy and adaptability of the prediction of charge and discharge constraint parameters.

[0060] In the implementation of step S101, controlling the energy storage battery to perform charging and discharging actions based on charging and discharging constraint parameters includes: Based on the preset charging and discharging constraint parameters-execution command mapping relationship, the upper limit of charging and discharging rate, the upper limit of charging and discharging depth, and the temperature operating range are converted into control parameters for the energy storage battery. The control parameters include the upper limit of current amplitude, the voltage protection threshold, and the heat dissipation start and stop threshold.

[0061] In this embodiment, the upper limit of the charge / discharge rate is converted into an upper limit of the current amplitude based on a preset mapping relationship between charge / discharge constraint parameters and execution commands. Based on the nominal rated capacity of the energy storage battery, and according to the physical relationship between current and rate: Current = Nominal Rated Capacity × Upper Limit of Charge / Discharge Rate, the upper limit of the current amplitude is determined. This obtained upper limit of the current amplitude is used as the BMS charge / discharge current limiting command to constrain the maximum allowable amplitude of the converter output current.

[0062] In this embodiment, the upper limit of charge / discharge depth is converted into a voltage protection threshold. Based on the state-of-charge-open-circuit voltage (SOC-OCV) mapping curve of the energy storage battery, the corresponding voltage value is obtained from a table at the SOC boundary point corresponding to the upper limit of charge / discharge depth. For the upper limit of charge depth, the charging cut-off voltage (corresponding to the upper limit of SOC) is obtained; for the upper limit of discharge depth, the discharging cut-off voltage (corresponding to the lower limit of SOC) is obtained. The charging cut-off voltage and the discharging cut-off voltage serve as the BMS voltage protection threshold, which are used to trigger overcharge or over-discharge protection actions when the battery terminal voltage reaches the threshold.

[0063] In this embodiment, the temperature operating range is converted into heat dissipation start / stop thresholds. The temperature operating range includes a lower limit temperature and an upper limit temperature. Based on a preset hysteresis parameter (e.g., hysteresis parameter = 5℃), the heat dissipation start temperature and heat dissipation stop temperature are determined, where the heat dissipation start temperature = upper limit temperature - hysteresis parameter, and the heat dissipation stop temperature = lower limit temperature - 2 × hysteresis parameter. Alternatively, direct mapping can be performed based on the thermal management parameter table provided by the battery manufacturer. The heat dissipation start temperature and heat dissipation stop temperature serve as BMS thermal management control commands to control the start / stop of the cooling fan, the start / stop of the coolant circulation pump, and the adjustment of cooling power, thereby confining the actual operating temperature of the battery within the temperature operating range.

[0064] Based on the above implementation method, by using a preset charging / discharging constraint parameter-execution command mapping relationship, the upper limit of charge / discharging rate, the upper limit of charge / discharging depth, and the temperature operating range are converted into control parameters executable by the BMS, such as the upper limit of current amplitude, the voltage protection threshold, and the heat dissipation start / stop threshold, thus realizing a closed-loop mapping from health status assessment to charge / discharge execution control. Through a rate-to-current conversion mechanism based on the battery's rated capacity, the upper limit of current amplitude is adaptively adjusted according to changes in the battery's health status. Through a depth-to-voltage conversion mechanism based on the SOC-OCV mapping curve, the voltage protection threshold is adaptively adjusted according to changes in the battery's health status. Through a temperature range-to-heat dissipation start / stop threshold conversion mechanism based on hysteresis parameters, the start / stop control of the heat dissipation equipment is linked to the battery's currently allowed temperature operating range.

[0065] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in every case. Furthermore, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional variations can be made to the above method.

[0066] It should be understood that in some embodiments, the components may be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods may be implemented using software or firmware stored in memory and executed by a suitable instruction execution system.

[0067] This embodiment also provides a computer program product 10, a computer-readable storage medium 20, and a computer device 30. Figure 7 This is a schematic diagram of a computer program product 10 according to an embodiment of the present invention. Figure 8 This is a schematic diagram of a computer-readable storage medium 20 according to an embodiment of the present invention. Figure 9 This is a schematic diagram of a computer device 30 according to an embodiment of the present invention. Figure 7 As shown, the computer program product 10 includes a computer program 11, which, when executed by the processor 32, implements the steps of any of the above-described intelligent control methods for substation energy storage batteries. For example... Figure 8 As shown, a computer-readable storage medium 20 stores the aforementioned computer program 11, which, when executed by the processor 32, implements the steps of the intelligent control method for the substation energy storage battery of any of the above embodiments. Figure 9 As shown, the computer device 30 may include a memory 31, a processor 32, and a computer program 11 stored on the memory 31 and running on the processor 32.

[0068] The computer program 11 used to perform the operations of this invention may be assembly instructions, Instruction Set Architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages. The computer program 11 may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computer. In some embodiments, to perform aspects of this invention, electronic circuits, including, for example, programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), may execute computer-readable program instructions to personalize the electronic circuits by utilizing status information of the computer-readable program instructions.

[0069] For the purposes of this embodiment, computer program product 10 is a related product that includes computer program 11.

[0070] For the purposes of this embodiment, computer-readable storage medium 20 is a tangible device capable of holding and storing a computer program 11. It can be any device capable of containing, storing, communicating, propagating, or transmitting the program 11 for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage medium 20 include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical encoding device, and any suitable combination thereof.

[0071] Computer device 30 can be, for example, a server, desktop computer, laptop computer, tablet computer, or smartphone. In some examples, computer device 30 can be a cloud computing node. Computer device 30 can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., that perform specific tasks or implement specific abstract data types. Computer device 30 can be implemented in a distributed cloud computing environment where tasks are performed by remote processing devices linked through a communication network. In a distributed cloud computing environment, program modules can reside on local or remote computing system storage media, including storage devices.

[0072] Computer device 30 may include a processor 32 adapted to execute stored instructions and a memory 31 that provides temporary storage space for the operation of said instructions during operation. The processor 32 may be a single-core processor, a multi-core processor, a computing cluster, or any other configuration. The memory 31 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0073] Computer device 30 may also include a network adapter / interface and an input / output (I / O) interface. The I / O interface allows external devices that can be connected to the computer device to input and output data. The network adapter / interface provides communication between the computer device and a network, typically represented as a communication network.

[0074] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A smart control method for a transformer substation energy storage battery, characterized in that, include: The remaining life of energy storage batteries in a distribution area is predicted using a SOH prediction model based on real-time operating data. This yields real-time multidimensional health indicators, including at least one of the following: battery health status estimate, internal resistance change rate, lithium plating risk parameter, and abnormal temperature rise parameter. The SOH prediction model is trained using federated learning on historical operating data labeled with battery health status. During training, a loss function is used to optimize the model parameters. This loss function is a weighted sum of at least one of the following: battery health status estimation loss, internal resistance evolution loss, lithium plating risk discrimination loss, and thermal behavior analysis loss. The battery health status estimation loss is determined using a mean square error loss function or a mean absolute error loss function. The internal resistance evolution loss is determined using a logarithmic mean square error loss function or a Huber loss function. The lithium plating risk discrimination loss is determined using a weighted cross-entropy loss function or a Focal loss function. The thermal behavior analysis loss is determined using a composite loss function combining the mean square error loss function and a temperature compensation correction term. The weighting coefficients of the weighted sum are preset according to the actual needs of the distribution area energy storage battery remaining life prediction task. The historical multidimensional health indicator sequence and the real-time multidimensional health indicator are fused under a self-attention mechanism to obtain the charge and discharge constraint parameters of the energy storage battery; the historical multidimensional health indicator sequence is a time series data formed by arranging the multidimensional health indicators recorded and stored by the energy storage battery in the past operating cycle in chronological order; the charge and discharge constraint parameters include at least one of the following: upper limit of charge and discharge rate, upper limit of charge and discharge depth, and temperature operating range. The energy storage battery is controlled to perform charging and discharging actions based on the charging and discharging constraint parameters.

2. The intelligent control method for substation energy storage batteries according to claim 1, characterized in that, The remaining life (SOH) prediction model is used to predict the remaining life of energy storage batteries in the distribution area based on real-time operating data, resulting in real-time multidimensional health indicators, including: The SOH prediction model includes a battery health state estimation branch, an internal resistance evolution branch, a lithium plating risk discrimination branch, and a thermal behavior analysis branch. Based on the battery health state estimation branch, the voltage, current and temperature time-series features in the real-time operating data are extracted and decoded to obtain the battery health state estimate. And / or, based on the internal resistance evolution branch, feature extraction and decoding are performed on the internal resistance time-series evolution component in the real-time running data to obtain the internal resistance change rate; And / or, based on the lithium plating risk discrimination branch, feature extraction and decoding are performed on the negative electrode potential time series component in the real-time operating data to obtain the lithium plating risk parameter; And / or, based on the thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain the temperature rise anomaly parameters.

3. The intelligent control method for the energy storage battery in the distribution area according to claim 2, characterized in that, Based on the battery health state estimation branch, feature extraction and decoding are performed on the voltage, current, and temperature time-series features in the real-time operating data to obtain the battery health state estimate, including: The battery health state estimation branch extracts and decodes the time-series features of voltage, current and temperature in the real-time operating data to obtain the first health feature map. The first health feature map is compressed into a one-dimensional feature vector by global average pooling. The one-dimensional feature vector is input into multiple fully connected layers for nonlinear transformation, and the result of the nonlinear transformation is mapped to the battery health state estimate through an activation function.

4. The intelligent control method for the energy storage battery in the distribution area according to claim 2, characterized in that, Based on the internal resistance evolution branch, feature extraction and decoding are performed on the internal resistance time-series evolution components in the real-time operating data to obtain the internal resistance change rate, including: The internal resistance time-series evolution component in the real-time running data is extracted and decoded by the internal resistance evolution branch to obtain the second health feature map. The internal resistance change trend is predicted by using a temporal convolutional layer for each time step feature in the second health feature map, and the internal resistance change trend is decoded by inverse normalization to obtain the internal resistance change rate.

5. The intelligent control method for substation energy storage batteries according to claim 2, characterized in that, Based on the lithium plating risk discrimination branch, feature extraction and decoding are performed on the negative electrode potential time-series component in the real-time operating data to obtain the lithium plating risk parameters, including: The negative electrode potential time-series component in the real-time operating data is extracted and decoded by the lithium plating risk discrimination branch to obtain the third health feature map. The lithium plating-sensitive features in the third health feature map are enhanced by an attention-weighted layer, and the enhanced features are mapped to lithium plating risk level or lithium plating risk probability value by a classification output layer to obtain the lithium plating risk parameter.

6. The intelligent control method for substation energy storage batteries according to claim 2, characterized in that, Based on the aforementioned thermal behavior analysis branch, feature extraction and decoding are performed on the multi-point temperature spatiotemporal matrix in the real-time operating data to obtain the temperature rise anomaly parameters, including: The thermal behavior analysis branch is used to extract and decode the features of the multi-point temperature spatiotemporal matrix in the real-time operation data to obtain the fourth health feature map. The temperature rise offset is predicted for each spatial location in the fourth health feature map by a regression convolutional layer, and the temperature rise offset is decoded and corrected by temperature compensation to obtain the temperature rise anomaly parameters.

7. The intelligent control method for substation energy storage batteries according to claim 1, characterized in that, By fusing the historical multidimensional health indicator sequence and the real-time multidimensional health indicator under a self-attention mechanism, the charge and discharge constraint parameters of the energy storage battery are obtained, including: The real-time multidimensional health indicator is appended to the last time step of the historical multidimensional health indicator sequence as the current time query vector to construct a complete input sequence containing historical and current time information. Each time step in the complete input sequence is position-encoded to obtain an encoded sequence carrying the temporal position. The feature vector at each time step in the encoded sequence is mapped to a query vector, a key vector, and a value vector through a query projection layer, a key projection layer, and a value projection layer, respectively. A query matrix, a key matrix, and a value matrix are obtained based on the query vector, the key vector, and the value vector. The query projection layer, the key projection layer, and the value projection layer are each independent linear mapping networks. Multi-head self-attention computation is performed on the query matrix, key matrix, and value matrix to obtain a self-attention enhanced output sequence; Extract the output vector corresponding to the last time step of the real-time multidimensional health indicator from the self-attention enhanced output sequence, and use it as a composite health feature vector that integrates historical decay trend information and current health status information. The composite health feature vector is input into a fully connected mapping network. The output layer of the fully connected mapping network contains the same number of neurons as the charge-discharge constraint parameters, and each neuron outputs a charge-discharge constraint parameter.

8. The intelligent control method for a distribution area energy storage battery according to claim 1, characterized in that, Controlling the energy storage battery to perform charging and discharging actions based on the charging and discharging constraint parameters includes: According to the preset charging and discharging constraint parameter-execution command mapping relationship, the upper limit of the charging and discharging rate, the upper limit of the charging and discharging depth, and the temperature operating range are converted into control parameters for the energy storage battery. The control parameters include the upper limit of the current amplitude, the voltage protection threshold, and the heat dissipation start / stop threshold.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent control method for the energy storage battery in the distribution area according to any one of claims 1 to 8.

Images