Method and device for evaluating health state of UPS (Uninterrupted Power Supply) storage battery based on physical information neural network, processor and storage medium thereof

Abstract

The invention relates to a method for evaluating the health state of a UPS (Uninterrupted Power Supply) storage battery based on a physical information neural network. The method comprises the following steps: establishing a physical guide mechanism according to battery discharge test data; establishing a multi-layer perceptron model; performing offline training according to the physical constraint and the data error; and realizing SOH online evaluation according to the real-time data of the battery. According to the method and the device for evaluating the state of health of the UPS storage battery based on the physical information neural network, the processor and the computer readable storage medium, a physical constraint fused VRLA storage battery SOH evaluation model and a training method are provided, a traditional data driven neural network is combined with a physical rule in a battery aging process, and the evaluation accuracy of the state of health of the UPS storage battery is improved. In the network training process, the physical constraint is introduced as a part of the loss function, so that the physical consistency of the model is effectively improved. According to the method, the black box problem of a traditional deep learning method is solved, and the interpretability and stability of the model are remarkably enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power guarantee device, especially to the field of UPS battery management technology, and in particular to a method and device for evaluating the state of health of a UPS battery based on a physical information neural network, a processor and a computer readable storage medium thereof. BACKGROUND

[0002] With the rapid development of digital economy, data centers as the core facilities to carry Internet basic services have been widely deployed. To ensure efficient storage and processing of data, the stability and reliability of its power supply system are crucial. Uninterruptible power supply (UPS) system as the key power guarantee device of data center, provides seamless switching and continuous power supply capability when the power is interrupted or the power grid fluctuates. The battery is the core energy storage component in the UPS system, which is directly related to the emergency power supply capability of the data center in the case of sudden power failure. Among them, the valve regulated lead acid (VRLA) battery is widely used in the UPS system of the data center due to its low cost, high safety, easy maintenance and other advantages. However, this type of battery is mostly in the floating state during daily operation, and has not been activated for a long time. It is prone to performance degradation, capacity decline or failure. It is particularly important to accurately evaluate the SOH of the battery online to ensure the safe and stable operation of the power system of the data center.

[0003] The existing evaluation methods of battery SOH can be divided into two categories: 1. Physical law method, including open circuit voltage method, internal resistance method, temperature attenuation method, etc. This kind of method uses the objective physical law of the battery to establish a health evaluation model, and its modeling process depends on the experience of experts. Due to the complexity of the physical mechanism of the battery, it is difficult to accurately describe its behavior through a clear state equation, and this method may ignore the potential nonlinear relationship and coupling mechanism between variables in the system, and its practicability is limited. Among them, "State Estimation of Energy Storage Battery Based on Model Parameter Identification" calculates the full discharge open circuit voltage through parameter identification to realize the battery SOH evaluation, but it needs to stand the battery for a long time and can only be estimated offline after each complete charge and discharge cycle. "Research on Detection and Grade Utilization of Power Battery Health State Based on Adaptive Unscented Kalman Filter" uses adaptive unscented Kalman filter algorithm to estimate the ohmic internal resistance and SOH of the battery in real time, but it needs to measure and analyze the ohmic internal resistance of the aged battery, and the test method is complex. "Research on High-precision Capacity Detection System of Mine Lithium Battery Based on Internal Resistance Method" uses the alternating current injection method to realize the online monitoring of the internal resistance of the battery, and estimates the capacity based on the obtained internal resistance parameters, but the internal resistance measurement device is expensive. 2. Data-driven method, including fuzzy logic, artificial neural network, support vector machine, etc. This kind of method does not need explicit modeling, has strong adaptability and simple implementation process, but has high requirements for the quantity and quality of data, poor model interpretability, and may produce "black box" prediction results that violate the physical law. Among them, "Fast Estimation Method of Lithium Ion Battery Health State Based on Improved Least Squares Support Vector Machine" proposes a fast and accurate SOH estimation method based on improved least squares support vector machine, but needs sufficient data sample size to train the model to enhance the precision. "Battery Health Degree Evaluation for Power Grid Based on Artificial Intelligence" proposes to realize the battery SOH evaluation by using the improved LMBP neural network algorithm based on data driving, which improves the convergence and convergence speed, but it is difficult to analyze the specific influence of each input parameter on the output, and lacks clear physical significance correlation. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a method, device, processor and computer readable storage medium for evaluating the health state of UPS battery based on physical information neural network, which has high prediction accuracy, good stability and wide application range.

[0005] In order to achieve the above purpose, the method, device, processor and computer readable storage medium for evaluating the health state of UPS battery based on physical information neural network are as follows: The method for evaluating the health state of UPS battery based on physical information neural network mainly comprises the following steps: (1) Establish a physical guidance mechanism according to the battery discharge test data; (2) Establish a multi-layer perception model; (3) Offline training according to physical constraints and data errors; (4) Realize SOH online evaluation according to real-time battery data.

[0006] Preferably, the step (1) specifically comprises the following steps: By collecting the data of the battery in the discharge test state, the internal resistance and actual capacity data points are extracted, and the cubic spline interpolation method is used to establish the mapping relationship between the internal resistance and the capacity.

[0007] Preferably, the step (3) specifically comprises the following steps: (3.1) Data preprocessing of battery characteristic parameters, using a standardizer to process all input features; (3.2) Randomly shuffle the battery characteristic data set, and divide the battery characteristic data set to obtain a training set, a validation set and a test set; (3.3) PINN model initialization; (3.4) Train the PINN model in an iterative optimization manner, continuously perform forward propagation and back propagation of the PINN model, and update the model parameters; (3.5) Save the PINN model parameters.

[0008] Preferably, the step (3.3) specifically comprises the following steps: (3.3.1) Define a combined loss function; (3.3.2) Dynamically adjust the weight coefficient of the physical constraint term during training; (3.3.3) Adopt a linear decay strategy to adjust the weight coefficient according to the dynamic change strategy, so that gradually decreases with the number of training rounds.

[0009] Preferably, the step (3.4) specifically comprises the following steps: (3.4.1) In each training round, perform forward propagation operation on the training set samples to calculate the model prediction output; (3.4.2) According to the defined composite loss function, perform back propagation to calculate the gradient; (3.4.3) Update the model parameters through the optimizer to gradually minimize the total loss function.

[0010] Preferably, the step (4) specifically comprises the following steps: ​(4.1) Real-time acquisition of the operating state information of each battery in each UPS power system, extraction of key characteristic parameters as model input; (4.2) Online estimation of the remaining capacity of each battery based on the PINN model; (4.3) Calculation of the battery SOH at each sampling point according to the battery remaining capacity evaluated by the PINN model; (4.4) Aggregation of several batteries belonging to the same battery pack in the UPS system according to the group, and calculation of the average SOH of each battery pack; (4.5) Summary of the predicted capacity values of all batteries, and calculation of the average SOH index of the system as a whole; (4.6) Summary of all statistical results and output.

[0011] The system for evaluating the health state of UPS batteries based on a physical information neural network has the following main features: A physical guidance module is used to explicitly embed physical constraints into the loss function of the neural network; A multilayer perception module is connected to the physical guidance module and uses several fully connected feedforward neural network layers to learn the complex nonlinear relationship between input features and battery remaining capacity; An offline training module is connected to the physical guidance module and is used to train the neural network model through a battery characteristic parameter data set; An online evaluation module is connected to the offline training module and is used to quickly evaluate the battery SOH state based on real-time collected battery characteristic parameters during the operation of the UPS system.

[0012] The device for evaluating the health state of UPS batteries based on a physical information neural network has the following main features: A processor configured to execute computer executable instructions; A memory storing one or more computer executable instructions, which, when executed by the processor, implement the steps of the method for evaluating the health state of UPS batteries based on a physical information neural network.

[0013] The processor for evaluating the health state of UPS batteries based on a physical information neural network has the following main features: the processor is configured to execute computer executable instructions, and the computer executable instructions, when executed by the processor, implement the steps of the method for evaluating the health state of UPS batteries based on a physical information neural network.

[0014] The computer readable storage medium is characterized in that a computer program is stored thereon, and the computer program can be executed by the processor to realize each step of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0015] The method, device, processor and computer readable storage medium for evaluating the state of health of the UPS battery based on the physical information neural network have the advantages that a SOH evaluation model and a training method for the VRLA battery are proposed, the traditional data-driven neural network is combined with the physical law in the battery aging process, the physical constraint is introduced as part of the loss function in the network training process, and the physical consistency of the model is effectively improved. The method overcomes the “black box” problem of the traditional deep learning method, and significantly enhances the explainability and stability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The figure is a structural block diagram of the system for evaluating the state of health of the UPS battery based on the physical information neural network.

[0017] Figure 2 The figure is a scatter plot of the estimation results of the method for evaluating the state of health of the UPS battery based on the physical information neural network on the test set.

[0018] Figure 3 The figure is a general flowchart of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0019] Figure 4 The figure is a neural network structure diagram of the PINN of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0020] Figure 5 The figure is a loss curve diagram in the offline training process of the embodiment of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0021] Figure 6 The figure is an average SOH distribution diagram of the UPS system as a whole of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0022] Figure 7 The figure is a group SOH distribution diagram of each group of batteries of the method for evaluating the state of health of the UPS battery based on the physical information neural network.

[0023] Figure 8 The figure is an evaluation SOH thermal diagram of each section of battery of the method for evaluating the state of health of the UPS battery based on the physical information neural network. DETAILED DESCRIPTION

[0024] In order to more clearly describe the technical content of the present application, further description will be made in combination with specific embodiments.

[0025] The method for evaluating the state of health of the UPS battery based on the physical information neural network of the present application comprises the following steps: (1) establishing a physical guidance mechanism according to battery discharge test data; (2) establishing a multi-layer perception model; (3) performing offline training according to physical constraints and data errors; (4) realizing SOH online evaluation according to real-time battery data.

[0026] As a preferred embodiment of the present application, the step (1) specifically comprises the following steps: By collecting the data of the battery in the discharge test state, the internal resistance and actual capacity data points are extracted, and the cubic spline interpolation method is used to establish the mapping relationship between the internal resistance and the capacity.

[0027] As a preferred embodiment of the present application, the step (3) specifically comprises the following steps: (3.1) performing data preprocessing of battery characteristic parameters, and using a standardizer to process all input features; (3.2) performing random shuffling operation on the battery characteristic data set, and dividing the battery characteristic data set to obtain a training set, a validation set and a test set; (3.3) performing PINN model initialization; (3.4) training the PINN model in an iterative optimization manner, continuously performing forward propagation and backward propagation of the PINN model, and updating the model parameters; (3.5) saving the PINN model parameters.

[0028] As a preferred embodiment of the present application, the step (3.3) specifically comprises the following steps: (3.3.1) defining a combined loss function; (3.3.2) dynamically adjusting the weight coefficient of the physical constraint term in the training process; (3.3.3) using a linear decay strategy to adjust the weight coefficient according to the dynamic change strategy, so that gradually decreases with the training round number.

[0029] As a preferred embodiment of the present application, the step (3.4) specifically comprises the following steps: ​​(3.4.1) In each round of training, forward propagation operation is performed on the training set samples to calculate the model prediction output; (3.4.2) According to the defined composite loss function, back propagation is performed to calculate the gradient; (3.4.3) The model parameters are updated by the optimizer to gradually minimize the total loss function.

[0030] As a preferred embodiment of the present application, the step (4) specifically comprises the following steps: (4.1) Real-time acquisition of the operating state information of each battery in each UPS power supply system, and extraction of key characteristic parameters as model input; (4.2) Online estimation of the remaining capacity of each battery based on the PINN model; (4.3) Calculation of the battery SOH at each sampling point according to the battery remaining capacity evaluated by the PINN model; (4.4) Aggregation of several batteries belonging to the same battery pack in the UPS system according to the group, and calculation of the average SOH of each battery pack; (4.5) Summary of the predicted capacity values of all batteries, and calculation of the average SOH index of the system as a whole; (4.6) Summary of all statistical results and output.

[0031] The system for evaluating the health state of UPS batteries based on the physical information neural network of the present application, wherein the system comprises: A physical guidance module for embedding physical constraints explicitly into the loss function of the neural network; A multilayer perception module connected to the physical guidance module, which adopts several fully connected feedforward neural network layers to learn the complex nonlinear relationship between input features and battery remaining capacity; An offline training module connected to the physical guidance module, which is used to train the neural network model through a battery characteristic parameter data set; An online evaluation module connected to the offline training module, which is used to quickly evaluate the battery SOH state based on the real-time collected battery characteristic parameters during the operation of the UPS system.

[0032] The device for evaluating the health state of UPS batteries based on the physical information neural network of the present application, wherein the device comprises: A processor configured to execute computer executable instructions; A memory storing one or more computer executable instructions, which, when executed by the processor, implement the steps of the method for evaluating the health state of UPS batteries based on the physical information neural network.

[0033] The processor for implementing the UPS battery health state evaluation based on the physical information neural network of the present application, wherein the processor is configured to execute computer executable instructions, and the computer executable instructions are executed by the processor to implement the steps of the method for implementing the UPS battery health state evaluation based on the physical information neural network.

[0034] The computer readable storage medium of the present application has a computer program stored thereon, and the computer program can be executed by a processor to implement the steps of the method for implementing the UPS battery health state evaluation based on the physical information neural network.

[0035] Currently, some researches have attempted to use machine learning methods to model and predict the SOH of the battery, but most of them focus on black box models trained based on a large amount of experimental data, ignoring the physical constraints and evolution mechanism in the battery aging process, resulting in insufficient model in generalization ability, stability and explainability. In addition, although some researches introduce equivalent circuit models or electrochemical models for SOH estimation, the model parameters are complex, dependent on specific scenarios, and difficult to adapt to large-scale, multi-condition data center environments. In the face of the real challenges of large-capacity VRLA batteries in the data center UPS system being in a floating state for a long time, lacking complete charge and discharge data, etc., the existing methods are difficult to effectively realize online and accurate health state evaluation. Therefore, there is an urgent need for a new method that can balance physical consistency, nonlinear modeling capability and data utilization efficiency to improve the accuracy and robustness of SOH estimation. To solve the above problems, the present application combines physical laws and data-driven methods and proposes a VRLA battery SOH evaluation method for data center UPS systems based on a physical information neural network.

[0036] The present application proposes and designs a VRLA battery SOH evaluation method for data center UPS systems based on a physical information neural network (PINN), and the system structure diagram is as shown in Figure 1 The physical guiding module, the multi-layer perception module, the offline training module and the online evaluation module.

[0037] The physical guiding module explicitly embeds the physical constraints into the loss function of the neural network, providing clear physical law guidance for the PINN offline training module, which not only helps to reduce the dependence on the number of experimental data, but also helps to speed up the model convergence speed and improve the accuracy and robustness of the prediction results.

[0038] The Multi-Layer Perceptron (MLP) module adopts several fully connected feedforward neural network layers to learn the complex nonlinear relationship between the input features (internal resistance, voltage, temperature, etc.) and the remaining capacity of the battery. This module has strong fitting and generalization capabilities while introducing physical information constraints, providing data errors for offline training of the PINN model.

[0039] The offline training module is responsible for training the neural network model based on the battery characteristic parameter dataset under the joint action of historical data and physical guidance. The model minimizes the data error term and the physical constraint term in the loss function during the training process to achieve accurate modeling of the remaining capacity of the VLRA battery. This module completes model optimization under the condition of sufficient computing resources and generates deployable model parameters.

[0040] The online evaluation module is used to quickly evaluate the SOH state of the battery based on real-time collected battery characteristic parameters during the actual operation of the UPS system. This module calls the trained PINN model to realize online prediction of the remaining capacity of the battery and calculate the SOH value of the battery based on the high-efficiency forward propagation characteristics of the neural network.

[0041] To alleviate problems such as gradient vanishing or gradient explosion, each hidden layer is equipped with a Batch Normalization operation, which is mathematically expressed as follows: ; where and are the mean and variance of the small batch samples; is a small constant to prevent division by zero errors; and represent the scaling factor and offset, respectively, which are learned parameters in the network and are automatically assigned values during initialization ( , ) and updated during the training process through backpropagation.

[0042] To prevent overfitting, the network uses a decreasing Dropout strategy, i.e., setting a higher Dropout probability in the front layers and decreasing it layer by layer in the back layers, which is expressed as follows: ; where represents the Hadamard product, is the Dropout mask, is the retention probability of the th layer, and increases layer by layer in the overall design.

[0043] In addition, two residual connections are introduced in the network to promote feature reuse and information fusion, which are expressed as follows: ; (3) Offline training combined with physical constraints and data errors (301) Complete the data preprocessing of battery characteristic parameters. Specifically, 1125 battery characteristic parameter data are selected as the data basis for model training and evaluation. Due to the existence of order of magnitude difference between each characteristic parameter, if directly input into the neural network model, it is easy to lead to uneven gradient distribution in the training process, thereby affecting the optimization efficiency and convergence effect of the model. To solve the above problem, the inventor introduces a standardization processing mechanism in the data preprocessing stage, and uses a standardizer (StandardScaler) to process all input features, so that they are converted to a distribution with a mean of 0 and a standard deviation of 1. The standardization conversion process is as follows: ; In the formula, is the original feature value, is the mean of the feature, is the standard deviation of the feature, is the standardized feature value. This processing aims to improve the numerical stability and convergence speed in the model training stage.

[0044] In addition, since the physical constraint term based on the physical characteristics of the battery is introduced in the model, in order to ensure that the physical constraint function is still based on the true physical parameters during calculation, the inventor performs inverse standardization processing on the features used for physical function calculation (i.e. internal resistance) when constructing the physical guided loss function. The inverse transformation form is as follows: ; This process can restore the standardized feature value to the original physical dimension, so as to match the pre-constructed interpolation function, realize that the model prediction result meets the data fitting at the same time, and also conforms to the basic physical law between battery capacity and internal resistance, and enhances the physical consistency and engineering adaptability of the model.

[0045] (302) Obtain the training set, validation set and test set by data set division. After completing the data preprocessing, in order to comprehensively evaluate the prediction performance of the model and its generalization ability for unknown samples, the inventor reasonably divides the battery characteristic data set. Specifically, all samples are divided into training set, validation set and test set in the ratio of 8:1:1, which are respectively used for model training, tuning and final performance evaluation.

[0046] To ensure the fairness and scientificity of the division process, the original data set is randomly shuffled before division, so as to avoid the interference of data order on the model training and evaluation results. The process can be expressed as follows: ; In the formula, represents the original data set, is the shuffled data set, is the set random seed. By fixing the random seed value, the model is ensured to be repeatable in different experiments.

[0047] Then, the subsets are divided in the following way: ; In the formula, the training set accounts for 80% of the total samples, which is used for optimization of model weight parameters; the validation set accounts for 10%, which is used to supervise the model training process and adjust the hyperparameters; and the test set accounts for 10%, which is used to verify the generalization ability of the model to new samples, ensuring the objectivity and scientificity of the final evaluation results.

[0048] (303) PINN model initialization. The network weight is loaded to the GPU or CPU computing resource (automatically judged according to the system environment), and the Adam optimizer is set, with the learning rate parameter set to 0.001. In order to embed the physical constraint into the training process, the present application defines a combined loss function, which is expressed as: ; In the formula, represents the mean square error loss (MSE) between the predicted value and the true value; represents the deviation between the model output and the physical prior (through the internal resistance-capacity relationship function); represents the total error loss; is the weight coefficient of the physical constraint loss.

[0049] In order to improve the training effect of the model and enhance its physical consistency, the present application dynamically adjusts the weight coefficient of the physical constraint term in the training process. Specifically, in the early stage of training, in order to make the model preferentially learn the physical mechanism of the battery system, suppress the noise interference in the experimental data, and enhance the compliance to the physical constraint, the is set to a higher initial value (0.8). In this stage, the model pays more attention to the consistency between the prediction result and the physical prior, which helps to build a reasonable physical mapping relationship.

[0050] With the gradual progress of training, the model gradually masters the basic physical laws. To further improve the fitting ability of complex nonlinear relationships in the data, the application adopts a linear decay strategy, so that With the gradual reduction of the number of training rounds, the minimum decay is reduced to 0.2, so as to guide the model to focus more on learning data-driven characteristics, mine complex mapping relationships that are difficult to represent by traditional physical models, and ultimately improve the overall prediction accuracy.

[0051] The dynamic change strategy of the weight coefficient can be expressed as: ; In the formula, is the initial weight, is the minimum weight, is the current training round, is the decay factor. This mechanism ensures that the training focus of the model is dynamically balanced at different stages, while maintaining physical consistency and effectively improving the model's fitting ability for actual engineering data.

[0052] (304) Iteratively perform forward propagation and back propagation of the PINN model to update the model parameters. After completing the model initialization, the PINN model is trained using an iterative optimization method. Specifically, in each training round, first perform the forward propagation operation on the training set samples to calculate the model prediction output; then, according to the defined composite loss function (including the data fitting term and the physical constraint term), perform back propagation to calculate the gradient, and update the model parameters through the optimizer, thereby gradually minimizing the total loss function.

[0053] The application uses the Adam optimization algorithm to iteratively update the model parameters, and the forward and backward propagation processes are continuously executed for several rounds (5000 rounds in this embodiment) to improve the model's fitting ability and generalization ability for the battery capacity prediction task. The training steps of each iteration are as follows: (3041) Forward propagation: ; In the formula, is the input feature of the i-th sample, is the model prediction output, is the current model parameter.

[0054] (3042) Loss calculation: ; ; ; In the formula, is the number of training batches. ​

[0055] (3043) Backpropagation and model parameter update: ; wherein, is the learning rate, is the gradient of the model parameters.

[0056] After each round of training, forward propagation and loss evaluation are also performed on the validation set to supervise the model training process and prevent overfitting. The change trend of data fitting loss, physical constraint loss and total loss is recorded during the training process as the basis for evaluating the convergence of the model. The loss curve of the embodiment of the present application is shown in Figure 5 .

[0057] (305) Save the PINN model parameters. In order to facilitate subsequent model deployment and calling, after the training of the PINN model is completed, the model parameters obtained by the training are persistently stored in the embodiment of the present application. Specifically, by calling the save interface of the model framework, the weight parameters of each layer of the neural network after the training are exported in a structured format and saved to the local file system, facilitating the loading and reuse of the model in other systems.

[0058] (4) Realize SOH online evaluation according to real-time battery data; (401) Obtain real-time monitoring data of battery characteristic parameters. Based on the data center environment control system and the UPS system monitoring platform, the running state information of each battery in each UPS power system is obtained in real time in the embodiment of the present application, and the key characteristic parameters are extracted as model inputs. The characteristic parameters include voltage (unit: volt, V), battery internal resistance (unit: milliohm, mΩ) and terminal temperature (unit: Celsius, ℃), which constitute a three-dimensional input feature vector: ; The collected data needs to be standardized before inputting into the neural network model, in order to unify the dimension and improve the prediction stability of the model. The processing process is as follows: ; wherein, and are the mean and standard deviation of each characteristic parameter, respectively, which are derived from the statistical data of the training data set.

[0059] The standardized data will be converted into tensor format and input into the trained PINN model to complete the capacity prediction.

[0060] (402) Online estimation of the remaining capacity of each battery based on the PINN model.

[0061] After obtaining the real-time characteristic parameters of each battery and completing data preprocessing, the PINN model trained is used to estimate the remaining capacity of the battery online. Specifically, the input characteristic vector after standardization: ; In the formula, represents the battery number, represents the sampling time sequence number.

[0062] is input into the PINN model, the forward propagation process is performed, and the estimated remaining capacity value is output: ; In the formula, is the remaining capacity estimated by the neural network (unit: ampere-hour, AH). represents the mapping function of the PINN model, is the model parameter obtained by training.

[0063] To ensure the physical credibility of the prediction value, the physical constraint function constructed by the embodiment of the present application is used to: ; check the neural network output, calculate the difference between the physical prediction value and the neural network output , to assist in judging the consistency and reliability of the model prediction, and finally realize the real-time online estimation of the remaining capacity of each battery.

[0064] (403) Single battery SOH calculation. According to the battery remaining capacity evaluated by the PINN model, the SOH of each sampling point is calculated according to the following formula: ; In the formula, is the rated capacity corresponding to each storage battery, unit AH. In the embodiment of the present application, the storage battery selected by the UPS system is Tangshu NP220-12H, The value is 220 AH.

[0065] Then, the average value of all sampling points of each battery is taken to obtain the average SOH index of the battery: ; In the formula, is the sampling number of the battery .

[0066] (404) Battery pack SOH statistics. The embodiment of the present application further aggregates several batteries belonging to the same UPS system battery pack according to groups, and calculates the average SOH of each battery pack. Let the first group contain The average SOH of each group is calculated as follows: ; The result is used to reflect the health level of a single group of UPS batteries in a day or a period.

[0067] (405) System overall SOH statistics. To reflect the overall operation health status of the entire UPS system, the predicted capacity values of all batteries are summarized, and the average SOH index of the entire system is calculated: ; In the formula, is the predicted capacity value of all batteries, is the total number of all predicted capacity samples.

[0068] (406) Result output. All statistical results are finally summarized into a structured Excel file, including: the average SOH of the entire UPS system, the group average SOH of each group of batteries, and the evaluation SOH of each battery, as shown in Figure 6 、 7 , 8 respectively. The analysis result can provide quantitative basis and technical support for the operation and maintenance and health warning of the UPS battery system.

[0069] In the specific embodiments of the present application, the battery state of health evaluation method based on physical information neural network includes: (1) Obtain the real-time operation characteristic parameters of the battery, including but not limited to voltage, internal resistance and temperature, and form a feature vector; (2) Standardize the feature vector to convert it into a distribution form with a mean of 0 and a standard deviation of 1, and obtain a standardized input; (3) Input the standardized input into the pre-trained PINN model, and output the predicted capacity value of the battery through forward propagation; (4) Based on the physical constraint function constructed in the training stage, calculate the capacity reference value under the physical constraint according to the battery internal resistance; (5) Compare the predicted capacity value of the battery in step 3) with the capacity reference value under the physical constraint in step 4), calculate the difference and judge the prediction credibility; (6) Calculate the battery state of health (SOH) based on the predicted capacity and the rated capacity.

[0070] The above steps (1) to (6) correspond to the operation steps of online evaluation in Figure 3 .

[0071] The PINN model adopts a deep neural network comprising an 8-layer full connection structure and a residual connection, and is trained through a combined loss function, wherein the combined loss function comprises a data fitting loss and a physical constraint loss, and the expression of the combined loss function is: ; In the formula, represents a mean square error loss (MSE) between a predicted value and an actual value; represents a deviation between a model output and a physical prior (through an internal resistance-capacity relationship function); represents a total error loss; is a weight coefficient of the physical constraint loss.

[0072] The physical constraint loss term in the formula (1) is generated based on an interpolation function constructed based on a physical correspondence between an internal resistance and a capacity, and is used to guide the neural network to output a capacity predicted value conforming to an actual battery physical characteristic.

[0073] In the model training process, the physical constraint weight linearly decays with training rounds, and the adjustment strategy is as follows: ; In the formula, is an initial weight, is a minimum weight, is a current training round number, is a decay factor.

[0074] The predicted result can be used for statistical calculation of average SOH of all batteries in the UPS system, SOH of a battery pack, and overall SOH of the system, and is output as a visual chart for operation state trend analysis and health assessment.

[0075] SOH represents a deterioration degree of a storage battery, reflects an expected service life, and is a relative quantity. With use, performance degradation and aging of the storage battery, the SOH gradually decreases. The SOH is usually defined as a ratio of a maximum remaining capacity that can be charged or discharged by the storage battery under specific conditions to a rated capacity of the storage battery, and is usually expressed in percentage, as shown in the following formula: ; In the formula, is a maximum remaining capacity that can be charged or discharged by the storage battery; is a rated capacity of the battery.

[0076] The application constructs an SOH online evaluation method based on a PINN, and in the running process, the trained PINN model parameters are called, so that the SOH can be quickly evaluated according to the real-time collected characteristic parameters such as battery voltage, internal resistance and temperature. The experimental results are as follows: Figure 2 ​The coefficient of determination (R2) of the proposed model on the test set is shown to be 0.9989, and the root mean square error (RMSE) is 1.5962, indicating that the prediction accuracy is high and the performance is stable and reliable.

[0077] The core requirement of the technical solution of the application is "long-term floating, accurate SOH evaluation under lack of complete charge-discharge data", and the data environment is "stable current, low charge-discharge frequency, and sparse sample data". The core of the technical solution of the application is a PINN model that combines physical constraints and data-driven. The physical guiding module of the technical solution of the application is constructed based on the physical characteristics of the UPS VRLA battery, i.e., a cubic spline interpolation is used to establish an internal resistance-capacity mapping.

[0078] The technical solution of the application realizes high-precision evaluation under sparse data by integrating physical constraints into PINN. In terms of model architecture, the technical solution of the application adopts a physical guiding module + a multi-layer perception + an offline training module + an online evaluation module to form a PINN model. In terms of constraint mechanism, the technical solution of the application adopts a dynamic combination loss function to integrate internal resistance-capacity physical mapping constraints. In terms of evaluation objects and scenarios, the technical solution of the application is aimed at UPS VRLA batteries in long-term floating and sparse data scenarios to realize single-cell-group-system multi-dimensional SOH quantitative evaluation. In terms of training and deployment processes, the technical solution of the application adopts offline training + online real-time evaluation. The application has good industrial applicability. To solve the problem of long-term floating of VRLA batteries in data center UPS systems and lack of complete charge-discharge data, the application introduces a physical constraint mechanism, effectively reducing the dependence on high-quality large-scale training data. Even in the case of less or incomplete battery data, high-precision SOH estimation can be achieved, providing technical support for the safe operation of data center power supply systems, and having good engineering promotion value and economic benefits.

[0079] The specific implementation scheme of the embodiment can be referred to the related description in the above embodiment, which will not be repeated here.

[0080] It can be understood that the same or similar parts in the above embodiments can be mutually referred to, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0081] It should be noted that, in the description of the application, the terms "first", "second", etc. are only for descriptive purposes, and cannot be understood as indicating or implying relative importance. In addition, in the description of the application, unless otherwise specified, "a plurality of" means at least two.

[0082] ​Any procedural or methodological descriptions in flow charts or otherwise described herein can be understood to represent modules, segments, or portions of code that include executable instructions for implementing the specific logical functions or steps, and the scope of preferred embodiments of the present application includes additional implementations in which the functions are performed in a different order, including substantially simultaneously, or in reverse order, as will be understood by those skilled in the art to which embodiments of the present application pertain.

[0083] It should be understood that portions of the present application can be realized with hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be realized as software or firmware to be executed by a proper

[0084] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be instructed by programs to complete the relevant hardware, and the corresponding programs can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0085] In addition, each functional unit in each embodiment of the present application can be integrated into one processing module, or each unit can exist physically independently, or two or more units can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software function module. When the integrated module is realized in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0086] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0087] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0088] The method, device, processor and computer readable storage medium thereof for evaluating the state of health of the UPS battery based on the physical information neural network adopt the application, a VRLA battery SOH evaluation model and a training method are proposed, which combines the traditional data-driven neural network with the physical law in the battery aging process, introduces the physical constraint as part of the loss function in the network training process, and effectively improves the physical consistency of the model. The method overcomes the "black box" problem of the traditional deep learning method, and significantly enhances the explainability and stability of the model.

[0089] In this specification, the application has been described with reference to its particular embodiments. It is clear, however, that various modifications and changes can be made without departing from the spirit and scope of the application. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for assessing the health status of a UPS battery based on a physical information neural network, characterized in that, The method includes the following steps: (1) Establish a physical guidance mechanism based on battery discharge test data; (2) Establish a multilayer perceptron model; (3) Offline training is performed based on physical constraints and data errors; (4) Realize online SOH assessment based on real-time battery data.

2. The method for assessing the health status of a UPS battery based on a physical information neural network according to claim 1, characterized in that, Step (1) specifically includes the following steps: By collecting data from the battery under discharge test conditions, its internal resistance and actual capacity data points are extracted, and the internal resistance is established using cubic spline interpolation. The mapping relationship between capacity and performance.

3. The method for assessing the health status of a UPS battery based on a physical information neural network according to claim 1, characterized in that, Step (3) specifically includes the following steps: (3.1) Perform data preprocessing of battery feature parameters, and use a normalizer to process all input features; (3.2) Randomly shuffle the battery feature dataset to divide it into a training set, a validation set, and a test set; (3.3) Initialize the PINN model; (3.4) The PINN model is trained by iterative optimization, and the forward and backward propagation of the PINN model is continuously executed to update the model parameters; (3.5) Save the PINN model parameters.

4. The method for assessing the health status of a UPS battery based on a physical information neural network according to claim 3, characterized in that, Step (3.3) specifically includes the following steps: (3.3.1) Define the combined loss function; (3.3.2) Weighting coefficients of physical constraint terms during training Dynamic adjustments were made; (3.3.3) A linear decay strategy is adopted, and the weighting coefficients are adjusted according to the dynamic change strategy, so that... The number of training rounds is gradually reduced.

5. The method for assessing the health status of a UPS battery based on a physical information neural network according to claim 3, characterized in that, Step (3.4) specifically includes the following steps: (3.4.1) In each round of training, a forward propagation operation is performed on the training set samples to calculate the model's predicted output; (3.4.2) Perform backpropagation to compute the gradient according to the defined composite loss function; (3.4.3) The model parameters are updated by the optimizer to gradually minimize the total loss function.

6. The method for assessing the health status of a UPS battery based on a physical information neural network according to claim 1, characterized in that, Step (4) specifically includes the following steps: (4.1) Real-time acquisition of the operating status information of each battery in each UPS power system, and extraction of key feature parameters as model input; (4.2) Online estimation of the remaining capacity of each battery based on the PINN model; (4.3) Calculate the battery SOH at each sampling point based on the remaining battery capacity evaluated by the PINN model; (4.4) Group several batteries belonging to the same UPS system battery pack according to their groups and calculate the average SOH of each battery pack; (4.5) Summarize the predicted capacity values ​​of all batteries and calculate the overall average SOH index of the system; (4.6) Summarize all statistical results and output them.

7. A system for assessing the health status of a UPS battery based on a physical information neural network, implementing the method of claim 1, characterized in that, The system includes: The physics-guided module is used to explicitly embed physical constraints into the loss function of a neural network; The multilayer perceptron module, connected to the physical guidance module, employs several fully connected feedforward neural network layers to learn the complex nonlinear relationship between input features and remaining battery capacity. An offline training module, connected to the physical guidance module, is used to train a neural network model using a battery feature parameter dataset. The online evaluation module, connected to the offline training module, is used to quickly evaluate the SOH state of the battery based on real-time collected battery characteristic parameters during the operation of the UPS system.

8. A device for assessing the health status of a UPS battery based on a physical information neural network, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for assessing the health status of a UPS battery based on a physical information neural network as described in any one of claims 1 to 6.

9. A processor for assessing the health status of a UPS battery based on a physical information neural network, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the method for assessing the health status of a UPS battery based on a physical information neural network as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method for assessing the health status of a UPS battery based on a physical information neural network, as described in any one of claims 1 to 6.

Images