Search

CN-122017578-A - Battery state prediction method based on lithium manganate characteristic modeling

CN122017578ACN 122017578 ACN122017578 ACN 122017578ACN-122017578-A

Abstract

The invention discloses a battery state prediction method based on lithium manganate characteristic modeling, which comprises the following steps of collecting battery data in real time, preprocessing to obtain preprocessed data, establishing an improved SOCNet model according to the electrochemical characteristics of the battery to obtain a residual electric quantity estimated value, an internal resistance estimated value, a health state estimated value and a capacity attenuation estimated value of the battery, calculating a comprehensive health state estimated value of the battery by adopting a weighted fusion model, dynamically adjusting a charge-discharge rate and a temperature control strategy to optimize the charge-discharge strategy of the battery, dynamically updating the improved SOCNet model after each charge-discharge through an incremental neural network, and predicting the future state of the battery through the improved TabNet model. The invention obviously improves the accuracy and adaptability of battery state prediction, provides an efficient optimization strategy for a battery management system, and is widely applied to battery management in the fields of electric automobiles, energy storage equipment and the like.

Inventors

  • HE LIQIANG
  • LIU XINGLI
  • LI ZENGPING
  • ZHANG YIFAN

Assignees

  • 赵县强能电源有限公司

Dates

Publication Date
20260512
Application Date
20260126

Claims (7)

  1. 1. The battery state prediction method based on lithium manganate characteristic modeling is characterized by comprising the following steps of: collecting voltage, current, temperature and charge-discharge cycle frequency data of a battery in real time, and preprocessing the data to obtain preprocessed data; According to the electrochemical characteristics of the lithium manganate battery, an improved SOCNet model is established, the preprocessing data are input into corresponding sub-models, and the residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity attenuation estimated value of the battery are obtained through calculation; calculating a comprehensive health state evaluation value of the battery by adopting a weighted fusion model based on the residual electric quantity evaluation value, the internal resistance evaluation value, the health state evaluation value and the capacity fading evaluation value; Based on the comprehensive health state evaluation value, optimizing the charge and discharge strategy of the battery by dynamically adjusting the charge and discharge rate, the temperature control strategy and the multi-objective optimization algorithm; Dynamically updating the improved SOCNet model after each charge and discharge through an incremental neural network, and correcting in real time to obtain updated residual electric quantity estimated value, internal resistance estimated value, health state estimated value and capacity attenuation estimated value; the future state of the battery is predicted by the modified TabNet model based on the updated remaining charge estimate, internal resistance estimate, state of health estimate, and capacity fade estimate.
  2. 2. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the method is characterized in that the voltage, current, temperature and charge-discharge cycle number data of the battery are collected in real time and are preprocessed to obtain preprocessed data, and specifically comprises the following steps: The method comprises the steps that battery data are collected in real time through a sensor, wherein the battery data comprise voltage, current, temperature and charge-discharge cycle number data of a battery, the sensor comprises a voltage sensor, a current sensor, a temperature sensor and a cycle counter, the voltage sensor collects battery terminal voltage, the current sensor collects battery charge-discharge current, the temperature sensor collects battery working temperature, and the cycle counter records the charge-discharge cycle number of the battery; Performing time sequence alignment processing on battery data, wherein the time sequence alignment processing comprises the steps of aligning voltage, current, temperature and charge-discharge cycle frequency data according to a time stamp, and performing interpolation processing on missing values, and the interpolation method comprises linear interpolation or spline interpolation; Resampling and normalizing the battery data, wherein the resampling is used for uniformly or non-uniformly resampling the collected data according to the working state of the battery, and the normalizing method comprises uniformly converting the battery data into a value between 0 and 1; the data enhancement technology is adopted to expand the battery data, and the data enhancement method comprises the operations of random disturbance, time window translation and time sequence data cutting on the current and voltage characteristics of the battery.
  3. 3. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the method is characterized in that an improved SOCNet model is built according to electrochemical characteristics of a lithium manganate battery, preprocessing data are input into corresponding sub-models, and residual electric quantity estimated values, internal resistance estimated values, health state estimated values and capacity fading estimated values of the battery are obtained through calculation, and specifically comprises the following steps: constructing an improved SOCNet model, inputting voltage, current, temperature and charge-discharge cycle number data of a battery, inputting the data into an input layer after standardized treatment, extracting preliminary features by adopting a full-connection layer, and inputting each feature into the model after standardization; constructing an electrochemical characteristic sub-network, calculating the voltage and current relationship, capacity attenuation and internal resistance dynamic change of the battery through a physical model, processing charge and discharge data of the battery by adopting a full-connection layer, fitting the nonlinear characteristics of the internal resistance and capacity attenuation of the battery by using a ReLU activation function, performing capacity attenuation prediction of the battery under time sequence data by using an LSTM layer, and calculating the residual electric quantity and the health state of the battery; Introducing a multi-task learning module, extracting common characteristics of a battery by using a sharing layer, extracting characteristics by using a multi-layer perceptron by using the sharing layer, respectively distributing different tasks to independent sub-networks for processing, respectively carrying out residual electric quantity prediction, health state evaluation and capacity attenuation prediction by independent regression sub-networks, and predicting residual electric quantity, health state and capacity attenuation values by using an output layer with a linear activation function; introducing a self-adaptive learning rate adjustment mechanism, adopting an Adam optimization algorithm as an optimizer, dynamically adjusting the learning rate according to the battery state change in the training process, and automatically adjusting the learning rate based on a learning rate scheduler; Introducing a self-attention mechanism, calculating the similarity between battery state features by using the scaling dot product attention, dynamically weighting, taking the voltage, current and internal resistance features of the battery as query, key and value, weighting and outputting the features by using an attention layer, and adaptively adjusting the contribution degree of each feature to prediction by the self-attention mechanism; the incremental learning module is added, an online learning algorithm is used, the model carries out fine adjustment on parameters according to new data when the battery state data are updated each time, and the model updates weight after each charge and discharge period by using a mini-batch gradient descent method; And adding a regularization term into the loss function, wherein the regularization term is used for restraining the improved SOCNet model parameters by using an L2 regularization technology, the regularization coefficient is selected by cross verification, and the weighting coefficient in the loss function is adjusted according to different battery working environments.
  4. 4. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the calculating the comprehensive state of health evaluation value of the battery by using a weighted fusion model based on the residual electric quantity evaluation value, the internal resistance evaluation value, the state of health evaluation value and the capacity fading evaluation value comprises the following steps: the residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity attenuation estimated value are used as input data to be transmitted into a weighted fusion model, the input data is subjected to feature extraction and linear transformation through a full connection layer, and intermediate features are output; The weighted fusion model distributes weight for each input data, nonlinear transformation is carried out on each input data by using Swish activation functions, the Swish activation functions carry out weighted processing on the input data, the weighted value of each feature is calculated by Swish activation functions, and the weighted value is dynamically adjusted based on the historical data and the current running state of the battery; adding the weighted residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity fading estimated value to obtain a comprehensive health state estimated value of the battery; Normalizing the comprehensive health state evaluation value, and converting the comprehensive health state evaluation value into a uniform numerical range; and optimizing the normalized comprehensive health state evaluation value by applying a regression analysis method, and optimizing model parameters by using a mean square error loss function.
  5. 5. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the optimizing the battery charge-discharge strategy by dynamically adjusting the charge-discharge rate, the temperature control strategy and the multi-objective optimization algorithm based on the comprehensive state of health evaluation value specifically comprises: Calculating a charge-discharge rate and a temperature control strategy of the battery based on the comprehensive health state evaluation value, wherein the charge rate is set to be maximum when the comprehensive health state evaluation value is higher than a preset threshold value, the charge rate is reduced to be smaller when the comprehensive health state evaluation value is lower than the preset threshold value so as to reduce damage to the battery, the charge current is reduced when the internal resistance evaluation value of the battery is higher than the preset value, and the temperature control strategy is used for reducing the charge rate and starting a cooling system when the capacity attenuation evaluation value exceeds the preset threshold value; Optimizing a charge-discharge rate and a temperature control strategy through a particle swarm optimization algorithm, and dynamically adjusting various parameters according to a comprehensive health state evaluation value and an internal resistance evaluation value by setting a plurality of objective functions, including maximizing charge efficiency, minimizing temperature rise and prolonging service life of a battery, and performing iterative computation on each particle in the optimization process to obtain optimal charge-discharge rate, charge current, charge duration and temperature control parameters; And adjusting the charging current, the discharging current, the charging power and the working mode of the temperature control system in real time according to the optimized charging and discharging rate, the charging current, the charging duration and the temperature control parameters.
  6. 6. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the method comprises dynamically updating the improved SOCNet model after each charge and discharge through an incremental neural network, and correcting in real time to obtain updated estimated remaining power, estimated internal resistance, estimated health state and estimated capacity attenuation, and specifically comprises: Dynamically updating through an incremental neural network, and correcting and updating an improved SOCNet model based on battery state data collected in real time after each charge and discharge to obtain an updated residual electric quantity estimated value, an internal resistance estimated value, a health state estimated value and a capacity attenuation estimated value; After each charge and discharge, collecting the voltage, current, temperature, charge and discharge time length, charge current and discharge current of the battery as incremental data, inputting the incremental data into the improved SOCNet model, and dynamically adjusting the internal weight of the improved SOCNet model by the model through processing the incremental data; Calculating an error between the output of the current improved SOCNet model and actual measured data, and updating the weight in the improved SOCNet model by using a back propagation algorithm in a back propagation mode of the error; Optimizing the network weight of the improved SOCNet model by using a random gradient descent method, calculating loss aiming at each incremental data, and adjusting the parameters of the improved SOCNet model according to the gradient of the loss; through incremental learning, the improved SOCNet model corrects the residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity attenuation estimated value in real time, and feeds back the updated estimated value in real time after each charge-discharge period is ended.
  7. 7. The battery state prediction method based on lithium manganate characteristic modeling according to claim 1, wherein the predicting the future state of the battery by the improved TabNet model based on the updated residual capacity estimation value, the internal resistance estimation value, the health state estimation value and the capacity fade estimation value specifically comprises: The updated residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity fading estimated value are used as input data to be input into the improved TabNet model; Carrying out standardized processing on input data, removing abnormal values by adopting a robust scaler method, and unifying the input data scale; Adding an adaptive weighted feature selection module into a feature selection layer, and dynamically distributing different weights for each feature by calculating the gradient contribution value of each input feature; Introducing a multi-head self-attention mechanism expansion module into a self-attention mechanism layer, and combining the outputs of a plurality of heads by calculating the relations among different feature subsets in parallel, wherein each attention head pays attention to different aspects of the battery state and gives different weights, and each head performs feature interaction independently in the training process; Introducing a differentiable decision tree module in a decision layer, learning nodes and splitting rules through a neural network, splitting by selecting the most relevant features at each node by using an information gain algorithm, outputting a future state predicted value of a battery at each leaf node, processing the output of the leaf node through a weighted average method of the neural network layer, combining a plurality of decision trees by adopting an integrated learning mode, fusing the output of each tree by adopting a weighted average method, dynamically adjusting the depth of the tree according to the complexity of the battery state, evaluating the complexity by calculating an information gain or a base index, and optimizing the depth by cross verification; jump connection is introduced, certain intermediate layers are skipped in the process of transmitting input data, and stable transmission of information in the training process is ensured; adding layering residual error connection between network layers, wherein the output of each layer is the calculation result of the current layer plus the input information of the previous layer; Training is performed by Adam optimizer and gradient descent method, and the weights of the network are updated by calculating the mean square error between the output value of the improved TabNet model and the actual battery state.

Description

Battery state prediction method based on lithium manganate characteristic modeling Technical Field The invention relates to the technical field of battery management systems, in particular to a battery state prediction method based on lithium manganate characteristic modeling. Background With the popularization of electric vehicles and energy storage devices, the real-time prediction and optimization control of battery states by a Battery Management System (BMS) becomes a research hotspot. Existing battery state prediction methods rely mainly on traditional statistical models or simple machine learning algorithms, such as Support Vector Machines (SVMs) or decision trees, which often face the following problems: The method has the advantages that the battery state prediction precision is low, the traditional method generally cannot capture complex nonlinear relations of battery state of health and capacity attenuation, the errors of prediction results of battery residual service time, state of health and the like are large, the model lacks dynamic adaptability, when the traditional method processes dynamic changes of battery state (such as fluctuation, temperature change and the like in the charging and discharging processes), the prediction results cannot be updated in real time, the adaptability of the battery in different working environments cannot be effectively processed, complex features cannot be efficiently processed, many methods depend on manual selection features, the complex relations of high-dimensional features in battery state data are ignored, the model cannot fully utilize potential information of battery data, prediction precision and efficiency are influenced, the model training process is slow, and when the traditional method processes large-scale battery data, the training process is complex, calculation cost is large, and real-time prediction requirements are difficult to meet. Therefore, how to provide a battery state prediction method based on lithium manganate characteristic modeling is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a battery state prediction method based on lithium manganate characteristic modeling, which fully utilizes battery electrochemical characteristic modeling, an improved SOCNet network structure, a dynamic updating mechanism and an improved TabNet model, calculates a comprehensive health state evaluation value through a weighted fusion method, and optimizes a battery charging and discharging strategy. The method can accurately predict the residual capacity, the internal resistance, the health state and the capacity attenuation of the battery, and has the advantages of high precision, high adaptability and real-time updating capability. According to the embodiment of the invention, the battery state prediction method based on lithium manganate characteristic modeling comprises the following steps: collecting voltage, current, temperature and charge-discharge cycle frequency data of a battery in real time, and preprocessing the data to obtain preprocessed data; According to the electrochemical characteristics of the lithium manganate battery, an improved SOCNet model is established, the preprocessing data are input into corresponding sub-models, and the residual electric quantity estimated value, the internal resistance estimated value, the health state estimated value and the capacity attenuation estimated value of the battery are obtained through calculation; calculating a comprehensive health state evaluation value of the battery by adopting a weighted fusion model based on the residual electric quantity evaluation value, the internal resistance evaluation value, the health state evaluation value and the capacity fading evaluation value; Based on the comprehensive health state evaluation value, optimizing the charge and discharge strategy of the battery by dynamically adjusting the charge and discharge rate, the temperature control strategy and the multi-objective optimization algorithm; Dynamically updating the improved SOCNet model after each charge and discharge through an incremental neural network, and correcting in real time to obtain updated residual electric quantity estimated value, internal resistance estimated value, health state estimated value and capacity attenuation estimated value; the future state of the battery is predicted by the modified TabNet model based on the updated remaining charge estimate, internal resistance estimate, state of health estimate, and capacity fade estimate. Optionally, the collecting voltage, current, temperature and charge-discharge cycle number data of the battery in real time, and preprocessing to obtain preprocessed data specifically includes: The method comprises the steps that battery data are collected in real time through a sensor, wherein the battery data comprise voltage, current, temperature and charge-discharge cycle number data of