Lithium ion battery temperature prediction based on hybrid model ConvLSTM-NARX

Abstract

The invention provides a lithium ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX, the model is based on a classic convolutional neural network and a long-short-term memory neural network, and a delay mechanism and a feedback cycle mechanism of a nonlinear autoregression network with exogenous input are combined, so that the learning ability of the model for historical data is enhanced, and the prediction accuracy of the temperature of a lithium ion battery is improved. The overall performance of the model is improved. In order to verify the accuracy and generalization of the model, two experiments are performed on the model by using a common data set, and the model is compared with a reference model LSTM and a reference model CNN-LSTM. Experiments show that the error of the hybrid model in a static long-term temperature data set is 0.1 DEG C, and the model fitting degree reaches 94.57%; the error in a temperature data set of dynamic driving is less than 0.13, and the model fitting degree is more than 98%. In conclusion, the hybrid model provided by the invention has good accuracy and generalization.

Description

Technical Field

[0001] This paper proposes a lithium-ion battery temperature prediction method based on a hybrid model, ConvLSTM-NARX. This model is based on the classic convolutional neural network (CNN) and long short-term memory (LSTM) neural network, and combines the delay mechanism and feedback loop mechanism of a nonlinear autoregressive network with external input (NARX) to improve the overall performance of the model. Background Art

[0002] For the past few decades, research on battery temperature has largely relied on numerical models. For example, thermoelectric models can determine a battery's heat generation and precise internal temperature distribution. However, physical models have limitations in their application. Given the diverse battery types, operating environments, and load conditions, simulation methods must calculate each case, resulting in a significant workload. Furthermore, the creation of physical models takes into account numerous factors and requires the calculation of numerous formulas, which increases the complexity of the prediction process and limits its practicality.

[0003] Secondly, the battery temperature will fluctuate dramatically during operation, which increases the difficulty of prediction. In addition, different load conditions and environmental conditions will cause the battery temperature to change, so battery temperature prediction has always been a technical difficulty.

[0004] In recent years, with the rise of machine learning, it has achieved widespread success in various fields. Artificial neural networks (ANNs), as a crucial machine learning tool, have demonstrated unique advantages in solving highly nonlinear problems and have been widely applied by researchers in many areas related to lithium-ion batteries, such as state of charge estimation, remaining useful life prediction, and battery temperature prediction. This approach can capture the complex nonlinear relationships between input and output parameters through relatively simple structures and algorithms, resulting in excellent performance.

[0005] Currently, in the field of lithium-ion battery temperature prediction, Shan Zhu et al. proposed using a long short-term memory (LSTM) model to predict battery temperature changes, and showed that temperature changes during battery operation can be regarded as time series data, making LSTM suitable for battery temperature prediction; Dong Wang et al. proposed a hybrid ConvLSTM model to predict the capacity and temperature of lithium-ion batteries and achieved good results; Jan Kleiner et al. compared a nonlinear autoregressive network (NARX) with exogenous input with a conventional feedforward network, proving that NARX has better accuracy but is more suitable for simple systems.

[0006] Question 1: During battery operation, the temperature fluctuates dramatically, making it difficult to accurately capture its changing trends. As time goes by, the predictive performance of most models will decline.

[0007] Problem 2: Many machine learning models today rely heavily on the data used for training, so generalization is often seen as a problem. Summary of the Invention

[0008] Neural networks are of great significance and development potential in the field of battery temperature prediction. This technology can effectively learn the complex nonlinear relationship between input and output data and complete battery temperature prediction, but it has problems with accuracy and generalization. To solve the above problems, the present invention proposes a lithium-ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX. This model is based on the classic CNN and LSTM models and combines the delay mechanism and feedback loop mechanism of NARX, including: S1: Data preprocessing, unified denoising and normalization of the experimental dataset; S2: Reconstruct the dataset using the delay mechanism and feedback loop mechanism in NARX; S3: Input the reconstructed data into the CNN and LSTM models to learn the historical feature data and predict the next temperature data; S4: Finally, the model is trained and the temperature prediction results are obtained by iteratively calculating each time step.

[0009] Furthermore, the pre-processing operation mentioned in step S1 includes: S1-1: Using multivariate input, extract temperature, voltage, and current to construct a multivariate dataset, where temperature is the target variable and voltage and current are exogenous variables; S1-2: Use wavelet threshold denoising to denoise the dataset. By decomposing the data into different frequency spaces, we obtain a detailed part containing high-frequency signals and an approximate part containing low-frequency signals. This approximate value can be further decomposed to obtain more details and approximate values. The basis function sym8 is used to retain more detailed features of the temperature data. S1-3: In order to better train the data, all features are normalized to the [0, 1] range using the MinMaxScaler method.

[0010] Furthermore, the specific contents of step S2 include: S2-1: Reconstruct the preprocessed data. First, initialize the parameters and define the historical window size as look_back=20. That is, use the data of the previous 20 time steps to predict the data of the next time step, and continuously update the window by sliding. The prediction step size pre_step=1 indicates a single-step prediction. feedback_delay=1 indicates feedback delay, that is, use the temperature of the time step before the current moment as the feedback feature. Initialize empty lists x and y to store features and labels, respectively. S2-2: Loop through the dataset, starting from "look_back + feedback_delay" and ending at "len(data) - pred_step". At each time point i, we extract the temperature, current, and voltage data from "i - look_back" to "i". For the feedback feature, we extract the temperature data from "i - look_back - feedback_delay" to "i - feedback_delay". These four components are combined into a matrix of shape (look_back, 4), where each row corresponds to a time step and the columns are: temperature (input), current, voltage, temperature (feedback). S2-3: Add the merged matrix to x for subsequent input into CNN and LSTM for learning and prediction. The reconstructed data provides rich temporal features and contextual information for subsequent steps, enabling the model to more accurately predict battery temperature changes.

[0011] Furthermore, the specific contents of step S3 include: S3-1: The CNN module contains a one-dimensional convolutional layer and a flatten layer. Each CNN layer is wrapped in a time-distributed layer. The time-distributed layer enables the CNN to better process time series data. It performs convolution processing on each time step in the data. After the flatten layer, the output data has a reduced dimensionality, making the data shape more suitable for the LSTM layer. S3-2: The LSTM module consists of two linked LSTM layers and two Dense layers. The first LSTM layer has 128 units and returns the output of each time step, preserving the time series structure ( return_sequences=True ). The second LSTM layer has 128 units and returns only the output of the last time step ( return_sequences=False ). The data then enters the fully connected layer, reducing the dimensionality from 128 to 1. Finally, a predicted value is output, which is the temperature forecast for the next time step.

[0012] Furthermore, the specific contents of step S4 include: S4-1: Use early stopping and dynamic learning rate adjustment to improve model convergence speed and training efficiency. In the early stopping mechanism, set patience = 10. This means that if the validation loss does not improve after 10 consecutive epochs, training is stopped and the optimal weights are restored to prevent model overfitting. In the dynamic learning rate adjustment, set factor = 0.5 and patience = 5. This means that if the validation loss does not improve after 5 consecutive epochs, the learning rate is multiplied by 0.5 (halved) to prevent the model from falling into a local optimum. S4-2: Update the temperature, voltage, current, and feedback temperature through the sliding window, and repeat the above steps to input the hybrid model to iteratively predict the temperature value of each time step of the data set.

[0013] Furthermore, the specific contents of step S3-1 include: In the CNN module, convolution is calculated using the following formula:

[0014] in, represents the feature j of the k+1th layer, represents the number of features in the kth layer, Represents the convolution kernel on the feature map f from the kth layer to the (k + 1)th layer. In this method, the size of the convolution kernel is 2×4, and the input data shape is (20, 4). Through the convolution operation, the local temporal pattern of these 20 time steps, that is, the relationship between two adjacent time steps, can be extracted, enhancing the feature representation ability and reducing the computational complexity of the subsequent LSTM.

[0015] Furthermore, the specific contents of step S3-2 include: The specific calculation formula for the LSTM module to process time series data is as follows:

[0016] in, represents the input data at the current time t, represents the output of the previous time step, , and They are the calculation formulas for the input gate, forget gate, and output gate, respectively. Represents the long-term state of the unit. Its calculation method is to calculate the long-term state of the previous time step through the forget gate and add the current unit state through the input gate. This can well learn the historical data. Finally, a vector is generated through the activation function and the output value is output through the output gate. Send to the next unit. In this method, the LSTM module uses two LSTM layers to link modules, which not only retains the time series structure but also extracts time features and learns long-term time patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solution of this specification, the following is a brief introduction to the drawings required for the technical solution.

[0018] Figure 1 This is the general architecture of the hybrid model ConvLSTM-NARX proposed in this paper. The model includes key components such as 1D-CNN, Flatten, LSTM, and Dense. The blue arrows represent the feedback loop mechanism in the NARX model, and the tapped delay line (TDL) represents the delay mechanism in the NARX model. By using TDL, the outputs of any number of previous time steps can be stored and fed back to the input layer.

[0019] Figure 2 The visualization results of the LSTM model and the ConvLSTM-NARX proposed in this paper were tested on a temperature dataset with long-term static cycling. The upper figure shows the visualization of the LSTM model, while the lower figure shows the visualization of the proposed model. Orange represents the true value curve, and green represents the predicted value curve. Through this comparison, we can clearly see that compared to the LSTM, the hybrid model proposed in this paper can better capture the complex temperature fluctuations during battery operation, and its accuracy does not decrease over time.

[0020] Figure 3 This is the visualization result of testing ConvLSTM-NARX of the present invention on data at 10°C in the Panasonic dataset. The data of different driving conditions at this temperature are tested. The blue curve is the true value curve and the orange curve is the predicted value curve. It can be seen that ConvLSTM-NARX can accurately capture temperature changes.

[0021] Figure 4This is the visualization result of testing the ConvLSTM-NARX of the present invention on the data at 25°C in the Panasonic dataset. By testing data of different driving conditions at this temperature, we can also observe the accuracy of the model's prediction.

[0022] Figure 5 This figure visualizes the RMSE and R-squared results of the experiments on the Panasonic dataset. We can see that ConvLSTM-NARX performs well under different ambient temperatures and driving conditions, with RMSE not exceeding 0.13 and R-squared above 98%. DETAILED DESCRIPTION

[0023] The present invention proposes a lithium-ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX, and conducts long-term temperature prediction experiments and dynamic driving temperature prediction experiments. The hybrid model is used to train and test data to verify its accuracy and generalization.

[0024] The first step is data preprocessing to better adapt it to the time series prediction model. In the long-term temperature prediction experiment, the temperature of each cycle is averaged, allowing the temperature data to be approximately treated as time series data. In the dynamic driving temperature prediction experiment, the non-uniform data is resampled by 30 seconds to uniform the data frequency. The dataset is then denoised, normalized, and partitioned.

[0025] In the second step, the dataset is reconstructed using the delay mechanism and feedback loop mechanism of NARX.

[0026] In the third step, the processed data enters the CNN module, which includes a one-dimensional convolutional layer and an expansion layer. The one-dimensional convolutional layer extracts patterns or features from the data through convolution operations and then expands the data for input to the next module. To better learn from historical data and make the next prediction, the data is input to the LSTM module. This module uses two LSTM layers and two fully connected layers, and finally outputs a single value: the predicted temperature value for the next time step.

[0027] The fourth step is to use loop iteration to train and test each sample data in the dataset, compare the predicted value with the true value, calculate the evaluation criteria RMSE and R square, and evaluate the performance of the model.

[0028] In summary, this paper proposes a lithium-ion battery temperature prediction method based on a hybrid model, ConvLSTM-NARX, and conducts an in-depth comparative analysis. The innovation of this hybrid model lies in its combination of the powerful large-scale feature extraction capabilities of CNN with the feedback loop and delay mechanism of NARX, enabling better processing and learning of time series data.

[0029] Furthermore, the method proposed in the present invention was tested on public datasets. The experimental results showed that the hybrid model ConvLSTM-NARX performed excellent performance on multiple public datasets and had generalization properties. Compared with the traditional LSTM model, this method can better capture temperature change trends and has higher accuracy.

[0030] Target Datasets: 1. A public dataset, which features lithium phosphate (LFP) / graphite cells cyclically charged and discharged in the horizontal cylindrical fixture of a 30-channel Arbin LBT potentiostat set at 48°C in a forced convection temperature chamber. These cells have a nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V. All cells are charged for a fixed 10-minute charge time, followed by a 20-second rest period, and cycling is terminated when the cells reach 80% of their nominal capacity. 2. A Panasonic 18650PF lithium-ion battery dataset, which contains data measured at various temperatures. Nine drive cycles are performed at each temperature: cycles 1-4, US06, HWFET, UDDS, LA92, and Neural Network (NN). Cycles 1-4 consist of random combinations of US06, HWFET, UDDS, and LA92 drive cycles. The NN cycle is a combination of US06 and LA92 drive cycles, with some additional dynamics to test the generalization of the neural network.

[0031] In summary, this paper proposes a lithium-ion battery temperature prediction method based on a hybrid model, ConvLSTM-NARX, and conducts an in-depth comparative analysis of this method. The innovation of this hybrid model lies in its combination of the powerful feature extraction capabilities of CNN with the feedback loop and delay mechanism of NARX, enabling better processing and learning of time series data. Experimental results show that the hybrid model, ConvLSTM-NARX, demonstrates excellent performance on multiple public datasets and significantly improves accuracy compared to traditional LSTM models. In a dataset obtained under long-term static battery operation conditions, the model better captures battery temperature trends than the LSTM model, and its performance does not show significant degradation over time. In a dataset of battery temperature under dynamic driving cycle conditions, the model demonstrates good performance across datasets with varying operating conditions and ambient temperatures, with R-squared values ​​exceeding 98%. These experimental results demonstrate that the combination of LSTM with CNN and NARX can achieve higher accuracy, providing an accurate and generalizable solution for battery temperature prediction. In addition, we also analyzed the decision-making process of the model through visual interpretation results, and the results showed that ConvLSTM-NARX can well capture the complex temperature fluctuations during battery operation under different conditions.

Claims

1. The present invention proposes a lithium-ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX, characterized in that: This hybrid model uses the LSTM network as its core, uses CNN for feature extraction, and combines the delay mechanism and loop feedback mechanism of NARX to improve the overall performance of the hybrid model. Specifically, the method includes the following: S1: Data preprocessing, unified denoising and normalization of the experimental dataset; S2: Reconstruct the dataset using the delay mechanism and feedback loop mechanism in NARX; S3: Input the reconstructed data into the CNN and LSTM models to learn the historical feature data and predict the next temperature data; S4: Train the model and obtain the temperature prediction results by iteratively calculating each time step.

2. A lithium-ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX according to claim 1, characterized in that: The specific operations of step S1 include: S1-1: Using multivariate input, extract temperature, voltage, and current to construct a multivariate dataset, where temperature is the target variable and voltage and current are exogenous variables; S1-2: Use wavelet threshold denoising to denoise the dataset. By decomposing the data into different frequency spaces, we obtain a detailed part containing high-frequency signals and an approximate part containing low-frequency signals. This approximate value can be further decomposed to obtain more details and approximate values. The basis function sym8 is used to retain more detailed features of the temperature data. S1-3: In order to better train the data, all features are normalized to the [0, 1] range using the MinMaxScaler method.

3. The lithium-ion battery temperature prediction method based on the hybrid model ConvLSTM-NARX according to claim 1 is characterized in that: The specific operations of step S2 include: S2-1: Reconstruct the preprocessed data and initialize the parameters. Define the historical window size as look_back=20, that is, use the data of the previous 20 time steps to predict the data of the next time step, and continuously update the window by sliding. The prediction step size pre_step=1 indicates a single-step prediction, feedback_delay=1 indicates feedback delay, that is, use the temperature of the time step before the current moment as the feedback feature, and initialize empty lists x and y to store features and labels respectively; S2-2: Loop through the dataset, starting from "look_back + feedback_delay" and ending at "len(data) - pred_step". At each time point i, we extract the temperature, current, and voltage data from "i - look_back" to "i". For the feedback feature, we extract the temperature data from "i - look_back - feedback_delay" to "i - feedback_delay". These four components are combined into a matrix of shape (look_back, 4), where each row corresponds to a time step and the columns are: temperature (input), current, voltage, and temperature (feedback). S2-3: Add the merged matrix to x for subsequent input into CNN and LSTM for learning and prediction. The reconstructed data provides rich temporal features and contextual information for subsequent steps, enabling the model to more accurately predict battery temperature changes.

4. The lithium-ion battery temperature prediction method based on the hybrid model ConvLSTM-NARX according to claim 1 is characterized in that: The specific operations of step S3 include: S3-1: The CNN module contains a one-dimensional convolutional layer and a flatten layer. Each CNN layer is wrapped in a time-distributed layer. The time-distributed layer enables the CNN to better process time series data. It performs convolution processing on each time step in the data. After the flatten layer, the output data has a reduced dimensionality, making the data shape more suitable for the LSTM layer. S3-2: The LSTM module consists of two linked LSTM layers and two Dense layers. The first LSTM layer has 128 units and returns the output of each time step, preserving the time series structure (return_sequences=True). The second LSTM layer has 128 units and returns only the output of the last time step (return_sequences=False). The data then enters the fully connected layer, reducing the dimension from 128 to 1. Finally, a predicted value is output, which is the temperature forecast for the next time step.

5. The lithium-ion battery temperature prediction method based on the hybrid model ConvLSTM-NARX according to claim 1 is characterized in that: The specific operations of step S4 include: S4-1: Use early stopping and dynamic learning rate adjustment to improve model convergence speed and training efficiency. In the early stopping mechanism, set patience = 10. This means that if the validation loss does not improve after 10 consecutive epochs, training is stopped and the optimal weights are restored to prevent model overfitting. In the dynamic learning rate adjustment, set factor = 0.5 and patience = 5. This means that if the validation loss does not improve after 5 consecutive epochs, the learning rate is multiplied by 0.5 (halved) to prevent the model from falling into a local optimum. S4-2: Update the temperature, voltage, current, and feedback temperature through the sliding window, and repeat the above steps to input the hybrid model to iteratively predict the temperature value of each time step of the data set.

6. A lithium-ion battery temperature prediction method based on a hybrid model ConvLSTM-NARX according to claim 4, characterized in that: In the CNN module of step S3-1, the specific calculation formula for extracting features by the convolution kernel is as follows: ； in, , represents the feature j of the k+1th layer and the kth layer, represents the activation function, represents the bias term of feature j at the kth layer, represents the number of features in the kth layer, Represents the convolution kernel on the feature map f from the kth layer to the (k + 1)th layer. In this method, the size of the convolution kernel is 2×4, and the input data shape is (20, 4). Through the convolution operation, the local temporal pattern of these 20 time steps, that is, the relationship between two adjacent time steps, can be extracted, enhancing the feature representation ability and reducing the computational complexity of the subsequent LSTM.

7. The lithium-ion battery temperature prediction method based on the hybrid model ConvLSTM-NARX according to claim 4 is characterized in that: In step S3-2, the specific calculation formula for the LSTM module to process time series data is as follows: ； in, represents the input data at the current time t, represents the output of the previous time step, represents the sgmoid activation function, , and They are the calculation formulas for the input gate, forget gate, and output gate, respectively. 、 、 and It represents the corresponding bias value, 、 Then it represents the corresponding weight matrix, represents the cell state at the current time t, It represents the long-term state of the unit, which is calculated by performing the forget gate calculation on the long-term state of the previous time step and adding the current unit state after the input gate. , so that we can learn the historical data well, and finally Generate a vector through the activation function and output the value through the output gate Send to the next unit. In this method, the LSTM module uses two LSTM layers to link modules, which not only retains the time series structure but also extracts time features and learns long-term time patterns.

Images