Lithium battery health state prediction method and system based on neural network

Abstract

The invention discloses a lithium battery health state prediction method and system based on a neural network, and the method comprises the steps: obtaining multi-source heterogeneous data of a lithium battery, and decomposing the multi-source heterogeneous data; key features are obtained according to the multi-source heterogeneous data; constructing a training set, wherein the training set comprises multi-source heterogeneous data, decomposed data and key features; constructing a multi-scale recurrent neural network model, and training the multi-scale recurrent neural network model through the training set to obtain an optimized model; and acquiring operation data of the lithium battery, and predicting the operation data through the optimized model to obtain lithium battery health state sequence data in future time.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery health state prediction, and in particular relates to a lithium battery health state prediction method and system based on neural network. Background Art

[0002] There is a key technical problem in the prediction of lithium battery health status, that is, how to minimize the computational complexity and resource consumption of the prediction model while ensuring the prediction accuracy. At present, the industry generally adopts the neural network-based method for lithium battery health prediction. However, as the working conditions of lithium batteries become increasingly complex and changeable, the traditional neural network model has been unable to meet the requirements of real-time and robustness.

[0003] Specifically, lithium battery health prediction requires the integration of multi-source heterogeneous data such as voltage, current, and temperature, and the consideration of external factors such as charge and discharge rate and ambient temperature. In addition, key parameters such as capacity and internal resistance of lithium batteries change dynamically over time, so the prediction model needs to be able to learn and update adaptively.

[0004] On the other hand, lithium battery health prediction also places high demands on the computational efficiency of the model. In practical applications, a large number of batteries often need to be monitored and predicted in real time, so the prediction model must be able to run efficiently on resource-constrained embedded devices. However, current deep neural network models usually have high computational complexity and are difficult to achieve real-time prediction on low-power devices.

[0005] Therefore, it is urgent to study a computationally efficient and robust lithium battery health prediction method. This method needs to be able to adaptively extract and fuse multi-source sensor data, build a dynamically updated prediction model, and minimize computational overhead while ensuring prediction accuracy to meet the needs of real-time prediction. This requires systematic innovation in neural network architecture design, adaptive learning algorithms, model compression, etc., in order to break through the existing technical bottlenecks and achieve efficient and reliable lithium battery health status prediction. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a method and system for predicting the health status of a lithium battery based on a neural network to solve the problems existing in the above prior art.

[0007] To achieve the above object, the present invention provides a method for predicting the health status of a lithium battery based on a neural network, comprising:

[0008] Acquire multi-source heterogeneous data of lithium batteries and decompose the multi-source heterogeneous data; acquire key features based on the multi-source heterogeneous data; construct a training set, wherein the training set includes the multi-source heterogeneous data, the decomposed data and the key features;

[0009] Construct a multi-scale recurrent neural network model, train the multi-scale recurrent neural network model through the training set to obtain an optimized model; obtain the operating data of the lithium battery, predict the operating data through the optimized model, and obtain the lithium battery health status sequence data.

[0010] Preferably, the multi-source heterogeneous data includes voltage data, current data and temperature data.

[0011] Preferably, the key features include the battery capacity decay rate and the area under the discharge voltage curve, wherein the battery capacity decay rate is calculated based on the initial charging capacity and the current capacity, and the area under the discharge voltage curve is obtained by statistically plotting the voltage data.

[0012] Preferably, the multi-scale recurrent neural network model adopts a long short-term memory network.

[0013] Preferably, the training set is further divided into data under different operating conditions, wherein different operating conditions are combinations of different discharge rates, temperatures and charge-discharge cycle strategies.

[0014] Preferably, the multi-source heterogeneous data are decomposed by a time series decomposition algorithm, and the decomposed data include trend items, period items and random items of the voltage data and the current data.

[0015] Preferably, the multi-scale recurrent neural network model also includes an attention mechanism, wherein the attention mechanism is added to the output end of the long short-term memory network, and weights are assigned to the health status sequence data output by the long short-term memory network to obtain the health status sequence data at a future time.

[0016] Preferably, the process of training the multi-scale recurrent neural network model includes:

[0017] The multi-scale recurrent neural network model is trained through an adaptive learning rate adjustment strategy, and the trained multi-scale recurrent neural network model is lightweight set and iteratively optimized to obtain an optimized model.

[0018] On the other hand, the present invention also provides a lithium battery health status prediction system based on a neural network, which is used to execute the above-mentioned lithium battery health status prediction method based on a neural network.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects:

[0020] The present invention discloses a method for predicting the health status of lithium batteries. A multi-scale recurrent neural network model integrating external factors is constructed to capture the dynamic change trend of battery capacity and internal resistance. On this basis, a lightweight long short-term memory network is designed to achieve selective memory and forgetting of information through a gating mechanism. In order to capture the long-term dependency of battery aging and degradation, a sequence prediction model of an attention mechanism is introduced. In response to changes in operating conditions, an adaptive learning mechanism is used to adjust model parameters. The present invention realizes accurate prediction of the health status of lithium batteries through multi-level deep learning model design, overcomes the problem that traditional methods are difficult to cope with complex operating conditions and long-term degradation, and improves the robustness and generalization ability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0022] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0023] Figure 2 The present invention is a flowchart of a specific implementation of the method of the embodiment of the present invention. DETAILED DESCRIPTION

[0024] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0025] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0026] like Figure 1-2 As shown, this embodiment provides a method for predicting the health status of a lithium battery based on a neural network, including:

[0027] Acquire multi-source heterogeneous data of lithium batteries and decompose the multi-source heterogeneous data; acquire key features based on the multi-source heterogeneous data; construct a training set, wherein the training set includes the multi-source heterogeneous data, the decomposed data and the key features;

[0028] Construct a multi-scale recurrent neural network model, train the multi-scale recurrent neural network model through the training set to obtain an optimized model; obtain the operating data of the lithium battery, predict the operating data through the optimized model, and obtain the lithium battery health status sequence data.

[0029] Specifically, the above contents include:

[0030] Step S1, obtaining multi-source heterogeneous data of lithium batteries including voltage data, current data and temperature data, and performing preprocessing and data decomposition operations on the multi-source heterogeneous data of lithium batteries to provide good data for the data model and provide a high-quality data foundation for subsequent data-related specific identification, thereby overcoming the problem that artificial features are difficult to characterize the battery health degradation mechanism.

[0031] Step S2, through the original data and decomposed data in the above step S1, combined with key features, construct a multi-scale recurrent neural network model that integrates the external factors of charge and discharge rate and ambient temperature to predict the current lithium battery health status sequence data and capture the dynamic change trend of the lithium battery health status.

[0032] Step S3, based on the output of the multi-scale recurrent neural network model, a lightweight long short-term memory network structure is designed to achieve selective memory and forgetting of information through a gating mechanism, thereby reducing the number of parameters and computational complexity of the model.

[0033] Step S4, in view of the long-term change trend of the health status of the lithium battery, an attention mechanism is introduced on the basis of the long short-term memory network to form a sequence prediction model, and the health status sequence data of the lithium battery in the future is predicted by weighted aggregation of historical health status information to capture the long-term dependency of battery aging and degradation.

[0034] Step S5, in response to changes in the working conditions of the lithium battery, an adaptive learning mechanism is introduced to adjust the training strategy and hyperparameters of the sequence prediction model according to the real-time working status of the battery to improve the robustness and generalization ability of the prediction.

[0035] The above technical solution is described in detail:

[0036] Step S1, obtaining multi-source heterogeneous data of lithium batteries including voltage data, current data and temperature data, and performing preprocessing and data decomposition operations on the multi-source heterogeneous data of lithium batteries to provide good data for the data model and provide a high-quality data foundation for subsequent data-related specific identification, thereby overcoming the problem that artificial features are difficult to characterize the battery health degradation mechanism.

[0037] For the multi-source heterogeneous time series data of lithium batteries, the missing values ​​and outliers are processed through data preprocessing to obtain normalized time series data. Then, the time series decomposition algorithm is used to decompose the normalized lithium battery time series data into trend items, period items and random items, and extract multi-scale time series features.

[0038] Specifically, first, it is necessary to collect multi-source heterogeneous time series data such as voltage, current, and temperature of lithium batteries. For example, the voltage, current, and temperature data of the battery can be collected every second and stored in the database. These data form the basis for evaluating the health status of the battery. The reason for collecting multi-source data is that a single data source may not be able to fully reflect the health status of the battery. Data such as voltage, current, and temperature reflect the electrochemical process inside the battery from different aspects. Only by combining them for analysis can the health status of the battery be more accurately evaluated. Next, the collected data is preprocessed. For example, if the temperature data at a certain time point is missing, it can be filled with the average value of the temperature data at the two time points before and after the time point. If the current data at a certain time point is found to be abnormally high, it may be caused by measurement error or external interference, and the data point can be removed. The purpose of data preprocessing is to eliminate noise and errors in the data, improve data quality, and lay the foundation for subsequent feature extraction and model construction. Then, the time series decomposition algorithm is used to decompose the preprocessed time series data into trend items, period items, and random items by additive decomposition. For example, the voltage data of a battery may show a slow downward trend, which is a reflection of the battery capacity decay; at the same time, the voltage data and current data may also have certain periodic fluctuations, which are caused by the battery charge and discharge cycle. Through time series decomposition, these features of different scales can be extracted to better understand the operating status of the battery. The decomposed trend term reflects the long-term decay trend of the battery, the periodic term reflects the charge and discharge cycle characteristics of the battery, and the random term reflects the short-term fluctuations and noise of the battery.

[0039] Step S2, through the original data and decomposed data in the above step S1, combined with key features, construct a multi-scale recurrent neural network model that integrates the external factors of charge and discharge rate and ambient temperature to predict the current lithium battery health status sequence data and capture the dynamic change trend of the lithium battery health status.

[0040] The key features that reflect the change trend of lithium battery capacity and internal resistance are extracted from the preprocessed lithium battery data. Based on the extracted features, a multi-scale recurrent neural network model that integrates external factors such as charge and discharge rate and ambient temperature is constructed. During the model training process, different time scales are set to capture the dynamic change trend of the health status of lithium batteries in the short and long term. The trained multi-scale recurrent neural network model is tested and evaluated to analyze the prediction performance of the model under different working conditions.

[0041] Specifically, obtaining data on lithium batteries under different operating conditions is the basis for building an accurate and reliable battery health status assessment model. The voltage, current and temperature data of the battery at different discharge rates (such as 0.5C, 1C, 2C) and ambient temperatures (such as 0℃, 25℃, 45℃) can be collected. In order to simulate real usage scenarios, different charge and discharge cycle strategies can be designed, such as constant current charging, constant voltage charging, and intermittent discharge. With different discharge rates, temperatures and charge and discharge cycle strategies, a lithium battery data set containing information on various operating conditions can be constructed, providing sufficient data support for subsequent feature extraction and model training. Preprocessing the collected lithium battery data is a key step to ensure data quality. Through the above preprocessing scheme, the above data is preprocessed and time series decomposed to obtain a standardized and high-quality lithium battery data set.

[0042] In addition, the battery capacity decay rate and the area under the discharge voltage curve are calculated as key features to supplement the data set, which can reflect the health status of the battery.

[0043] The multi-scale recurrent neural network model can capture the dynamic change trend of lithium battery capacity and decomposed features at different time scales. For example, the short-term time scale can be set as one charge and discharge cycle, and the long-term time scale can be set as multiple charge and discharge cycles. On the short-term time scale, the model can learn the voltage, current and temperature change rules of the battery in a single charge and discharge cycle. On the long-term time scale, the model can learn the decay trend of battery capacity and the characteristic law of decomposed features. By integrating information at different time scales, the health status of the battery can be more comprehensively characterized. In order to verify the effectiveness of the model, the trained multi-scale recurrent neural network model needs to be tested and evaluated. The data set can be divided into a training set and a test set, the training set is used to train the model, and the test set is used to evaluate the prediction performance of the model. Common evaluation indicators include mean square error, mean absolute error, and determination coefficient.

[0044] By analyzing the prediction performance of the model under different working conditions, the generalization ability and robustness of the model can be evaluated. For example, the prediction accuracy of the model under different discharge rates and ambient temperatures can be compared. If the model can maintain a high prediction accuracy under various working conditions, it means that the model has good generalization ability.

[0045] Based on the characteristics obtained above, a multi-scale recurrent neural network model is constructed that integrates external factors such as charge and discharge rate and ambient temperature.

[0046] The original data, key features, and preprocessed decomposed features are used as multi-scale features. A multi-scale recurrent neural network model is constructed through a recurrent neural network, and the multi-scale features are used as input to predict the health status and remaining life of the battery. During the model training process, an attention mechanism is introduced to adaptively adjust the weights of different features to improve the generalization ability of the model. An adaptive learning rate adjustment strategy is designed to dynamically adjust the learning rate according to the performance of the model on the validation set to accelerate model convergence and avoid overfitting. A fusion mechanism is constructed to weightedly combine the output of the recurrent neural network with the estimated results of other physical models to obtain a more robust prediction result.

[0047] Specifically, the key features extracted from the original data and the decomposed features are the basis for predicting the health status and remaining life of the battery. These characteristic parameters need to be able to reflect the changing trend of battery performance. The above original features, key features and decomposed features include: battery voltage data, current data, temperature data, data after voltage data and current data are decomposed through time series, area under the discharge voltage curve, capacity decay rate, etc. These features are not independent of each other. For example, the decrease in the area under the discharge voltage curve is usually accompanied by capacity decay. At the same time, the ambient temperature will also affect the performance of the battery. High or low temperature will accelerate the aging of the battery. Suppose a set of data is collected, showing that at an ambient temperature of 25 degrees Celsius, a battery is discharged at a rate of 1C, with an initial voltage of 4.2V. After 1 hour of discharge, the voltage drops to 3.7V, and the discharge capacity is 2Ah. For the same battery, the initial voltage is 4.2V when discharged at the same rate at an ambient temperature of 45 degrees Celsius, but the voltage drops to 3.7V after 45 minutes of discharge, and the discharge capacity is only 1.8Ah. This shows that the high temperature environment accelerates the decay of the battery capacity.

[0048] In order to capture the changing trends of battery performance at different time scales, a time series prediction model is constructed using a recurrent neural network, which can effectively learn the changing patterns of battery performance over time. Using multi-scale features as input to the recurrent neural network allows the model to capture both short-term and long-term trends in the battery. For example, a recurrent neural network can use a long short-term memory network (LSTM) to construct a prediction model, using multi-scale features as input to the LSTM to predict the health status and remaining life of the battery.

[0049] The adaptive learning rate adjustment strategy can dynamically adjust the learning rate according to the performance of the model on the validation set, accelerate model convergence and avoid overfitting. For example, when the performance of the model on the validation set improves slowly, the learning rate can be reduced; and when the performance of the model on the validation set decreases, the learning rate can be increased. The fusion mechanism can weightedly combine the output of the recurrent neural network with the estimated results of other physical models to obtain a more robust prediction result. For example, the battery life model based on the empirical formula can be fused with the output of the recurrent neural network model to improve the prediction accuracy. The online incremental learning algorithm enables the model to be updated in time when new charging and discharging data arrives, adapting to the long-term change trend of battery performance. For example, algorithms such as online gradient descent can be used to update model parameters so that the model can continuously learn new data and maintain the accuracy of prediction. Assuming that the performance of the battery has changed after a period of use, through online incremental learning, the model can adjust parameters according to the new data to adapt to the new performance characteristics of the battery.

[0050] Step S3, based on the output of the multi-scale recurrent neural network model, a lightweight long short-term memory network structure is designed to achieve selective memory and forgetting of information through a gating mechanism, thereby reducing the number of parameters and computational complexity of the model.

[0051] According to the output results of the multi-scale recurrent neural network model, the initial LSTM network structure parameters are obtained; for the initial LSTM network structure parameters, the selective memory and forgetting process of information is controlled through the gating mechanism, and the memory and forgetting rate of the network are dynamically adjusted; under the control of the gating mechanism, the LSTM network structure is iteratively optimized to gradually reduce the number of parameters and computational complexity of the model; the gradient pruning algorithm is used to further lightweight the optimized LSTM network structure and compress redundant parameters; through the knowledge distillation technology, the knowledge of the lightweight LSTM network structure is transferred to a smaller student model to further reduce the model size; on the premise of ensuring the performance of the model, the student model is quantized to reduce the precision of the model parameters and intermediate features from floating point numbers to fixed point numbers to reduce storage and computing overhead; the quantized student model is deployed on the target hardware platform to achieve efficient reasoning of the lightweight LSTM network structure.

[0052] Specifically, the multi-scale recurrent neural network model will output a large amount of prediction result data in the process of predicting the health status and remaining life of lithium batteries. These prediction results can be used as a reference for the initial long short-term memory network (LSTM) structure parameters. For example, the multi-scale recurrent neural network model predicts that the battery capacity decay trend in the next week is relatively gentle, while the battery capacity decay trend in the next week is more obvious. Then, the initial forgetting rate and memory strength of the memory unit in the LSTM network can be set according to these trends. For example, for short-term predictions, a higher forgetting rate can be set to quickly respond to short-term changes in battery status; for long-term predictions, a lower forgetting rate can be set to better capture the long-term evolution trend of battery status. The core of the LSTM network is the gating mechanism, which can control the selective memory and forgetting process of information. For example, suppose the LSTM network is used to predict the capacity change of the battery at different temperatures. When the ambient temperature rises from 25℃ to 30℃, the input gate will open, allowing new temperature information to enter the memory unit and update the predicted value of the battery capacity. When the temperature changes little, the input gate will be closed to avoid unnecessary calculation and storage overhead. At the same time, the forget gate will selectively forget the temperature information of the past period of time according to the amplitude and duration of the temperature change. For example, if the temperature continues to remain at around 30°C, the forget gate will gradually forget the previous temperature information of 25°C. The output gate will output the final battery capacity prediction value based on the current temperature information and the state of the memory unit.

[0053] Under the control of the gating mechanism, the LSTM network structure can be iteratively optimized. For example, the initial LSTM network may contain multiple hidden layers and a large number of parameters. Through iterative optimization, the number of hidden layers and the scale of parameters can be gradually reduced, thereby reducing the computational complexity of the model. For example, the initial network has four hidden layers, each with 100 neurons. After optimization, it can be reduced to two hidden layers, each with 50 neurons, while ensuring the prediction accuracy, while reducing the amount of computation and storage requirements.

[0054] The gradient pruning algorithm can further lighten the LSTM network structure. The core idea of ​​the algorithm is to judge the importance of the parameter based on its gradient size and prune those redundant parameters with small gradients. For example, if the gradient of a weight parameter in a LSTM unit is consistently close to zero, it means that the parameter has little effect on the prediction result of the model and can be pruned to reduce the size of the model.

[0055] Knowledge distillation technology can transfer the knowledge of the lightweight LSTM network structure to a smaller student model. For example, an LSTM network with two hidden layers and 50 neurons in each layer is used as the teacher model, and an LSTM network with only one hidden layer and 25 neurons is trained as the student model. During the training process, the student model not only learns the characteristics of the training data, but also learns the output of the teacher model, thereby obtaining stronger generalization ability. Under the premise of ensuring the performance of the model, the student model can be quantized. For example, the accuracy of model parameters and intermediate features can be reduced from 32-bit floating point numbers to 16-bit fixed point numbers, or even lower 8-bit fixed point numbers. This can significantly reduce the storage and computing overhead of the model. For example, the parameter size is directly reduced by three quarters when 32-bit floating point numbers are reduced to 8-bit fixed point numbers. Finally, the quantized student model is deployed on the target hardware platform, such as an embedded system or mobile device. Due to the small size of the model and low computing overhead, efficient reasoning can be achieved. For example, on resource-constrained mobile devices, millisecond-level battery capacity and internal resistance prediction can be achieved, providing real-time data support for battery health management and maintenance.

[0056] Step S4, in view of the long-term change trend of the health status of the lithium battery, an attention mechanism is introduced on the basis of the long short-term memory network to form a sequence prediction model, and the health status sequence data of the lithium battery in the future is predicted by weighted aggregation of historical health status information to capture the long-term dependency of battery aging and degradation.

[0057] The multi-scale recurrent neural network model using the long short-term memory network is used as the basic model, and the gating mechanism is used to model and memorize long-term dependencies. An attention mechanism is introduced on the basis of the long short-term memory network, and the importance of different time steps is dynamically adjusted by weighted aggregation of historical health status information to form a sequence prediction model. According to the aging and degradation characteristics of lithium batteries, a decay factor is designed to give a smaller weight to historical states that are far away from the current moment, highlighting the impact of recent states. A loss function is constructed, and the sparsity of prediction errors and attention weights is comprehensively considered to optimize the model parameters through the gradient descent algorithm. During the training process, an early stopping mechanism and regularization technology are used to avoid overfitting of the model and improve generalization ability. The trained model is used to predict the new lithium battery health status sequence, and the remaining life and health level of the battery are evaluated based on the prediction results to provide decision support for battery management and maintenance.

[0058] Specifically, the long short-term memory network (LSTM) is trained using a data set. The long short-term memory network (LSTM) can effectively process time series data because it has a gating mechanism that can selectively remember and forget information. The input gate controls which new information needs to be remembered, the forget gate controls which old information needs to be forgotten, and the output gate controls which information needs to be output. Through these gating mechanisms, LSTM can effectively capture the long-term dependencies of the battery health state. At the output end of the long short-term memory network, an attention mechanism is added after the output end, and a fully connected layer is added after the attention mechanism. The final result is output through the fully connected layer to form a sequence prediction model, so that the sequence prediction model can be used to obtain the prediction results of the health state in a later period of time. At this time, the prediction results are more accurate, and the attention mechanism can further improve the performance of the LSTM model. It can perform weighted aggregation on historical health status information and dynamically adjust the importance of different time steps. For example, if the battery's recent health state has experienced abnormal fluctuations, the attention mechanism will give these time steps higher weights so that the model pays more attention to these important information. A specific example is that if the battery is overcharged on a certain day, then the health status data of that day is very important, and the attention mechanism will give a higher weight to that day. Considering the aging and degradation characteristics of lithium batteries, a decay factor can be designed. For example, an exponential decay function can be used to give a smaller weight to the historical state that is far away from the current moment, highlighting the impact of the recent state. For example, today's health status has a weight of 1, yesterday's weight is 0.9, the weight of the day before yesterday is 0.81, and so on. This can better reflect the impact of the battery's recent state on the future health state. The design of the loss function of the sequence prediction model needs to comprehensively consider the sparsity of the prediction error and the attention weight. For example, the mean square error can be used as a measure of the prediction error, and the L1 regularization term can be added to constrain the sparsity of the attention weight. This can encourage the model to focus only on the most important historical state information and avoid overfitting. During the training process, early stopping mechanisms and regularization techniques, such as L2 regularization, can be used to avoid model overfitting and improve generalization ability. The early stopping mechanism means that when the performance of the model on the validation set no longer improves, training is stopped. Regularization technology can penalize the complexity of the model and prevent the model from overfitting the training data. The trained model can be used to predict new lithium battery health status sequences. For example, by inputting the battery usage plan for the next week, the model can predict the battery health status for the next week and evaluate the remaining life and health level of the battery based on the prediction results. These prediction results can provide decision support for battery management and maintenance, such as when the battery needs to be replaced and how to optimize the battery usage strategy. A specific example is that if the model predicts that the battery health status will drop below a critical value in the next week, the system can issue an early warning to remind the user to replace the battery in time.To construct the loss function, the sequence prediction model also comprehensively considers the sparsity of prediction error and attention weight, and optimizes the model parameters through the gradient descent algorithm.

[0059] Step S5, in response to changes in the working conditions of the lithium battery, an adaptive learning mechanism is introduced to adjust the training strategy and hyperparameters of the sequence prediction model according to the real-time working status of the battery to improve the robustness and generalization ability of the prediction.

[0060] At the same time, in order to better adapt to different working conditions, the real-time working status data of the lithium battery is obtained to determine whether the current working conditions have changed. If so, the adaptive learning mechanism is triggered. After the adaptive learning mechanism is triggered, the training strategy of the sequence prediction model is dynamically adjusted, including adjusting the time window size and update frequency of the training data set. At the same time, the adaptive learning mechanism also dynamically adjusts the hyperparameters of the sequence prediction model, such as the learning rate and regularization coefficient, to adapt to the current working conditions. The adjusted training strategy and hyperparameters are applied to the sequence prediction model, and the model is retrained and optimized using the real-time working status data of the lithium battery. Through adaptive learning and dynamic tuning, the sequence prediction model can adapt to changes in the working conditions of the lithium battery and improve the robustness of the prediction. The optimized sequence prediction model is evaluated under different working conditions to verify its generalization ability to ensure that the model can accurately predict the state and performance of the lithium battery.

[0061] Specifically, the real-time working status data of the lithium battery, such as voltage, current, temperature, etc., are obtained and arranged in chronological order to form a time series data set. For example, the voltage, current and temperature data of the battery can be collected every 1 second, and these data can be stored in a database to form a time series. According to the real-time data obtained, it is judged whether the current working condition has changed. The change of working condition can refer to the change of the working environment of the battery, such as a sharp increase or decrease in temperature, a sudden increase or decrease in load, or a change in the above working condition combination. There are many ways to judge whether the working condition has changed. For example, a threshold can be set. When a parameter exceeds the set threshold, it is considered that the working condition has changed. For example, if the battery temperature rises by 10 degrees Celsius in a short period of time, or the current suddenly increases by 2 amperes, it is considered that the working condition has changed. When the working condition changes, the adaptive learning mechanism is triggered. The core idea of ​​the adaptive learning mechanism is to dynamically adjust the parameters and training strategy of the model according to the current working condition. For example, if the current working condition is high temperature and high load, then the training data set of the model should focus on the data under high temperature and high load. After the adaptive learning mechanism is triggered, the training strategy of the sequence prediction model needs to be adjusted dynamically. For example, the time window size of the training data set can be adjusted. If the working condition changes drastically, the time window should be reduced so that the model can adapt to the new working condition more quickly. On the contrary, if the working condition changes slowly, the time window can be appropriately enlarged so that the model can learn longer-term laws. Assuming that the battery switches from a light-load condition to a heavy-load condition, the time window can be shortened from the previous 1 hour to 15 minutes so that the model can learn the battery characteristics under the heavy-load condition more quickly. If the battery is in a stable condition of constant temperature and constant current for a long time, the time window can be extended from 1 hour to 2 hours so that the model can learn longer-term attenuation trends. The update frequency also needs to be adjusted. The faster the working condition changes, the higher the update frequency should be. For example, under heavy-load conditions, the model can be updated every 5 minutes. Under light-load conditions, the model can be updated every 1 hour. In addition, the adaptive learning mechanism also needs to dynamically adjust the hyperparameters of the sequence prediction model. For example, the learning rate and the regularization coefficient. The learning rate controls the speed at which the model parameters are updated. When the operating conditions change dramatically, the learning rate should be increased so that the model can quickly adapt to the new conditions. Otherwise, the learning rate should be decreased. Assuming that the battery switches from a low temperature environment to a high temperature environment, the learning rate can be increased from 0.001 to 0.01 so that the model can adapt to the battery characteristics in a high temperature environment more quickly. If the battery is in a stable operating condition for a long time, the learning rate can be reduced from 0.01 to 0.001 to avoid model oscillation. The regularization coefficient controls the complexity of the model and can prevent the model from overfitting. When the operating conditions change dramatically, the regularization coefficient should be reduced so that the model can learn more complex laws. Otherwise, the regularization coefficient should be increased.Apply the adjusted training strategy and hyperparameters to the sequence prediction model, and retrain and optimize the model using the real-time working status data of the lithium battery. For example, the data in the new time window can be added to the training set, and the model can be retrained using the updated learning rate and regularization coefficient. Through adaptive learning and dynamic tuning, the sequence prediction model can adapt to the changes in the working conditions of the lithium battery and improve the robustness of the prediction. For example, the model can accurately predict the health status of the battery under different working conditions such as temperature and load. Evaluate the optimized sequence prediction model under different working conditions. For example, the model can be applied to different working conditions such as high temperature, low temperature, high load, and low load, and the prediction accuracy of the model can be evaluated. The generalization ability of the model is verified by the evaluation results to ensure that the model can accurately predict the state and performance of the battery under various working conditions. For example, the prediction errors of the model under different working conditions can be compared. If the errors are all within an acceptable range, it means that the model has good generalization ability.

[0062] On the other hand, the present invention also provides a lithium battery health status prediction system based on a neural network, characterized in that it is used to execute the above-mentioned lithium battery health status prediction method based on a neural network.

[0063] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting the health status of a lithium battery based on a neural network, characterized in that: include: Obtain multi-source heterogeneous data of lithium batteries and decompose the multi-source heterogeneous data; obtain key features based on the multi-source heterogeneous data; Constructing a training set, wherein the training set includes multi-source heterogeneous data, decomposed data, and key features; Construct a multi-scale recurrent neural network model, train the multi-scale recurrent neural network model through the training set to obtain an optimized model; obtain the operating data of the lithium battery, predict the operating data through the optimized model, and obtain the health status sequence data of the lithium battery in the future.

2. The method according to claim 1, characterized in that The multi-source heterogeneous data includes voltage data, current data and temperature data.

3. The method according to claim 1, characterized in that The key features include the battery capacity decay rate and the area under the discharge voltage curve, wherein the battery capacity decay rate is calculated based on the initial charging capacity and the current capacity, and the area under the discharge voltage curve is obtained by statistically plotting the voltage data.

4. The method according to claim 1, characterized in that: The multi-scale recurrent neural network model adopts a long short-term memory network.

5. The method according to claim 1, characterized in that: The training set is further divided into data under different working conditions, wherein different working conditions are combinations of different discharge rates, temperatures, and charge-discharge cycle strategies.

6. The method according to claim 1, characterized in that The multi-source heterogeneous data are decomposed by time series decomposition algorithm, and the decomposed data include trend items, period items and random items of voltage data and current data.

7. The method according to claim 1, characterized in that An attention mechanism is also introduced after the multi-scale recurrent neural network model, wherein the attention mechanism is added to the output end of the multi-scale recurrent neural network model, and weighted aggregation is performed on the health status sequence data output by the multi-scale recurrent neural network model to obtain the health status sequence data at a future time.

8. The method according to claim 1, characterized in that: The process of training a multi-scale recurrent neural network model includes: The multi-scale recurrent neural network model is trained through an adaptive learning rate adjustment strategy, and the trained multi-scale recurrent neural network model is lightweight set and iteratively optimized to obtain an optimized model.

9. A lithium battery health status prediction system based on neural network, characterized in that: Used to execute the neural network-based lithium battery health status prediction method described in any one of claims 1-8.

Images