Low-voltage power station area energy storage overall RUL prediction method

Abstract

The invention discloses an overall RUL prediction method for energy storage in a low-voltage power station area, belongs to the technical field of battery health management, and aims to solve the problems of low precision, weak generalization ability and lack of physical interpretability of an existing prediction method. According to the method, an IC curve is obtained through ampere-hour integral and differential capacity analysis on the basis of voltage and current data in the battery charging process, and characteristic parameters reflecting the aging state of the battery are extracted through Lorentz function fitting and comprise the peak area, the peak width and the symmetric center voltage of each fitting peak; constructing a deep neural network model, taking the feature parameters as input for training, and establishing a nonlinear mapping relation between the feature parameters and RUL; and deploying the trained model in an actual transformer area, and outputting an RUL predicted value of the battery on line through real-time data acquisition, feature extraction and model reasoning. The method has both physical interpretability and high prediction precision, and can effectively meet the online health monitoring requirements of the energy storage battery in the low-voltage transformer area.

Description

Technical Field

[0001] This invention belongs to the field of battery health management and predictive maintenance technology, specifically involving a method for online prediction of the remaining service life (RUL) and health status assessment of a low-voltage distribution substation energy storage system that integrates a battery electrochemical mechanism model and data-driven deep learning. Background Technology

[0002] With the increasing penetration of fluctuating loads such as distributed photovoltaic power and electric vehicle charging facilities in distribution networks, low-voltage distribution areas, as the final stage of power supply, face severe challenges to operational stability. Against this backdrop, energy storage battery systems have become indispensable key equipment in low-voltage distribution areas due to their crucial role in peak shaving, valley filling, improving power quality, and ensuring power supply reliability. However, under the complex operating conditions of distribution areas, battery performance inevitably degrades. Accurate prediction of the remaining useful life (RUL) directly affects the operation and maintenance costs, operational safety, and economic efficiency of the distribution area's energy storage system.

[0003] Currently, methods for predicting the rated uptime (RUL) of batteries can be mainly divided into three categories: models based on physical mechanisms, models based on statistical experience, and purely data-driven machine learning models. However, these methods all have significant limitations when applied to the specific scenario of low-voltage distribution areas. 1. Mechanism models rely on precise electrochemical equations and a large number of internal parameters. The models are complex and the parameters are difficult to obtain. They have poor versatility in practical engineering and are difficult to apply to battery groups with different brands, batches and aging levels in the distribution area.

[0004] 2. Statistical and empirical models are usually based on simplifying assumptions and are not adaptable enough to the complex and ever-changing aging path of batteries. They cannot accurately depict the nonlinear degradation behavior caused by factors such as load fluctuations and temperature changes in actual transformer operation, and their prediction accuracy is limited.

[0005] 3. While purely data-driven deep learning models (such as LSTM networks) can automatically learn degradation patterns from historical data, they are essentially "black box" models, lacking physical interpretability in their prediction process. Furthermore, these models rely directly on raw voltage and current time-series data, making them highly susceptible to noise and outliers from field data collection, resulting in weak generalization ability and significant prediction errors when operating conditions change.

[0006] In summary, existing technologies struggle to achieve a good balance between prediction accuracy, model generalization ability, and physical interpretability. Therefore, there is an urgent need for a new RUL prediction method that can adapt to the complex operating environment of low-voltage distribution areas, fully utilize operational data, and possess clear physical implications. Summary of the Invention

[0007] To address the shortcomings of existing Remaining Life (RUL) assessment methods for low-voltage power distribution area energy storage batteries in complex and variable operating environments, such as low prediction accuracy, weak generalization ability, and lack of physical interpretability, this invention proposes a battery RUL prediction method that integrates battery mechanism modeling and deep learning. This method enables real-time and high-precision assessment of the health status of batteries in low-voltage power distribution areas, thereby improving the operation and maintenance efficiency and power supply security of power distribution area equipment.

[0008] The technical solution adopted in this invention is as follows: A method for predicting the overall RUL (Recovery Limit) of low-voltage radio station energy storage includes the following steps: Step 1: Voltage-Capacity Model Construction and Physical Feature Extraction Based on the voltage and current data during the battery charging process, the voltage-capacity curve is calculated, and then the differential capacity curve is fitted with the Lorentz function to extract characteristic parameters reflecting the battery aging state. The characteristic parameters include the peak area, peak width and symmetry center voltage of each fitted peak. Step 2: Construct and train a deep neural network mapping model: The feature parameters obtained in step one are used as input to construct a deep neural network model and train it to establish a nonlinear mapping relationship between the feature parameters and the remaining battery life, and finally output the RUL prediction value. Step 3, Online Remaining Service Life Prediction: Based on real-time collected battery operation data, its feature parameters are extracted and input into the deep neural network model trained in step two, and the remaining service life prediction value of the battery is calculated and output online.

[0009] Furthermore, in step one, the data acquisition and preprocessing process based on the voltage and current data during the battery charging process is as follows: During battery charging, voltage and current data are sampled simultaneously, and the capacity is calculated using the ampere-hour integration method. The calculation formula is as follows: ; In the formula, The cumulative capacity represents the amount of time from the initial voltage to a specific voltage. At that time, the total amount of electricity accumulated into the battery; This is the instantaneous voltage, representing the battery terminal voltage at a specific moment during the charging process; The lower voltage limit represents the starting voltage of the charging process, which is the lowest voltage at which the battery is allowed to discharge or the voltage point at which constant current charging begins. For instantaneous current, representing the time interval... The charging current value obtained by sampling at any time; Let be time, and be the integral variable.

[0010] Furthermore, in step one, the process of calculating the voltage-capacity curve is as follows: Differentiating the voltage-capacity curve yields a differential capacity curve, whose differential capacity... The calculations can amplify and capture peak characteristics related to the electrochemical phase transition process inside the battery; Differential capacity The calculation formula is: ; In the formula, The differential capacity is represented by the vertical axis of the IC curve, which indicates the change in battery capacity caused by a unit voltage change. It can amplify and reveal the electrochemical phase transition process inside the battery. The number of peaks is the number of main peaks that can be identified in the differential capacity curve within the fitted voltage window. The peak number is the index of the summation symbol; For Peak The area of ​​represents the area of ​​the first . The area under each peak curve is directly related to the mass of the active material involved in the phase transition reaction and is a key parameter reflecting the state of health (SOH) of the battery. For Peak The width represents the width of the first... The full width of a peak at half peak height reflects the polarization and reversibility of the electrochemical reaction. As the battery ages, the peak width usually increases. For Peak The symmetrical center voltage represents the first The center voltage corresponding to each peak is closely related to the thermodynamic equilibrium potential of the battery material, and its drift can indicate changes in the electrode material structure.

[0011] Furthermore, in step one, a model consisting of multiple superimposed Lorentz functions is used to fit the differential capacity curve. The integral form of the model is expressed as: ; In the formula, The cumulative capacity calculated by the model represents the capacity calculated based on the fitted model at voltage. The cumulative battery capacity below; This is the capacity constant / offset, used to adjust the vertical position of the entire model curve so that it matches the measured voltage-capacity curve on the vertical axis.

[0012] Furthermore, in step two: the network of the deep neural network mapping model is a fully connected neural network structure, and its forward propagation process is represented as follows: and ; In the formula, Indicates the first The output vector of the hidden layer, and The input feature vector; This represents the total number of hidden layers. , The first Layer weight matrix and bias terms; The ReLU function is used as the activation function. This is a scalar value output by the network, i.e., the RUL prediction value.

[0013] Furthermore, in step two, the training process of the deep neural network mapping model is as follows: optimization is performed by minimizing the mean squared error loss function between the predicted value and the true value. This loss function is defined as: ; In the formula, This represents the set of all parameters to be optimized in the network, including all weights and biases; The total number of training samples; For the first The true RUL value of each sample; For the first RUL predicted values ​​for each sample.

[0014] Furthermore, in step three, the trained deep neural network model is applied to the actual transformer area for online and automated health monitoring: New data input: For the battery to be predicted, voltage-current data are collected in real time during the new charging cycle; Real-time feature extraction: Repeat the process in step one to extract the 9-dimensional IC peak parameter feature vector for the current period in real time; Online inference and prediction: The extracted feature vectors are input into the pre-trained DNN model; Output results: The DNN model automatically calculates and outputs the current RUL prediction value of the battery, providing a direct basis for operation and maintenance decisions.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the above-described method for predicting the overall RUL (Recovery Limits) of low-voltage radio station energy storage.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting the overall RUL (Recovery Limits) of low-voltage radio station energy storage.

[0017] The beneficial effects of this invention are: 1. High prediction accuracy and strong generalization ability: Through an interpretable physical model—Lorentz fitting—feature parameters directly related to the battery aging mechanism are extracted from the raw data, including peak area, width, and center voltage. These parameters can more fundamentally characterize the battery's health status, avoiding the sensitivity of purely data-driven models to surface noise and outliers. The combination of physical features and deep learning utilizes both the noise resistance and stability of the physical model in feature extraction and the powerful nonlinear mapping capability of deep neural networks, thus maintaining high prediction accuracy even under complex and variable actual transformer operating conditions.

[0018] 2. Excellent physical interpretability, breaking the black box of the model: Compared to end-to-end black box models such as LSTM, this invention directly links the battery aging process with internal electrochemical phase transitions through IC curve analysis and Lorentz function fitting. The extracted feature parameters have clear physical meanings, enabling maintenance personnel not only to obtain RUL prediction results but also to understand why the battery ages and which aging mechanism is dominant. This provides a direct basis for precise maintenance and greatly improves the scientific nature of decision-making.

[0019] 3. Strong anti-interference capability and suitable for complex field environments: The core of this method lies in parametric dimensionality reduction of the original data, transforming the noise-sensitive voltage-current time-series data into stable, low-dimensional physical feature vectors. This process effectively filters out noise and fluctuations that are unavoidable in field acquisition. Therefore, this method exhibits stronger adaptability and robustness to complex operating conditions commonly encountered in low-voltage distribution areas, such as load fluctuations, temperature changes, and irregular charging and discharging. The predictive performance is far less affected by changes in operating conditions than that of the comparative models.

[0020] 4. It achieves online and automated lifespan prediction with high engineering practical value: the methodology is clear and easy to deploy and automate in engineering projects. Once the model training is complete, the online prediction phase only requires real-time data collection, feature extraction, and model inference to quickly output the RUL (Lifetime Limit), meeting the real-time requirements of online health monitoring and early warning for distribution network energy storage systems. By predicting battery lifespan in advance, it can guide preventative battery replacement and tiered utilization planning, which is expected to significantly reduce the maintenance and replacement costs of batteries in distribution networks, improving the economy and safety of low-voltage distribution network operation.

[0021] In summary, this invention, through the deep integration of mechanism and data-driven approaches, successfully addresses the pain points of existing methods in achieving a balance between accuracy, interpretability, and robustness. It provides an efficient and reliable technical solution for the refined, intelligent, and interpretable health management of low-voltage distribution area energy storage batteries, and has significant engineering application value and market prospects. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart of the overall RUL prediction method for low-voltage radio station area energy storage according to the present invention; Figure 2 This is a general block diagram of the low-voltage radio station area energy storage overall RUL prediction method of the present invention; Figure 3 This is a schematic diagram illustrating the development of the deep neural network (DNN) of this invention; Figure 4 This is a schematic diagram of the Lorentz function fitting of the IC curve in this invention; Figure 5 This is a flowchart of the model training process of the present invention; Figure 6 This is a comparison chart of the RUL prediction errors of the present invention; Figure 7 This is a schematic diagram of the online deployment of the model of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To address the shortcomings of existing Remaining Usage (RUL) assessment methods for low-voltage power grid energy storage batteries operating in complex and variable environments, such as low prediction accuracy, weak generalization ability, and lack of physical interpretability, this embodiment proposes a comprehensive RUL prediction method for low-voltage power grid energy storage. The core innovation of this method lies in integrating a battery electrochemical mechanism model with a data-driven deep learning approach. By extracting feature parameters with clear physical meaning, the accuracy, robustness, and interpretability of RUL prediction are improved.

[0026] like Figure 1 As shown in the overall flowchart, this method includes three core steps: first, extracting feature parameters through a voltage-capacity physical model; then, constructing and training a deep neural network to establish the mapping relationship between the parameters and RUL; and finally, deploying the trained model in actual transformer substations for online prediction. Figure 2The overall flowchart further clarifies the inputs, processing, and outputs of each step, demonstrating the complete system-level process from raw data to the final prediction result.

[0027] The specific steps of this method for predicting the overall RUL (Range Limitless Usage) of low-voltage radio station energy storage are as follows: Step 1, Voltage-Capacity Model Construction and Physical Feature Extraction: The original voltage and current time series data are converted into low-dimensional, stable, and interpretable physical feature parameters that can directly reflect the internal electrochemical phase transition and aging state of the battery.

[0028] Voltage-capacity model construction and physical feature extraction include: Step 1.1 Data Acquisition and Preprocessing: During the constant current charging phase of the battery, the battery terminal voltage V and charging current I are synchronously acquired at a fixed frequency through the battery management system (BMS) to obtain the raw data required for model construction.

[0029] The specific operation is as follows: the capacity is calculated using the ampere-hour integration method to generate a voltage-capacity QV curve.

[0030] The calculation formula is: ; In the formula, The cumulative capacity represents the amount of time from the initial voltage to a specific voltage. At that time, the total amount of electricity accumulated into the battery; This is the instantaneous voltage, representing the battery terminal voltage at a specific moment during the charging process; The lower voltage limit represents the starting voltage of the charging process, which is the lowest voltage at which the battery is allowed to discharge or the voltage point at which constant current charging begins. For instantaneous current, representing the time interval... The charging current value obtained by sampling at any time; Let be time, and be the integral variable.

[0031] in, The charging start voltage is used; in actual operation, the voltage sequence is obtained through discrete sampling. And the corresponding capacity values, forming a data point sequence: ; in, The voltage sampling step size represents the interval when discretely sampling the voltage. The number of sampling points indicates the number of points from... arrive How many voltage points were collected in total? For the first The voltage at each sampling point represents the voltage sequence. The first in There are elements, among which =0,1,2,⋯ , ; For the first The cumulative capacity of each sampling point represents the capacity from the initial voltage. Charged to the voltage points At that time, the total amount of battery charge accumulated; ( , () represents voltage-capacity data points, which are coordinate pairs, indicating the voltage... At that time, the corresponding cumulative capacity is .

[0032] Step 1.2 Differential Capacitance IC Analysis: Numerically differentiate the QV curve and calculate... The differential capacity curve, i.e., the IC curve, is obtained.

[0033] Differential capacity The calculation formula is: ; In the formula, The differential capacity is represented by the vertical axis of the IC curve, which indicates the change in battery capacity caused by a unit voltage change. It can amplify and reveal the electrochemical phase transition process inside the battery. The number of peaks is the number of main peaks that can be identified in the differential capacity curve within the fitted voltage window. The peak number is the index of the summation symbol; For Peak The area of ​​represents the area of ​​the first . The area under each peak curve is directly related to the mass of the active material involved in the phase transition reaction and is a key parameter reflecting the state of health (SOH) of the battery. For Peak The width represents the width of the first... The full width of a peak at half peak height reflects the polarization and reversibility of the electrochemical reaction. As the battery ages, the peak width usually increases. For Peak The symmetrical center voltage represents the first The center voltage corresponding to each peak is closely related to the thermodynamic equilibrium potential of the battery material, and its drift can indicate changes in the electrode material structure.

[0034] Differential capacity IC analysis transforms smooth voltage plateaus into easily identifiable peaks that correspond one-to-one with the phase transition reactions of the active materials within the battery. For example... Figure 4 As shown in the measured IC curve, the actual measured original IC curve clearly shows multiple peaks.

[0035] Step 1.3 Lorentz Function Fitting and Parameter Identification: A model using multiple superimposed Lorentz functions is employed to fit the IC curve. Based on observations of the differential capacity curves over the battery's lifespan, such as... Figure 3 As shown, three main peaks are typically identified. .

[0036] To avoid noise introduced by numerical differentiation, parameter identification is performed directly using integral form: ; In the formula, The cumulative capacity calculated by the model represents the capacity calculated based on the fitted model at voltage. The cumulative battery capacity below; This is the capacity constant / offset, used to adjust the vertical position of the entire model curve so that it matches the measured voltage-capacity curve on the vertical axis.

[0037] The purpose of Lorentz function fitting is to mathematize and parameterize the entire IC curve and extract features with physical meaning.

[0038] Parameter identification: Nonlinear least squares (NLS) is used as the benchmark, and the optimal parameter set is found through optimization algorithms. Minimize the objective function: And satisfy the following constraints: .

[0039] In the formula, For about Minimize; This is the model parameter vector, which contains all the unknown parameters that need to be identified in the voltage-capacity model; `datapoint` is the data point index, and `sum` is the index. This iterates through all sampled points, from point 0 to point 1. One point; The total number of sampling points is represented by the total number of sampling points minus one, starting from 0 in the discretized voltage-capacity curve. For the first The measured capacity at each sampling point, under voltage The capacity value actually calculated using the ampere-hour integration method is the true observed value. For the first The voltage at the nth sampling point represents the voltage at the nth sampling point. The voltage value corresponding to each sampling point; For the capacity predicted by the model, the voltage-capacity model under given parameters and input voltage The calculated capacity prediction value; The lower bound vector of parameters is a vector whose dimension is the same as the lower bound vector of parameters. Similarly, it specifies physically reasonable lower limits for each model parameter; The parameter upper limit vector is a vector that specifies the physically reasonable upper limit value for each model parameter.

[0040] The output of this step is: a 9-dimensional feature vector for each charging cycle. .like Figure 4 As shown in the Lorentz function fitting, these parameters are directly related to the battery's health state, and their physical meaning is clear: peak area The peak width reflects the mass of the active substance participating in the phase transition reaction; This reflects the degree of polarization in the electrochemical reaction; the center voltage. It reflects the thermodynamic equilibrium potential of the material.

[0041] Step 2: Construct and train a deep neural network mapping model: Establish a complex nonlinear mapping relationship from physical characteristics to remaining battery life (RUL).

[0042] Building and training a deep neural network mapping model includes: Step 2.1 Network Design: The structure of the deep neural network mapping model is as follows: it adopts a fully connected neural network FCNN, and its structure is as follows. Figure 3 As shown.

[0043] The input layer of a deep neural network mapping model receives a 9-dimensional feature vector. ; The hidden layers of a deep neural network mapping model consist of several layers and use the ReLU activation function to introduce non-linearity. The output layer of the deep neural network mapping model outputs scalar RUL predictions. The forward propagation process of a deep neural network mapping model is as follows: ; ; In the formula, Indicates the first The output vector of the hidden layer, and The input feature vector; This represents the total number of hidden layers. , The first Layer weight matrix and bias terms; The ReLU function is used as the activation function. This is a scalar value output by the network, i.e., the RUL prediction value.

[0044] Step 2.2 Model Training: Prepare a historical dataset containing feature vectors of a large number of battery cycles and their corresponding true RUL values.

[0045] like Figure 5 As shown, the training process includes: inputting features and ground truth RUL labels; forward propagation to calculate predicted values; calculating mean squared error (MSE) loss; and backpropagation to update network parameters. Optimization objective: Minimize the loss function: In the formula, This represents the set of all parameters to be optimized in the network, including all weights and biases; The total number of training samples; For the first The true RUL value of each sample; For the first RUL predicted values ​​for each sample.

[0046] Ultimately, a DNN model that can accurately map feature parameters to RUL is obtained.

[0047] Step 3, Online Remaining Life Prediction: Achieve online and automated prediction of the Remaining Life (RUL) of energy storage batteries in actual transformer substation environments.

[0048] like Figure 7 The online deployment diagram of the model is shown below: The system architecture includes the energy storage battery in the distribution area, the data acquisition unit, the edge computing server (which deploys this prediction model), and the cloud platform monitoring center.

[0049] The workflow is as follows: New data input: Real-time acquisition of voltage-current data during the new charging cycle; Real-time feature extraction: Repeat step one on the edge server to extract the 9-dimensional feature vector for the current period; Online inference and prediction: Input the feature vectors into the trained DNN model; Output results: The model outputs RUL predicted values, which are uploaded to the cloud platform for display.

[0050] This implementation method was verified on an actual energy storage battery system in a low-voltage distribution substation. The substation was equipped with a set of lithium-ion energy storage batteries, with a nominal capacity of 100Ah per cell and an operating voltage range of 3.0V–4.2V.

[0051] During operation, the system deployed intelligent data acquisition devices that automatically recorded battery voltage, current, and other operating parameters every 5 minutes. In a periodic health assessment experiment, battery voltage-capacity (VQ) curve data were continuously collected for a complete charging cycle.

[0052] After preprocessing, 51 data points were obtained during the voltage rise from 3.0 V to 4.2 V, as shown below: (V0=3.00V, Q0=0 Ah), (V1=3.05V, Q1=10.2 Ah), ..., (V50=4.20V, Q50=98.8Ah), etc.

[0053] Subsequently, differential capacity curves were obtained through differential calculations, and preliminary fitting was performed using the least squares method to obtain three main Lorentz peaks and the fitting parameters: Area of ​​peak 1 =11.3, width =0.094, center of symmetry =3.35V; Area of ​​peak 2 =13.7, width =0.185, center of symmetry =3.76V; Area of ​​peak 3 =7.5, width =0.157, center of symmetry =4.05V.

[0054] These parameters, after normalization, are arranged into a 9-dimensional feature vector: [0.25, 0.34, 0.19, 0.31, 0.42, 0.12, 0.22, 0.36, 0.17].

[0055] Next, these parameters are input into a pre-trained deep neural network (DNN) model. This DNN model is trained based on historical failure data and typical transformer substation operating characteristics, minimizing the mean squared error between the predicted Residual Lifetime (RUL) and the actual decommissioning life. In this case, after receiving the input features, the model outputs the predicted RUL of the energy storage battery for the current cycle through activation mapping of three fully connected hidden layers. After inference, the current battery RUL is estimated to be approximately 975 days. Compared to the actual long-term observed health degradation trend, the model's prediction error is less than 5%. Figure 6 As shown in the figure, the number of cycles is used as the horizontal axis and the RUL prediction error is used as the vertical axis. The two curves show the changes in prediction error of the proposed method and the traditional deep learning method throughout the entire battery life cycle. Compared with the traditional deep learning method, the prediction error of the proposed method is significantly reduced, and it exhibits higher accuracy and stability throughout the entire battery life cycle.

[0056] Furthermore, to further verify the effectiveness, generalization ability, and superiority over existing technologies of the method of this invention, this embodiment designed and conducted a systematic comparative experiment. The experiment aimed to compare the performance of the method of this invention with current mainstream prediction models under the same test environment.

[0057] 1. Experimental setup: Comparison Models: Three representative mainstream prediction models were selected for comparison: Traditional empirical models: linear regression, support vector machine.

[0058] Single deep learning model: Long Short-Term Memory Network.

[0059] Test dataset: A single test set containing 100 sets of periodic charge-discharge data was used. All data covered a variety of real-world operating conditions to ensure fairness in the comparison. All models adopted the same standardized data preprocessing procedure.

[0060] Evaluation metrics: The quantitative evaluation is based on three key metrics: root mean square error (RMSE) of remaining useful life (RUL) prediction, mean absolute error (MAE), and accuracy of health status (SOH) identification.

[0061] To visually demonstrate the prediction details, Table 1 lists the prediction results of each model and the health status judgment for 10 representative samples in the test set.

[0062] Table 1. Performance comparison of different models on representative test samples: Sample number Actual RUL (days) The method of this invention predicts RUL (days). Linear regression prediction of RUL (days) SVM predicts RUL (days) LSTM predicts RUL (days) Health status assessment method of the present invention Linear regression for health status assessment SVM health status assessment LSTM Health Status Assessment Actual health status 1 945 962 1010 980 970 normal normal abnormal normal normal 2 876 861 820 840 864 normal abnormal normal normal normal 3 803 799 860 820 808 normal abnormal abnormal normal normal 4 745 751 698 721 770 normal abnormal abnormal normal normal 5 612 600 670 640 630 abnormal normal normal normal abnormal 6 482 470 520 500 495 abnormal abnormal abnormal abnormal abnormal 7 394 390 420 410 405 abnormal abnormal normal abnormal abnormal 8 302 316 235 280 310 abnormal abnormal abnormal abnormal abnormal 9 210 202 250 240 220 abnormal normal abnormal normal abnormal 10 105 110 130 98 115 abnormal abnormal abnormal abnormal abnormal 2. Experimental Results and Analysis: Based on the detailed cases in Table 1 and the overall statistics in Table 2, the following conclusions can be drawn: Table 2 Overall performance indicators of different models on all 100 test samples. Model RUL predicts RMSE (days) RUL predicts MAE (days) Health status recognition accuracy Method of the present invention 27.6 21.3 96.2% Linear Regression 52.8 43.5 82.5% SVM 47.1 36.9 85.8% LSTM 34.5 27.8 89.4% In terms of RUL prediction accuracy, our method significantly outperforms all comparative models; its predicted RMSE is 27.6 days and MAE is 21.3 days. In contrast, the linear regression model has an RMSE of 52.8 days and a MAE of 43.5 days; the support vector machine model has 47.1 days and 36.9 days; even the best-performing comparative model, LSTM, has an RMSE of 34.5 days and a MAE of 27.8 days, which are significantly higher than our method. The data in Table 1 provide intuitive evidence; for example, in Sample 5, our method is closest to the true RUL of 612 days over 600 days, while other models have larger prediction deviations.

[0063] In terms of the reliability of health status identification, the method of this invention demonstrates an absolute advantage, with an accuracy rate as high as 96.2%. As shown in Table 1, in the key case of Sample 5, only the method of this invention correctly identified the battery status as abnormal, while the other three comparative models all made misjudgments, which could lead to serious failure risks in actual operation and maintenance. The accuracy rate of this method, reaching 96.2%, is far higher than that of linear regression (82.5%), SVM (85.8%), and LSTM (89.4%). This result highlights that features based on the physical parameters of the IC curve can more sensitively and accurately capture changes in the battery's health status.

[0064] 3. Robustness verification under complex working conditions: To test the stability of the model in real-world complex operating environments, this experiment additionally constructed an enhanced dataset simulating extreme operating conditions: ambient temperature fluctuates between 10 and 40°C, charge / discharge rates randomly vary between 0.5C and 2C, and intermittent irregular charge / discharge occurs.

[0065] On this dataset, the performance of traditional models and single data-driven models declined significantly: the RUL prediction RMSE of linear regression and SVM rose sharply to 74.2 days and 68.5 days, respectively, and the RMSE of the LSTM model also increased to 42.7 days.

[0066] In stark contrast, the RMSE of this method is only 31.5 days, and the accuracy of health status identification remains above 93%. This fully demonstrates that the differential capacity curve parameters extracted in this invention possess extremely strong noise resistance and adaptability to operating conditions. The fundamental reason is that the IC parameters themselves originate from the electrochemical phase transition process inside the battery, which can effectively filter out surface acquisition noise and operational disturbances, mapping the complex aging mechanism into stable and interpretable feature variables, thereby endowing the subsequent neural network model with stronger generalization ability and robustness.

[0067] In summary, a series of comparative experiments and data strongly demonstrate that this method not only leads in various indicators under standard working conditions, but also exhibits excellent stability and reliability in simulating complex real-world scenarios. Its overall performance is significantly better than traditional empirical models and single deep learning models, and it has extremely high engineering application and promotion value.

[0068] The low-voltage radio station area energy storage overall RUL prediction method of the present invention can be fully implemented by computer programs and can be deployed on various computing devices.

[0069] Computer device examples In one embodiment, a computer device is provided, the device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements all the steps of the low-voltage radio station area energy storage overall RUL prediction method as described in the foregoing method embodiments of the present invention, including: constructing a voltage-capacity model based on charging data and extracting IC peak physical characteristic parameters; constructing and training a deep neural network to establish a mapping relationship between the characteristic parameters and RUL; and using the trained model to perform online RUL prediction.

[0070] The computer device can be a cloud server, an edge computing server, or a local terminal deployed in the distribution area. Its core function is to execute specific programs through a processor to implement the aforementioned methods and processes, thereby achieving automated and intelligent assessment of the remaining lifespan of the energy storage batteries in the distribution area.

[0071] Computer-readable storage medium embodiments In one embodiment, a computer-readable storage medium is provided on which a computer program is stored. When executed by a processor, the program is capable of performing all the steps in the foregoing method embodiments of the present invention.

[0072] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above.

[0073] The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, solid-state drive (SSD), or any combination thereof. The processor can be a general-purpose processor, central processing unit (CPU), graphics processing unit (GPU), digital signal processor (DSP), or programmable logic device.

[0074] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the overall RUL (Range Limit Usage) of low-voltage radio station area energy storage, characterized in that, Includes the following steps: Step 1: Voltage-Capacity Model Construction and Physical Feature Extraction Based on the voltage and current data during the battery charging process, the voltage-capacity curve is calculated, and then the differential capacity curve is fitted with the Lorentz function to extract characteristic parameters reflecting the battery aging state. The characteristic parameters include the peak area, peak width and symmetry center voltage of each fitted peak. Step 2: Construct and train a deep neural network mapping model: The feature parameters obtained in step one are used as input to construct a deep neural network model and train it to establish a nonlinear mapping relationship between the feature parameters and the remaining battery life, and finally output the RUL prediction value. Step 3, Online Remaining Service Life Prediction: Based on real-time collected battery operation data, its feature parameters are extracted and input into the deep neural network model trained in step two, and the remaining service life prediction value of the battery is calculated and output online.

2. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 1, characterized in that, In step one, the data acquisition and preprocessing process based on the voltage and current data during battery charging is as follows: During battery charging, voltage and current data are sampled simultaneously, and the capacity is calculated using the ampere-hour integration method. The calculation formula is as follows: ； In the formula, Cumulative capacity; It is the instantaneous voltage; This is the lower limit of voltage. It is the instantaneous current; For time.

3. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 2, characterized in that, In step one, the process of calculating the voltage-capacity curve is as follows: Differentiating the voltage-capacity curve yields a differential capacity curve, whose differential capacity... The calculations can amplify and capture peak characteristics related to the electrochemical phase transition process inside the battery; Differential capacity The calculation formula is: ； In the formula, The capacity is the differential capacity. The number of peaks; The peak number; For Peak The area; For Peak The width; For Peak The symmetrical center voltage.

4. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 3, characterized in that, In step one, a model consisting of multiple superimposed Lorentz functions is used to fit the differential capacity curve. The integral form of the model is expressed as: ； In the formula, The cumulative capacity calculated for the model; This represents the capacity constant / offset.

5. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 1, characterized in that, In step two: the deep neural network mapping model has a fully connected neural network structure, and its forward propagation process is represented as follows: and ; In the formula, Indicates the first The output vector of the hidden layer, and The input feature vector; This represents the total number of hidden layers. , The first Layer weight matrix and bias terms; For activation functions; This is a scalar output from the network.

6. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 5, characterized in that, In step two, the training process of the deep neural network mapping model is as follows: optimization is performed by minimizing the mean squared error loss function between the predicted and true values. This loss function is defined as: ； In the formula, This represents the set of all parameters to be optimized in the network. The total number of training samples; For the first The true RUL value of each sample; For the first RUL predicted values ​​for each sample.

7. The method for predicting the overall RUL of low-voltage radio station area energy storage according to claim 1, characterized in that, In step three, the trained deep neural network model is applied to the actual transformer area for online and automated health monitoring: New data input: For the battery to be predicted, voltage-current data are collected in real time during the new charging cycle; Real-time feature extraction: Repeat the process in step one to extract the 9-dimensional IC peak parameter feature vector for the current period in real time; Online inference and prediction: The extracted feature vectors are input into the pre-trained DNN model; Output results: The DNN model automatically calculates and outputs the current RUL prediction value of the battery, providing a direct basis for operation and maintenance decisions.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the low-voltage radio station area energy storage overall RUL prediction method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the low-voltage radio station area energy storage overall RUL prediction method as described in any one of claims 1 to 7.

Images