LSTM-BP power battery temperature prediction method and system based on fusion average pooling

Abstract

The application claims a kind of LSTM-BP power battery temperature prediction method and system based on fusion average pooling, belong to new energy automobile battery thermal management technical field, the method includes the following steps: obtaining the time series data of battery operation and pretreatment;Sequence sample is constructed by sliding window;The prediction model is connected by double-layer LSTM network, average pooling layer and shallow BP network in turn;Model is trained using sample;Temperature prediction is carried out using trained model.The application captures long-term time series dependence of temperature evolution through LSTM network, globally fuses all time step information through average pooling mechanism, and then carries out nonlinear mapping through shallow BP network, realizes high-precision, strong generalization ability temperature prediction.The method avoids overfitting that complex attention mechanism may bring, balances model performance and complexity, and provides effective technical support for intelligent thermal management of power battery.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle battery thermal management technology, specifically to an LSTM-BP power battery temperature prediction method based on fused average pooling. Background Technology

[0002] Power batteries are the core component of new energy vehicles, and their operating temperature directly affects their performance, safety, and lifespan. Accurate temperature prediction is crucial for improving the performance of battery thermal management systems and can effectively prevent the risk of thermal runaway. The thermal process of power batteries exhibits strong nonlinearity, time-varying characteristics, and long-term time dependence. Existing technologies for battery temperature prediction mainly include physical mechanism-based modeling and data-driven modeling. Physical models rely on precise partial differential equations, whose parameters are difficult to determine under complex real-world conditions, limiting their applicability. Data-driven methods, especially neural network models, have become the mainstream research direction in this field. However, existing data-driven prediction schemes still have many limitations.

[0003] Traditional static neural network models suffer from insufficient prediction accuracy. For example, backpropagation neural networks lack temporal modeling capabilities, making it difficult to effectively capture the long-term dependencies of battery temperature changes. This results in significant errors and insufficient robustness when handling temperature prediction tasks under dynamic operating conditions. For instance, even though the publicly disclosed CN116933651A is applied to sensor layout optimization, its core still relies on static BP networks for spatial interpolation prediction, failing to address the fundamental problem of temporal prediction. Single temporal models do not fully utilize global information. While temporal models such as Long Short-Term Memory networks can capture long-term dependencies, traditional applications typically rely only on the hidden state of the last time step for prediction, failing to fully integrate the global historical information of the entire input time series. This limits the model's ability to grasp the overall temperature evolution trend and makes it sensitive to data noise. Hybrid models have complex structures, singular optimization objectives, and are prone to overfitting. Existing technologies attempt to improve neural network prediction performance by optimizing algorithms. For example, CN119598840A discloses the use of an improved PSO-ABC algorithm to optimize the initial weights and thresholds of a BP neural network. However, its core is to optimize the parameters of a BP network, without embedding temporal modeling capabilities. Essentially, it is still a static mapping, and its complex PSO-ABC-Q-learning hybrid optimization mechanism brings high computational costs and the risk of overfitting. CN116029183A uses an improved particle swarm optimization algorithm to optimize the hyperparameters of an LSTM. Its optimization algorithm is itself rather than the model structure, and it does not address the inherent defects of LSTM in feature fusion.

[0004] Therefore, there is an urgent need for a power battery temperature prediction method that can efficiently integrate global time series information, has a simple and efficient model structure, and combines high prediction accuracy, strong generalization ability, and good practicality, in order to overcome one or more of the defects in the existing technologies mentioned above. Summary of the Invention

[0005] This invention aims to solve the problems of the prior art. It provides a method and system for predicting the temperature of power batteries with high prediction accuracy, strong generalization ability, and moderate model complexity. The technical solution of this invention is as follows:

[0006] A method for predicting the temperature of a power battery based on fusion average pooling using LSTM-BP includes the following steps:

[0007] Step 1: Obtain timing data of the power battery during operation;

[0008] Step 2: Perform data preprocessing on the time-series data, including data cleaning, feature filtering, and standardization preprocessing.

[0009] Step 3: Divide the preprocessed time series data into fixed-length input sequence samples using a sliding window;

[0010] Step 4: Construct a Long Short-Term Memory-Backpropagation Neural Network (LSTM-BP) model that integrates average pooling; the LSTM-BP model includes a Long Short-Term Memory network layer, an average pooling layer, and a shallow backpropagation neural network layer connected in sequence;

[0011] Step 5: Train the LSTM-BP model using the input sequence samples;

[0012] Step 6: Input the time series data to be predicted into the trained LSTM-BP model, and output the predicted value of the power battery temperature at future time.

[0013] Furthermore, the Long Short-Term Memory (LSTM) network layer has a two-layer structure, which is used to extract temporal features from the input sequence and output the hidden state at each time step.

[0014] Furthermore, the average pooling layer is used to perform a global arithmetic average of the hidden states of all time steps output by the long short-term memory network layer. That is, the element-wise summation of each hidden state vector in the time dimension is divided by the total number of time steps to generate a feature vector that integrates full-time information.

[0015] Furthermore, the shallow backpropagation neural network layer includes a first fully connected layer, a random deactivation layer, and a second fully connected layer connected in sequence; the first fully connected layer uses a modified linear unit activation function and is used to implement nonlinear feature transformation; the random deactivation layer is used to randomly discard some neuron outputs to prevent overfitting; and the second fully connected layer is used to map features to the output dimension to obtain normalized prediction values.

[0016] Furthermore, in step 5, the model training uses the Adam optimizer and introduces weight decay and gradient clipping techniques. Weight decay refers to adding an L2 regularization term to the loss function to suppress model complexity, while gradient clipping refers to limiting the maximum norm of the gradient vector during backpropagation to prevent gradient explosion.

[0017] Furthermore, in step 5, an early stopping strategy is adopted to control the training process: when the validation set loss does not decrease within 20 consecutive training cycles, training is stopped and the optimal model parameters are restored.

[0018] Furthermore, the preprocessing in step 2 includes feature screening, which employs a combination of mutual information evaluation and random forest feature importance evaluation. The screened features include at least battery current, voltage, and vehicle speed.

[0019] Furthermore, the preprocessing in step 2 also includes data standardization, using the Z-score standardization method.

[0020] A power battery temperature prediction system employing any one of the methods described herein, comprising:

[0021] The data acquisition and preprocessing module is used to acquire and preprocess the runtime sequence data of the power battery.

[0022] The sample construction module is used to construct input sequence samples from preprocessed data using a sliding window.

[0023] The model building module is used to build the long short-term memory-backpropagation neural network model that incorporates average pooling.

[0024] A model training module is used to train the model using samples;

[0025] The temperature prediction module is used to predict temperatures using a trained model.

[0026] The advantages and beneficial effects of this invention are as follows:

[0027] 1. This invention achieves high prediction accuracy: It effectively captures the long-term temporal dependence of battery temperature evolution through a dual-layer LSTM network, then globally fuses information from all time steps using an average pooling mechanism, and finally performs precise nonlinear mapping through a shallow BP network, significantly improving prediction accuracy. Results show that its coefficient of determination R² can reach over 0.996.

[0028] 2. Strong generalization ability: The average pooling mechanism equally weights the information of all time steps in the sequence, avoiding the attention mechanism's potential over-focus on specific noise or abnormal time steps, thus enhancing the model's stability and generalization ability on different vehicle datasets. The cross-dataset test R² remains above 0.98.

[0029] 3. Balance between model complexity and performance: Average pooling is used instead of a complex attention mechanism, requiring no additional trainable parameters. A shallow backpropagation (BP) structure is employed instead of a deep network, effectively avoiding overfitting and vanishing gradient problems. The model maintains high performance while exhibiting low computational complexity and good deployment potential.

[0030] 4. The innovation of this invention lies in the following: For the task of battery temperature time series prediction, a simple LSTM-BP fusion model is proposed, which does not require a complex attention mechanism or residual connection. Global time series information fusion is achieved through average pooling, which avoids the limitation of traditional LSTM relying only on the output of the last time step. At the same time, it overcomes the problem of overfitting caused by the attention mechanism. While ensuring high accuracy, it significantly improves the generalization ability and practicality of the model. Attached Figure Description

[0031] Figure 1 This is a general framework diagram of the method according to a preferred embodiment of the present invention;

[0032] Figure 2 This is a scatter plot comparing the prediction results of different models (BP, LSTM, LSTM-BP) in the embodiments of the present invention.

[0033] Figure 3 This is a comparison chart of the prediction results of the LSTM-BP pooling model on multiple test sets in this embodiment of the invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0035] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0036] This embodiment provides a method for predicting the temperature of a power battery based on a long short-term memory-backpropagation neural network with fused average pooling.

[0037] 1. Data Preparation

[0038] The data comes from a real-vehicle operating dataset of ternary lithium-ion batteries. The original data sampling period is 10 seconds. First, data cleaning is performed to remove data points with a vehicle speed of 0 and a total current of less than 2A. Data points with a time interval greater than 60 seconds are considered different driving segments, and only segments longer than 10 minutes are retained. Valid data records are obtained after cleaning.

[0039] For the cleaned data, the mutual information (MI) and random forest (RF) feature importance assessment methods were used to screen out the features most relevant to battery temperature prediction. Through correlation analysis, the input features were finally determined to be the current current I(t), the minimum voltage V_min, and the vehicle speed Speed.

[0040] The selected feature data is processed using the Z-score normalization method, and the calculation formula is as follows:

[0041]

[0042] 2. Sample Construction

[0043] To construct training samples with temporal dependencies, a sliding window mechanism is employed. The time step is set to T=60, meaning each sample contains feature data from 60 consecutive time steps. The input of the i-th sample is... The corresponding output label is the temperature value at time T+1. A large number of sequence samples are generated using a sliding window. The overall data is then divided into multiple parts in chronological order, and further divided into training, validation, and test sets in a 6:2:2 ratio.

[0044] 3. Model Building

[0045] like Figure 1 As shown, the core model of this invention is constructed.

[0046] (1) LSTM layer: A two-layer LSTM structure is adopted. The first layer LSTM receives the input sequence. The calculation process is as follows:

[0047] Use the Forgotten Gate The formula for calculating the degree to which the cell state is preserved from the previous moment is:

[0048]

[0049] in This is the Sigmoid activation function.

[0050] Use input gates The formula for determining the update level of the current input information is:

[0051]

[0052] Where W and b are the weight matrix and bias vector, respectively. It is in a hidden state.

[0053] use The candidate cell state is determined by the following formula:

[0054]

[0055] Where tanh is the hyperbolic tangent activation function.

[0056] use The cell renewal state is determined by the following formula:

[0057]

[0058] The output gate controls the amount of information output from the cell state to the hidden layer, and the calculation formula is as follows;

[0059] ,

[0060] The first LSTM layer outputs the hidden state sequence at each time step, which serves as the input to the second LSTM layer. The second LSTM layer further processes the hidden state and finally outputs the hidden state at each time step. .

[0061] (2) Average pooling layer: receives the hidden states of all time steps from the LSTM layer output. This layer contains no trainable parameters; its operation is to perform an element-wise arithmetic average of the T hidden state vectors, calculated as follows:

[0062]

[0063] in This is the feature vector after average pooling.

[0064] (3) Shallow BP neural network layer: receives the feature vector output by the average pooling layer. This part of the structure is designed to be shallow to control complexity, specifically including:

[0065] First fully connected layer: ,in , Here are the weights and biases, and the activation function is ReLU.

[0066] Random deactivation layer: To prevent overfitting, some neurons are randomly discarded with a low probability.

[0067] Second fully connected layer (output layer): Output the normalized predicted value.

[0068] (4) Destandardization: Normalize the predicted values ​​output by the model. The standard deviation of the temperature labels calculated during the training phase was used. and mean Perform inverse standardization to obtain the final temperature prediction value. .

[0069] 4. Model Training and Optimization

[0070] The model is implemented in the PyTorch framework. During training, the mean squared error loss function is used. The optimizer is Adam, the learning rate is set to 0.0005, and the weight decay coefficient is set. To improve training stability, gradient pruning is employed to limit the gradient norm. To prevent overfitting, an early stopping strategy is introduced: if the validation set loss does not decrease within 20 consecutive training epochs, training is stopped, and the model parameters that minimize the validation loss are restored.

[0071] 5. Analysis of Experimental Results

[0072] To verify the model's performance, ablation experiments and generalization tests were conducted.

[0073] Ablation experiments: On the same car dataset, the prediction results were compared with BP, LSTM, and the LSTM-BP pooling model proposed in this invention. For example... Figure 2 As shown, the prediction results of this invention are close to the true values. The model of this invention achieved the best performance on the test set with MSE=0.1049, MAE=0.1537, and R²=0.9962, outperforming other variants.

[0074] Generalization ability test: The trained model was tested on a dataset of 5 additional cars, and the prediction results were as follows. Figure 3 As shown, R² is above 0.98 for all datasets, while MSE and MAE remain at low levels, demonstrating the model's excellent generalization ability.

[0075] The method proposed in this invention can accurately and stably predict the temperature of power batteries and has good cross-scenario adaptability, thus possessing high engineering application value.

[0076] It should be noted that the user information (including but not limited to user device information, personal user information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the laws, regulations and standards of relevant countries and regions.

[0077] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0078] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0079] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for predicting the temperature of a power battery based on fusion average pooling using LSTM-BP, characterized in that, Includes the following steps: Step 1: Obtain timing data of the power battery during operation; Step 2: Perform data preprocessing on the time-series data, including data cleaning, feature filtering, and standardization preprocessing. Step 3: Divide the preprocessed time series data into fixed-length input sequence samples using a sliding window; Step 4: Construct a Long Short-Term Memory-Backpropagation Neural Network (LSTM-BP) model that integrates average pooling; the LSTM-BP model includes a Long Short-Term Memory network layer, an average pooling layer, and a shallow backpropagation neural network layer connected in sequence; Step 5: Train the LSTM-BP model using the input sequence samples; Step 6: Input the time series data to be predicted into the trained LSTM-BP model, and output the predicted value of the power battery temperature at future time.

2. The method according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network layer is a two-layer structure used to extract temporal features from the input sequence and output the hidden state at each time step.

3. The method according to claim 2, characterized in that, The average pooling layer is used to perform a global arithmetic average of the hidden states of all time steps output by the long short-term memory network layer. That is, the element-wise summation of each hidden state vector in the time dimension is divided by the total number of time steps to generate a feature vector that integrates full temporal information.

4. The method according to claim 3, characterized in that, The shallow backpropagation neural network layer includes a first fully connected layer, a random deactivation layer, and a second fully connected layer connected in sequence. The first fully connected layer uses a modified linear unit activation function to implement nonlinear feature transformation. The random deactivation layer is used to randomly discard some neuron outputs to prevent overfitting. The second fully connected layer is used to map features to the output dimension to obtain normalized prediction values.

5. The method according to claim 1, characterized in that, In step 5, the model is trained using the Adam optimizer, and weight decay and gradient clipping techniques are introduced. Weight decay refers to adding an L2 regularization term to the loss function to suppress model complexity, and gradient clipping refers to limiting the maximum norm of the gradient vector during backpropagation to prevent gradient explosion.

6. The method according to claim 1, characterized in that, In step 5, an early stopping strategy is used to control the training process: when the validation set loss does not decrease within 20 consecutive training cycles, training is stopped and the optimal model parameters are restored.

7. The method according to claim 1, characterized in that, The preprocessing in step 2 includes feature selection, which uses a combination of mutual information evaluation and random forest feature importance evaluation. The selected features include at least battery current, voltage, and vehicle speed.

8. The method according to claim 7, characterized in that, The preprocessing in step 2 also includes data standardization, using the Z-score standardization method.

9. A power battery temperature prediction system employing the method described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess the runtime sequence data of the power battery. The sample construction module is used to construct input sequence samples from preprocessed data using a sliding window. The model building module is used to build the long short-term memory-backpropagation neural network model that incorporates average pooling. A model training module is used to train the model using samples; The temperature prediction module is used to predict temperatures using a trained model.

Images