Multi-physics field coupled virtual power plant energy storage health state monitoring and predicting method

Abstract

The invention belongs to the technical field of virtual power plants, and particularly relates to a multi-physics field coupled virtual power plant energy storage health state monitoring and predicting method, which comprises the following steps: acquiring electrochemical parameters and thermodynamic parameters of an energy storage facility in real time, preprocessing the acquired data, and storing the preprocessed data into a database; calculating the current SOH of the energy storage facility, constructing a physical information neural network (PINN), and predicting the SOH of the energy storage facility based on the current SOH and historical data in the database; displaying a curve graph of data acquired in real time, the current SOH, an SOH prediction chart and alarm information in a visual mode; and with maximization of SOH and minimization of energy loss and thermal risk as targets, a Pareto optimization problem is constructed, and an optimal charging and discharging strategy is solved. According to the method, multi-parameter perception, physical mechanism and deep learning are fused, and high-precision real-time monitoring and prediction of the energy storage health state of the virtual power plant can be realized.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant technology, specifically relating to a method for monitoring and predicting the health status of virtual power plant energy storage using multi-physics coupling. Background Technology

[0002] With the large-scale grid connection of renewable energy and the rapid development of the electric vehicle industry, the role of virtual power plant (VPS) energy storage forecasting in the power system is becoming increasingly prominent. VPS, with its ability to aggregate diverse resources such as distributed energy storage, electric vehicles, and adjustable loads, plays a crucial role in scenarios such as grid peak and frequency regulation, renewable energy consumption, and electric vehicle energy management. However, VPS energy storage faces complex operating challenges during operation, such as frequent charge-discharge cycles of multiple devices accelerating battery aging, differences in operating environments among different devices increasing the difficulty of temperature control, and high-dynamic load aggregation and scheduling exacerbating the non-uniformity of the system's internal electrochemical state. Against this backdrop, real-time status monitoring of VPS energy storage is crucial for ensuring its safe operation, achieving intelligent operation and maintenance, and full lifecycle management.

[0003] Currently, virtual power plant energy storage status monitoring mainly adopts a layered and collaborative monitoring architecture. At the device level, the battery management system (BMS) is used to monitor the basic parameters of individual cells; at the edge, distributed monitoring units are used to achieve regional data collection and processing; and in the cloud, a unified management platform performs global status analysis and evaluation.

[0004] However, existing technologies have many limitations. On the one hand, traditional monitoring methods often focus on threshold monitoring of single parameters, such as judging over-discharge by voltage drops or triggering overheat protection based on temperature thresholds. This monitoring approach is difficult to comprehensively capture the complex coupling effects (such as the interaction between electrochemical-thermodynamic-mechanical stress) between various types of energy storage devices, diverse loads, and distributed power sources in a virtual power plant, leading to large errors in equipment health status (SOH) assessment. On the other hand, existing technologies are insufficient in data processing capabilities, mainly using linear filtering or simple statistical algorithms to process data. They are not robust to noise interference and sensor drift in the massive heterogeneous data of virtual power plants, and cannot effectively extract implicit features from high-dimensional data. Furthermore, traditional methods perform poorly in terms of dynamic adaptability. In highly dynamic scenarios such as virtual power plants participating in real-time frequency regulation of the power grid and responding to peak loads, parameters such as current and temperature of energy storage devices will fluctuate rapidly. Traditional methods, due to algorithmic lag (such as calculating the average value within a fixed time window), cannot capture transient anomalies in a timely manner, which can easily lead to false alarms or missed alarms. Moreover, they lack the ability to predict degradation trends based on historical data and optimize control schemes, making it difficult to meet the needs of preventive maintenance and collaborative scheduling.

[0005] Furthermore, existing technologies typically rely on monitoring a single or a few parameters, such as voltage, current, and temperature, neglecting or underestimating the effects of multi-physics coupling. In actual battery operation, complex interactions exist between electrical, thermal, and mechanical fields. For example, temperature changes affect the battery's internal resistance and electrochemical reaction rate, while changes in current load cause heat accumulation and mechanical stress changes within the battery. Poor interpretability and difficulty in integrating with physical mechanisms mean that machine learning models (such as neural networks) are often considered black boxes, with their internal mechanisms difficult to interpret, leading to low user trust in the model results. Existing methods fail to closely integrate with the battery's physical mechanisms (such as electrochemical reactions and thermodynamic behavior), resulting in poor reliability and adaptability of the models in practical applications. High cost and complexity also limit the widespread application of high-precision sensor-based monitoring systems, despite their ability to provide high accuracy. The high cost of hardware such as high-precision sensors and high-speed data acquisition equipment hinders their adoption in low-cost scenarios. Digital twin-based monitoring systems require complex software and hardware support, increasing the difficulty of system deployment and maintenance.

[0006] In summary, virtual power plant energy storage status monitoring technology urgently needs improvement in areas such as multi-physics coupling, data processing capabilities, dynamic adaptability, and prediction and decision support to meet the reliability and security requirements of complex virtual power plant application scenarios. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method for monitoring and predicting the health status of virtual power plant energy storage by multi-physics coupling, which integrates multi-parameter sensing, physical mechanism and deep learning, and can realize high-precision real-time monitoring and prediction of the health status of virtual power plant energy storage.

[0008] To achieve the above objectives, this invention provides a method for monitoring and predicting the health status of virtual power plant energy storage using multi-physics coupling, comprising the following steps: S1. Deploy a distributed sensor network in the energy storage facility of the virtual power plant to collect the electrochemical and thermodynamic parameters of the energy storage facility in real time, preprocess the collected data, and store the preprocessed data in the database. S2. Calculate the current SOH of the energy storage facility, and construct a physical information neural network PINN based on LSTM, physical model and joint loss function. Based on the current SOH and historical data in the database, predict the SOH of the energy storage facility, compare the prediction result with the current calculation result, and determine whether to issue an alarm. S3. Display real-time collected data in a visual manner, including graphs, current SOH, SOH prediction charts, and alarm information; S4. With the goal of maximizing SOH and minimizing energy loss and thermal risk, construct a Pareto optimization problem, solve for the optimal charging and discharging strategy, and send the optimal charging and discharging strategy to the battery management system (BMS) in real time to dynamically adjust the operating strategy.

[0009] As a preferred embodiment of the present invention, in S1, the electrochemical parameters include voltage, current, internal resistance, and state of charge, and the thermodynamic parameters include battery surface / internal temperature and heat flux density. The specific process of preprocessing is as follows: S1.1. Anomaly data is identified by combining dynamic thresholding with the isolated forest algorithm. In the dynamic thresholding method, the threshold is dynamically calculated and adjusted using the 3σ criterion. When the proportion of abnormal data exceeds the set alarm threshold, the alarm mechanism is triggered to notify maintenance personnel to conduct an inspection. When the proportion of abnormal data does not exceed the set alarm threshold, data correction is performed, using interpolation, historical data filling, or sliding window averaging to repair the abnormal data. S1.2 Perform spatiotemporal alignment on the repaired data, including spatial alignment and time alignment. Spatial alignment is achieved based on the Kriging interpolation algorithm, and time alignment is achieved by using timestamp calibration to eliminate the time series deviation of multi-source data. S1.3 Construct the spatiotemporally aligned data into tensor form: ; In the formula, Represents the constructed tensor; V is voltage; I is current; Indicates the surface temperature of the battery; Indicates the internal temperature of the battery; Indicates internal resistance; Indicates heat flux density; Let N represent an N×6 dimensional real space, where N represents the number of samples; S1.4, For Multi-scale features are extracted by wavelet transform, and dimensionality reduction is performed by principal component analysis (PCA). Kalman filtering and wavelet transform are then used for filtering and denoising. S1.5. Normalize the data processed in S1.4 to complete the data preprocessing.

[0010] As a preferred embodiment of the present invention, in S2, the architecture of PINN includes: The LSTM temporal feature extraction layer employs a dual-hidden-layer structure. The input is historical data from a database, and the output is the hidden state h. t ; The physical mechanism embedding layer directly embeds the physical model, which includes both electrochemical and thermodynamic models, as hard constraints into PINN. Specifically, the residual terms of the solid-phase diffusion equation in the electrochemical model... Constraints on SOC predictions, constraints on overpotentials using the Butler-Volmer equation, and residual terms in the heat balance equation of the thermodynamic model. Limiting the temperature prediction range, surface heat transfer boundary conditions ensure that the temperature gradient conforms to the actual heat dissipation law; The joint loss function L is expressed as: ; In the formula, , The weighting coefficients are determined through Bayesian optimization. Indicates mean square error; Represents the L2 norm; Indicates the predicted value; Represents the actual value.

[0011] As a preferred embodiment of the present invention, the electrochemical model specifically includes: Solid-phase diffusion equation based on Fick's second law; The internal boundary conditions are spherically symmetric constraints: ; In the formula, c s Indicates lithium ion concentration; r represents the radial position of the particle; Surface boundary conditions, correlation between lithium-ion flux and reaction current density: ; In the formula, D s ρ is the solid-phase lithium diffusion coefficient; j is the surface reaction current density; a is the specific surface area; F represents the Faraday constant; R represents the particle radius; The surface reaction follows the Butler-Volmer equation; The thermodynamic model specifically includes: Based on the law of conservation of energy, the heat balance equation of a battery is expressed as: ; In the formula, m represents the battery mass; c p T represents the battery's specific heat capacity; T represents the battery temperature; t represents time. This represents the total heat generated inside the battery. This represents the total heat lost by the battery through convection, conduction, and radiation. The internal boundary is represented as: ; The outer boundary is represented as: ; In the formula, h represents the convective heat transfer coefficient; The ambient temperature is represented by k, which represents the thermal conductivity of the battery material.

[0012] As a preferred embodiment of the present invention, in PINN, during the forward propagation phase, the LSTM temporal feature extraction layer outputs the hidden state h. t Then, h t Input is fed into the physical mechanism embedding layer to solve. and During the backpropagation phase, and Participating in gradient calculation forces the PINN network parameters to satisfy data fitting and physical laws while updating; The predicted result is compared with the current calculation result, and the difference or error percentage between the two is calculated. Based on the set error threshold, it is determined whether an alarm is required. If the error exceeds the corresponding error threshold, an alarm signal is issued to remind maintenance personnel to check. Otherwise, the visualization display stage is entered.

[0013] As a preferred embodiment of the present invention, a dynamic adaptive training strategy is constructed for staged training of PINN, specifically as follows: Pre-trained LSTM: uses only the data loss term, i.e. Preliminary learning of temporal characteristics; Physical constraint fine-tuning: gradually increasing This ensures that the prediction results gradually conform to the laws of physics; Online update mechanism: The model is fine-tuned every 24 hours based on the latest data, and a flexible weight solidification algorithm is used to retain key physical features; When comparing the predicted results with the current calculation results, if the error exceeds the corresponding error threshold, the online retraining of PINN is triggered simultaneously, and the parameters of the electrochemical model and the thermodynamic model are updated.

[0014] As a preferred embodiment of the present invention, a transfer learning-driven physical parameter decoupling mechanism is introduced into PINN, the process of which is as follows: Step 1: Calculate in real time the gradient sensitivity of the physical model parameters embedded in PINN to the joint loss function L. , is represented as: ; In the formula, This is a set of physical model parameters, including the solid-phase lithium diffusion coefficient D. s Convection heat transfer coefficient h, surface reaction current density j; For gradient operators, Indicates L pair The gradient vector; Step 2: When there are K consecutive training cycles When the gradient threshold is exceeded, it is determined to be a physical parameter drift, triggering transfer learning. Freeze all weight parameters of the LSTM temporal feature extraction layer; Building a lightweight migration network It includes an input layer, an intermediate layer, and an output layer, wherein the intermediate layer includes a BN layer and a ReLU activation function; Will The output is reconstructed into physical residual terms: ; In the formula, , These are the weighting coefficients; Represents the physical residual term of the reconstruction; Step 3: Add a migration stability term to L, expressed as: ; In the formula, Indicate the new joint loss function; This is the migration regularization strength coefficient; Indicates KL divergence; This represents the distribution of historical physical parameters.

[0015] As a preferred embodiment of the present invention, in S3, during the visualization process, the remaining number of cycles before the predicted capacity decays to 80% and the root cause analysis of failure are also displayed. The remaining number of cycles before the predicted capacity decay to 80% is obtained by first obtaining the lithium loss caused by the SEI layer based on the SEI growth kinetics equation, and then calculating the total capacity change. ; In the formula, This represents the battery capacity at time t; Q0 represents the initial capacity. represents the lithium loss caused by the SEI layer at time t; q and n are empirical fitting parameters, representing the decay rate and the degree of nonlinearity of the decay curve, respectively; Q represents the decay term adjustment coefficient, used to control the contribution of the SEI layer to the total capacity change in the capacity decay model. The SOH sequence output by PINN is input into a particle filter, and a state-space model is constructed by combining the SEI growth kinetic equation and battery capacity change to predict the remaining number of cycles until the capacity decays to 80%. The root cause analysis of the failure was to locate the dominant factor of attenuation through the physical residual terms of PINN.

[0016] As a preferred embodiment of the present invention, in S3, during the visualization display, optimization suggestions are generated based on the displayed content. Specifically, a network database containing historical operating data of energy storage facilities, physical model parameters, and optimization suggestion templates is constructed, and the data displayed in real time is compared with the data stored in the network database to obtain matching optimization suggestions.

[0017] As a preferred embodiment of the present invention, in S4, the Pareto optimization problem is expressed as: ; In the formula, Indicates energy loss; Indicates the temperature index of thermal risk; The optimal charging and discharging strategy is obtained by solving the Pareto optimization problem using the NSGA-II algorithm.

[0018] The beneficial effects of this invention are: This invention proposes a method that integrates multi-parameter sensing, physical mechanism embedding deep learning, and adaptive early warning decision-making. This method can assess the health status of virtual power plant energy storage in real time, comprehensively, and accurately, effectively solving the problems of traditional monitoring technologies such as single monitoring dimensions, insufficient data processing capabilities, poor dynamic adaptability, and lack of prediction and decision support.

[0019] This invention achieves closed-loop optimization management by visually displaying monitoring data and evaluation results and generating optimization suggestions. It improves the accuracy of battery health assessment, extends the lifespan of virtual power plant energy storage, reduces maintenance costs, and enhances the safety and reliability of the system. It is applicable to a variety of energy storage devices and has broad application prospects and significant economic value. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the principle of the method of the present invention; Figure 2 These are the SOH prediction curves of batteries B0005 and B0006 used as the training set during the verification process of this invention; Figure 3 These are the SOH prediction curves of batteries B0006 and B0007 used as the training set during the verification process of this invention; Figure 4 These are the SOH prediction curves of batteries B0005 and B0007 used as the training set during the verification process of this invention. Detailed Implementation

[0021] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, the method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling includes the following steps: S1. Deploy a distributed sensor network in the energy storage facility of the virtual power plant to collect the electrochemical and thermodynamic parameters of the energy storage facility in real time, preprocess the collected data, and store the preprocessed data in the database. S2. Calculate the current SOH of the energy storage facility, and construct a physical information neural network PINN based on LSTM, physical model and joint loss function. Based on the current SOH and historical data in the database, predict the SOH of the energy storage facility, compare the prediction result with the current calculation result, and determine whether to issue an alarm. S3. Display real-time data curves (voltage, current, temperature, etc.), current SOH, SOH prediction charts (battery performance changes over a future period), and alarm information (including relevant time points) in a visual manner. S4. With the goal of maximizing SOH and minimizing energy loss and thermal risk, construct a Pareto optimization problem to solve for the optimal charging and discharging strategy (including the optimal charging current, cutoff temperature, etc.). The optimal charging and discharging strategy is then sent to the battery management system (BMS) in real time to dynamically adjust the operating strategy.

[0022] In S1, electrochemical parameters include voltage, current, internal resistance, and state of charge, while thermodynamic parameters include battery surface / internal temperature and heat flux density. Sensor data is transmitted to edge computing nodes via a redundant communication protocol (CAN bus + LoRa), and the sampling frequency is dynamically adjusted according to the operating conditions (1Hz in steady state and 10Hz in transient state).

[0023] The specific process of preprocessing is as follows: S1.1. Anomaly data is identified by combining dynamic thresholding with the isolated forest algorithm. In the dynamic thresholding method, the threshold is dynamically calculated and adjusted using the 3σ criterion. When the percentage of abnormal data exceeds the set alarm threshold (5%), the alarm mechanism is triggered to notify maintenance personnel to conduct an inspection. When the proportion of abnormal data does not exceed the set alarm threshold, data correction is performed, using interpolation, historical data filling, or sliding window averaging to repair the abnormal data. S1.2 Perform spatiotemporal alignment on the repaired data, including spatial alignment and time alignment. Spatial alignment is achieved based on the Kriging interpolation algorithm, and time alignment is achieved by using timestamp calibration to eliminate the time series deviation of multi-source data. S1.3 Construct the spatiotemporally aligned data into tensor form: ; In the formula, Represents the constructed tensor; V is voltage; I is current; Indicates the surface temperature of the battery; Indicates the internal temperature of the battery; Indicates internal resistance; Indicates heat flux density; Let N represent an N×6 dimensional real space, where N represents the number of samples; The state of charge (SOC) is not included in the data because it is a derived parameter. It needs to be calculated using the original electrochemical parameters such as voltage (V) and current (I) in the tensor, combined with the ampere-hour integration method or Kalman filtering. Therefore, it is not included in the tensor dimension as the original acquired data.

[0024] S1.4, For Multi-scale features are extracted by wavelet transform, and dimensionality reduction is performed by principal component analysis (PCA). Kalman filtering and wavelet transform are then used for filtering and denoising. S1.5. Normalize the data processed in S1.4 to complete the data preprocessing.

[0025] In S2, the architecture of PINN includes: The LSTM temporal feature extraction layer employs a dual-hidden-layer structure. The input is historical data from a database, and the output is the hidden state h. t ; The physical mechanism embedding layer directly embeds the physical model, which includes both electrochemical and thermodynamic models, as hard constraints into PINN. Specifically, the residual terms of the solid-phase diffusion equation in the electrochemical model... Constraining the predicted SOC value to ensure the lithium-ion concentration distribution conforms to Fick's law; constraining the overpotential using the Butler-Volmer equation to prevent non-physical decomposition of the electrochemical reaction rate; and addressing the residual terms of the heat balance equation in the thermodynamic model. Limit the temperature prediction range to avoid energy non-conservation, and ensure that the surface heat transfer boundary conditions conform to the actual heat dissipation law by limiting the temperature prediction range. The joint loss function L is expressed as: ; In the formula, , The weighting coefficients are determined through Bayesian optimization (in this embodiment). , ); Indicates mean square error; Represents the L2 norm; Indicates the predicted value; Represents the actual value.

[0026] The electrochemical models specifically include: Solid-phase diffusion equation based on Fick's second law; The internal boundary conditions are spherically symmetric constraints: ; In the formula, c s Indicates lithium ion concentration; r represents the radial position of the particle; Surface boundary conditions, correlation between lithium-ion flux and reaction current density: ; In the formula, D s ρ is the solid-phase lithium diffusion coefficient; j is the surface reaction current density; a is the specific surface area; F represents the Faraday constant; R represents the particle radius; The surface reaction follows the Butler-Volmer equation; The thermodynamic model specifically includes: Based on the law of conservation of energy, the heat balance equation of a battery is expressed as: ; In the formula, m represents the battery mass; c p T represents the battery's specific heat capacity; T represents the battery temperature; t represents time. This represents the total heat generated inside the battery, including ohmic heat, reaction heat, and concentration polarization heat (electrochemical heat). This represents the total heat lost by the battery through convection, conduction, and radiation. The internal boundary is represented as: ; The outer boundary is represented as: ; In the formula, h represents the convective heat transfer coefficient; The ambient temperature is represented by k, which represents the thermal conductivity of the battery material.

[0027] In PINN, during the forward propagation phase, the LSTM temporal feature extraction layer outputs the hidden state h. t Then, h t Input is fed into the physical mechanism embedding layer to solve. and During the backpropagation phase, and Participating in gradient calculation forces the PINN network parameters to satisfy data fitting and physical laws while updating; The predicted result is compared with the current calculation result, and the difference or error percentage between the two is calculated. Based on the set error threshold, it is determined whether an alarm is required. If the error exceeds the corresponding error threshold, an alarm signal is issued to remind maintenance personnel to check. Otherwise, the visualization display stage is entered.

[0028] A dynamic adaptive training strategy is constructed for phased training of PINN, specifically as follows: Pre-trained LSTM: uses only the data loss term, i.e. Preliminary learning of temporal characteristics; Physical constraint fine-tuning: gradually increasing The value was increased from 0 to 0.3 to ensure that the prediction results gradually conformed to the physical laws. Online update mechanism: The model is fine-tuned every 24 hours based on the latest data, and a flexible weight solidification algorithm is used to retain key physical features and prevent catastrophic amnesia; When comparing the predicted results with the current calculation results, if the error exceeds the corresponding error threshold (5%), the online retraining of PINN is triggered simultaneously, and the parameters of the electrochemical model and thermodynamic model (solid-phase lithium diffusion coefficient D) are updated. s (e.g., convective heat transfer coefficient h, surface reaction current density j).

[0029] In S3, the visualization process also displays the remaining number of cycles before the predicted capacity decays to 80% and the root cause analysis of failure. The remaining number of cycles before the predicted capacity decays to 80% is obtained by first calculating the lithium loss caused by the SEI layer based on the SEI growth kinetics equation, and then calculating the total capacity change. ; In the formula, This represents the battery capacity at time t; Q0 represents the initial capacity. represents the lithium loss caused by the SEI layer at time t; q and n are empirical fitting parameters, representing the decay rate and the degree of nonlinearity of the decay curve, respectively; Q represents the decay term adjustment coefficient, used to control the contribution of the SEI layer to the total capacity change in the capacity decay model. The SOH sequence output by PINN is input into a particle filter, and a state-space model is constructed by combining the SEI growth kinetic equation and battery capacity change to predict the remaining number of cycles until the capacity decays to 80%. The root cause analysis of the failure was to locate the dominant factors of decay (such as lithium deposition, thermal runaway, etc.) by using the physical residual terms of PINN.

[0030] During visualization, optimization suggestions are generated based on the displayed content. Specifically, a network database is constructed that includes historical operating data of energy storage facilities, physical model parameters, and optimization suggestion templates (predefined optimization suggestion templates, such as heat dissipation adjustment suggestions for abnormal temperatures, charge and discharge depth limit suggestions for SOH decay, and maintenance plan formulation). The real-time displayed data is compared with the data stored in the network database (using a well-known similarity matching algorithm) to obtain matching optimization suggestions.

[0031] Users can view data in a visual interface and adjust their operations based on suggestions. Simultaneously, optimized parameters are fed back to PINN to continuously improve the health management of the virtual power plant's energy storage.

[0032] In S4, the Pareto optimization problem is represented as: ; In the formula, Indicates energy loss; Indicates the temperature index of thermal risk; The optimal charging and discharging strategy is obtained by solving the Pareto optimization problem using the NSGA-II algorithm.

[0033] The verification process is as follows: Two schemes are used to predict and simulate the battery state: Scheme 1: prediction using only an LSTM model; Scheme 2: prediction using LSTM coupled with a physical model (Physics-Informed Neural Network) (i.e., the method in this embodiment), in order to compare the prediction accuracy of the two schemes.

[0034] The core objective of this experiment is to evaluate the performance of two methods in predicting SOC (State of Charge), SOH (State of Health), and temperature. By comparing experimental measurement data, it aims to verify whether the method of coupling LSTM with a physical model can improve prediction accuracy, enhance physical consistency, and better adapt to battery operating conditions under different circumstances. NASA batteries B0005, B0006, and B0007 were used for validation. Two of these batteries were used as the training set, and the remaining set was used as the validation dataset. Specifically, batteries B0005 and B0006 were used as the training set, and battery B0007 as the test set; batteries B0006 and B0007 were used as the training set, and battery B0005 as the test set; and batteries B0005 and B0007 were used as the training set, and battery B0006 as the test set.

[0035] In data preprocessing, all data were normalized to eliminate outliers, and a sliding window method was used to construct time series data to ensure that the LSTM model could fully learn the dynamic changes in battery state.

[0036] During the model training phase, the LSTM model is first trained using historical data such as current, voltage, and temperature as input features. Two hidden layers and a fully connected layer are configured in the LSTM to output predicted future SOC, SOH, and temperature values. Mean squared error (MSE) is used as the loss function during training, and the Adam optimizer is used for gradient descent optimization. The LSTM model can fit historical data well under normal operating conditions, but under extreme conditions (such as low temperature or high load changes), its predictions often deviate from the true values, and it is prone to physically unreasonable SOC fluctuations or abnormal temperature changes. Because LSTM relies solely on data and lacks physical constraints, its generalization ability is weak, especially when the data distribution changes, the error increases significantly.

[0037] Option 2 introduces a physical model on top of LSTM, constructing the PINN method, which couples LSTM and the physical model. This method incorporates physical constraints into the loss function of LSTM, ensuring that the prediction results conform to the electrochemical and thermodynamic principles of the battery. In its implementation, a single-event model (SPM) is used to describe the diffusion process of lithium ions in the electrode, combined with a thermal model of the battery to constrain the rationality of temperature changes. The loss function of the PINN model consists of a data error term, a physical model constraint term, and a thermal model constraint term. The data error term is responsible for fitting the training data, the physical model constraint term ensures that the change in SOC conforms to the lithium-ion diffusion equation, and the thermal model constraint term ensures that the temperature change conforms to the principle of energy conservation. To improve training stability, a staged training strategy is adopted. First, the LSTM part is pre-trained to learn basic time-series features. Then, the weight of the physical loss term is gradually increased, allowing the model to still fit the experimental data well while conforming to physical constraints.

[0038] In the experimental testing section, independent datasets were used for validation, and the performance of the two methods under different operating conditions was compared and analyzed. Regarding SOC prediction, the LSTM model showed relatively small errors within the normal operating range, but the errors increased significantly under complex conditions such as low temperatures and high-speed charging and discharging. The PINN model, due to its integration of a single-event model, accurately describes the diffusion behavior of lithium ions, thus maintaining low prediction errors under all test conditions. Regarding SOH prediction, the LSTM model relies on historical data trends; when the battery aging mode changes (e.g., accelerated aging rate or enhanced polarization), the prediction error increases significantly. The PINN model, because its physical model includes the internal degradation mechanism of the battery, can better adapt to changes in battery health. Regarding temperature prediction, the LSTM model's prediction results are often limited by the data distribution, and in some cases, unreasonable temperature drops or rises may occur. The PINN model, by introducing thermal model constraints, ensures that temperature changes conform to the law of conservation of energy, resulting in a smoother prediction curve that conforms to actual physical laws. Figure 2 , Figure 3 , Figure 4 As shown.

[0039] Comprehensive error analysis shows that the average error of the PINN method in predicting SOC, SOH, and temperature is significantly lower than that of the LSTM method. In particular, under non-standard operating conditions (such as extreme temperatures, high-rate charge and discharge, and aging conditions), the PINN method can maintain more stable prediction accuracy.

[0040] LSTM methods, because they do not involve physical model calculations, have faster inference speeds and are suitable for computationally limited applications, such as real-time BMS (Battery Management Systems). PINN methods, on the other hand, require the calculation of physical loss terms, increasing training time by approximately 30% compared to LSTM. However, their computational load during inference remains lower, making them suitable for offline optimization or cloud-based analysis. Despite the longer training time, the higher accuracy of PINN methods in long-term prediction and battery life management can reduce maintenance costs and improve system safety and reliability. Therefore, PINN methods are more advantageous in practical applications.

[0041] In summary, by comparing LSTM and LSTM+physical model coupling methods, the effectiveness of the Physical Information Neural Network (PINN) in battery state prediction was verified. The PINN method not only outperforms SOC, SOH, and temperature predictions but also ensures that the prediction results conform to the battery's physical mechanisms, improving the model's adaptability and robustness under complex operating conditions. Therefore, the LSTM method combined with a physical model can serve as a preferred solution for future battery state prediction and health management, particularly suitable for scenarios requiring high-precision prediction and long-term health management, such as electric vehicles and grid energy storage.

[0042] Example 2: Introducing a transfer learning-driven physical parameter decoupling mechanism in PINN, the process is as follows: Step 1: Calculate in real time the gradient sensitivity of the physical model parameters embedded in PINN to the joint loss function L. , is represented as: ; In the formula, This is a set of physical model parameters, including the solid-phase lithium diffusion coefficient D. s Convection heat transfer coefficient h, surface reaction current density j; For gradient operators, Indicates L pair The gradient vector; Step 2: When there are K consecutive training cycles When the gradient threshold is exceeded, it is determined to be a physical parameter drift, triggering transfer learning. Freeze all weight parameters of the LSTM temporal feature extraction layer; Building a lightweight migration network It includes an input layer, an intermediate layer, and an output layer, wherein the intermediate layer includes a BN layer and a ReLU activation function; Will The output is reconstructed into physical residual terms: ; In the formula, , For weighting coefficients (which can be set to adaptive weighting coefficients and dynamically adjusted through residual contribution rate); Represents the physical residual term of the reconstruction; Step 3: Add a migration stability term to L, expressed as: ; In the formula, Indicate the new joint loss function; This is the migration regularization strength coefficient; Indicates KL divergence; This represents the distribution of historical physical parameters.

[0043] Gradient-sensitive coefficient of variation detection of parameter drift is more in line with the actual degradation state than fixed-period retraining and can better reduce prediction error; the transfer network output is reconstructed into physical residual terms, which solves the problem of failure of hard constraints of physical equations in traditional transfer learning, and at the same time, the physical consistency of the transfer process is guaranteed by KL divergence constraints.

[0044] Example 3: A multi-physics coupled virtual power plant energy storage health status monitoring and prediction device, comprising: One or more processors; Memory, used to store one or more computer programs; When one or more programs are executed by one or more processors, the one or more processors perform the method in Embodiment 1 or Embodiment 2.

[0045] Example 4: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, implement the method in Example 1 or Example 2.

Claims

1. A method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling, characterized in that... Includes the following steps: S1. Deploy a distributed sensor network in the energy storage facility of the virtual power plant to collect the electrochemical and thermodynamic parameters of the energy storage facility in real time, preprocess the collected data, and store the preprocessed data in the database. S2. Calculate the current SOH of the energy storage facility, and construct a physical information neural network PINN based on LSTM, physical model and joint loss function. Based on the current SOH and historical data in the database, predict the SOH of the energy storage facility, compare the prediction result with the current calculation result, and determine whether to issue an alarm. S3. Display real-time collected data in a visual manner, including graphs, current SOH, SOH prediction charts, and alarm information; S4. With the goal of maximizing SOH and minimizing energy loss and thermal risk, construct a Pareto optimization problem, solve for the optimal charging and discharging strategy, and send the optimal charging and discharging strategy to the battery management system (BMS) in real time to dynamically adjust the operating strategy.

2. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 1, characterized in that, In S1, the electrochemical parameters include voltage, current, internal resistance, and state of charge, and the thermodynamic parameters include battery surface / internal temperature and heat flux density. The specific process of preprocessing is as follows: S1.

1. Anomaly data is identified by combining dynamic thresholding with the isolated forest algorithm. In the dynamic thresholding method, the threshold is dynamically calculated and adjusted using the 3σ criterion. When the proportion of abnormal data exceeds the set alarm threshold, the alarm mechanism is triggered to notify maintenance personnel to conduct an inspection. When the proportion of abnormal data does not exceed the set alarm threshold, data correction is performed, using interpolation, historical data filling, or sliding window averaging to repair the abnormal data. S1.2 Perform spatiotemporal alignment on the repaired data, including spatial alignment and time alignment. Spatial alignment is achieved based on the Kriging interpolation algorithm, and time alignment is achieved by using timestamp calibration to eliminate the time series deviation of multi-source data. S1.3 Construct the spatiotemporally aligned data into tensor form: ； In the formula, Represents the constructed tensor; V is voltage; I is current; Indicates the surface temperature of the battery; Indicates the internal temperature of the battery; Indicates internal resistance; Indicates heat flux density; Let N represent an N×6 dimensional real space, where N represents the number of samples; S1.4, For Multi-scale features are extracted by wavelet transform, and dimensionality reduction is performed by principal component analysis (PCA). Kalman filtering and wavelet transform are then used for filtering and denoising. S1.

5. Normalize the data processed in S1.4 to complete the data preprocessing.

3. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 1, characterized in that, In S2, the architecture of PINN includes: The LSTM temporal feature extraction layer employs a dual-hidden-layer structure. The input is historical data from a database, and the output is the hidden state h. t ; The physical mechanism embedding layer directly embeds the physical model, which includes both electrochemical and thermodynamic models, as hard constraints into PINN. Specifically, the residual terms of the solid-phase diffusion equation in the electrochemical model... Constraints on SOC predictions, constraints on overpotentials using the Butler-Volmer equation, and residual terms in the heat balance equation of the thermodynamic model. Limiting the temperature prediction range, surface heat transfer boundary conditions ensure that the temperature gradient conforms to the actual heat dissipation law; The joint loss function L is expressed as: ； In the formula, , The weighting coefficients are determined through Bayesian optimization. Indicates mean square error; Represents the L2 norm; Indicates the predicted value; Represents the actual value.

4. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 3, characterized in that, The electrochemical models specifically include: Solid-phase diffusion equation based on Fick's second law; The internal boundary conditions are spherically symmetric constraints: ； In the formula, c s Indicates lithium ion concentration; r represents the radial position of the particle; Surface boundary conditions, correlation between lithium-ion flux and reaction current density: ； In the formula, D s ρ is the solid-phase lithium diffusion coefficient; j is the surface reaction current density; a is the specific surface area; F represents the Faraday constant; R represents the particle radius; The surface reaction follows the Butler-Volmer equation; The thermodynamic model specifically includes: Based on the law of conservation of energy, the heat balance equation of a battery is expressed as: ； In the formula, m represents the battery mass; c p T represents the battery's specific heat capacity; T represents the battery temperature; t represents time. This represents the total heat generated inside the battery. This represents the total heat lost by the battery through convection, conduction, and radiation. The internal boundary is represented as: ； The outer boundary is represented as: ； In the formula, h represents the convective heat transfer coefficient; The ambient temperature is represented by k; k represents the thermal conductivity of the battery material.

5. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 3, characterized in that, In PINN, during the forward propagation phase, the LSTM temporal feature extraction layer outputs the hidden state h. t Then, h t Input is fed into the physical mechanism embedding layer to solve. and During the backpropagation phase, and Participating in gradient calculation forces the PINN network parameters to satisfy data fitting and physical laws while updating; The predicted result is compared with the current calculation result, and the difference or error percentage between the two is calculated. Based on the set error threshold, it is determined whether an alarm is required. If the error exceeds the corresponding error threshold, an alarm signal is issued to remind maintenance personnel to check. Otherwise, the visualization display stage is entered.

6. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 5, characterized in that, A dynamic adaptive training strategy is constructed for phased training of PINN, specifically as follows: Pre-trained LSTM: uses only the data loss term, i.e. Preliminary learning of temporal characteristics; Physical constraint fine-tuning: gradually increasing This ensures that the prediction results gradually conform to the laws of physics; Online update mechanism: The model is fine-tuned every 24 hours based on the latest data, and a flexible weight solidification algorithm is used to retain key physical features; When comparing the predicted results with the current calculation results, if the error exceeds the corresponding error threshold, the online retraining of PINN is triggered simultaneously, and the parameters of the electrochemical model and the thermodynamic model are updated.

7. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 6, characterized in that, In PINN, a transfer learning-driven physical parameter decoupling mechanism is introduced, as follows: Step 1: Real-time calculation of the gradient sensitivity of the physical model parameters embedded in PINN to the joint loss function L. , represented as: ； In the formula, This is a set of physical model parameters, including the solid-phase lithium diffusion coefficient D. s Convection heat transfer coefficient h, surface reaction current density j; For gradient operators, Indicates L pair The gradient vector; Step 2: When there are K consecutive training cycles When the gradient threshold is exceeded, it is determined to be a physical parameter drift, triggering transfer learning. Freeze all weight parameters of the LSTM temporal feature extraction layer; Building a lightweight migration network It includes an input layer, an intermediate layer, and an output layer, wherein the intermediate layer includes a BN layer and a ReLU activation function; Will The output is reconstructed into physical residual terms: ； In the formula, , These are the weighting coefficients; Represents the physical residual term of the reconstruction; Step 3: Add a migration stability term to L, expressed as: ； In the formula, Indicate the new joint loss function; This is the migration regularization strength coefficient; Indicates KL divergence; This represents the distribution of historical physical parameters.

8. The method for monitoring and predicting the health status of virtual power plant energy storage using multiphysics coupling as described in claim 3, characterized in that, In S3, the visualization process also displays the remaining number of cycles before the predicted capacity decays to 80% and the root cause analysis of failure. The remaining number of cycles before the predicted capacity decay to 80% is obtained by first determining the lithium loss caused by the SEI layer based on the SEI growth kinetics equation, and then calculating the total capacity change. ； In the formula, This represents the battery capacity at time t; Q0 represents the initial capacity. represents the lithium loss caused by the SEI layer at time t; q and n are empirical fitting parameters, representing the decay rate and the degree of nonlinearity of the decay curve, respectively; Q represents the decay term adjustment coefficient, used to control the contribution of the SEI layer to the total capacity change in the capacity decay model. The SOH sequence output by PINN is input into a particle filter, and a state-space model is constructed by combining the SEI growth kinetic equation and battery capacity change to predict the remaining number of cycles until the capacity decays to 80%. The root cause analysis of the failure was to locate the dominant factor of attenuation through the physical residual terms of PINN.

9. The method for monitoring and predicting the health status of virtual power plant energy storage using multi-physics coupling as described in claim 1, characterized in that, In S3, during the visualization display, optimization suggestions are generated based on the displayed content. Specifically, a network database containing historical operating data of energy storage facilities, physical model parameters, and optimization suggestion templates is constructed. The real-time displayed data is compared with the data stored in the network database to obtain matching optimization suggestions.

10. The method for monitoring and predicting the health status of virtual power plant energy storage using multi-physics coupling as described in claim 1, characterized in that, In S4, the Pareto optimization problem is expressed as: ； In the formula, Indicates energy loss; Indicates the temperature index of thermal risk; The optimal charging and discharging strategy is obtained by solving the Pareto optimization problem using the NSGA-II algorithm.

Images