A Battery Health Prediction Method Based on AGA-Elman Algorithm

Abstract

This invention relates to a battery health prediction method based on the AGA-Elman algorithm, comprising the following steps: Step 1: Data acquisition; Step 2: Data preprocessing, dividing the highly correlated incremental capacity data and battery health data into training and test sets, with 50% of the data set as the training set and 50% as the test set; Step 3: AGA-Elman prediction model training, importing the training data from Step 2 into the AGA-Elman model for training. This application can provide accurate prediction of battery health even with limited data.

Description

Technical Field

[0001] This invention relates to the field of battery health prediction technology, and in particular to a battery health prediction method based on the AGA-Elman algorithm. Background Technology

[0002] 18650 batteries are widely used in portable electronic devices, power tools, vehicles, rechargeable battery packs, LED lighting, security equipment, and many other technological fields due to their long lifespan, high energy density, and lightweight portability. However, as the number of charge-discharge cycles increases, the battery ages, its impedance increases, and it may even cause serious safety problems. Therefore, predicting battery health is crucial. Accurate prediction methods can inform users of the battery's aging status and capacity degradation, allowing for timely battery inspection and replacement.

[0003] In recent years, with the rapid development of big data and cloud computing, data-driven modeling methods for predicting battery health have emerged. These methods utilize machine learning to establish suitable algorithmic models to predict battery health, achieving good predictive results. This approach can eliminate the need for a battery model and complete the prediction using purely data-driven modeling. In contrast, typical backpropagation (BP) neural networks require a large amount of cyclic charge-discharge data as training data to accurately predict battery health, and cannot achieve accurate predictions with limited data. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a battery health prediction method based on the AGA-Elman algorithm, which solves the problem that battery health cannot be accurately predicted when the amount of data is limited.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a battery health prediction method based on the AGA-Elman algorithm, comprising the following steps:

[0006] Step 1: Data Acquisition. The specific steps for data acquisition are as follows:

[0007] S11. Obtain battery test data from multiple cycles of charge and discharge and perform preliminary extraction. Select the data to be extracted, including the battery's nominal capacity data, voltage data, current data, capacity data, and discharge time data for all cycles during the discharge phase.

[0008] S12. Based on the obtained nominal capacity data and capacity data, the battery health data is obtained using the first calculation formula;

[0009] S13. Based on the obtained voltage data, current data, and discharge time data, the second calculation formula is used to obtain the battery incremental capacity data, and the obtained battery incremental capacity data is processed for noise reduction.

[0010] S14. Perform Pearson correlation analysis on the incremental capacity data and battery health data after noise reduction processing, and select the incremental capacity data with high correlation based on the correlation analysis results.

[0011] Step 2: Data preprocessing; Divide the incremental capacity data with high relevance obtained in step S14 and the battery health data obtained in step S12 into training set and test set, with 50% of the data set as training set and 50% of the data set as test set.

[0012] Step 3: AGA-Elman prediction model training. Import the training data from Step 2 into the AGA-Elman model for training. The specific training steps are as follows:

[0013] S31. Determine the network structure; Based on the training data, construct an Elman neural network, including an input layer, hidden layers, an output layer, and a relay layer. The relay layer is used to carry the memory and transmission of time data. Set the initial training times to 1000 times, the target error to 0.000001, the learning rate to 0.01, the number of nodes in the input layer to 168, and the number of nodes in the output layer to 1.

[0014] S32. Initialize the Elman neural network and obtain the initial weights and thresholds of the Elman neural network;

[0015] S33, AGA optimizes Elman weights and thresholds; AGA obtains the initial weights and thresholds of the Elman neural network and encodes them, sets the initial population size and the maximum number of iterations, constructs the corresponding fitness value function, and writes the adaptive crossover and mutation probabilities determined by the individual fitness value and the number of algorithm iterations;

[0016] S34. Perform crossover, mutation, and selection operations;

[0017] S35. Calculate the fitness value and determine whether the optimization goal has been achieved. If the optimization goal has been achieved, determine the optimal weight and threshold. If the optimization goal has not been achieved, proceed to step S34.

[0018] S36. After determining the optimal weights and thresholds, assign the optimal weights and thresholds to the Elman neural network. Train the Elman neural network with training data to obtain an accurate data prediction model and complete the accurate prediction of battery health.

[0019] Optionally, the first calculation formula in step S12 is:

[0020]

[0021] Where Q0 is the nominal capacity of the battery, Q tLet t represent the battery capacity at time t, and SOH represent the battery health.

[0022] Optionally, the second calculation formula in step S12 is:

[0023]

[0024] Where Q is the discharge capacity, I is the current, t is the time, and V is the voltage.

[0025] Optionally, the correlation analysis in step S14 uses a third calculation formula to obtain the Pearson correlation analysis results. The third calculation formula includes:

[0026]

[0027] Where, ρ X,Y X is the Pearson correlation coefficient. i , Corresponding to the characteristics and their average values, Y i , It represents battery health and its average value, ρ X,Y The value of is in the range of [-1, 1]. The closer the value of this coefficient is to 1, the higher the correlation is considered to be.

[0028] Optionally, the number of hidden layer nodes in step S31 is determined according to the fourth calculation formula, which includes:

[0029]

[0030] Where h is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is generally a constant between 1 and 10.

[0031] Optionally, the adaptive crossover probability in S33 is determined according to the fifth calculation formula, which includes:

[0032]

[0033] Where k1 and k2 are constants with a range of [0,1], i is the current iteration number, fitnessgbest is the optimal individual fitness value, fitnessbest is the globally optimal adaptive value, maxgen is the total number of iterations of the algorithm, and P cj Let be the crossover probability of the j-th individual at the i-th time;

[0034] The adaptive mutation probability is determined according to the sixth calculation formula, which includes:

[0035]

[0036] Where k3 is a constant with a value range of [0, 0.95], P mj Let be the mutation probability of the j-th individual in the i-th iteration.

[0037] Beneficial effects: 1. In terms of data selection, this patent uses the battery incremental capacity curve to predict the battery health. The voltage plateau of battery discharge is transformed into a clearly identifiable capacity increment (dQ / dV) peak on the capacity increment curve, which to some extent amplifies the characteristic that the voltage change is not obvious in each cycle of the original battery charge and discharge cycle. In addition, the data used for training adopts the incremental capacity part data that is highly correlated with the battery health change curve, thereby reducing the amount of original data to some extent.

[0038] 2. In terms of algorithm model construction, this patent uses the Elman neural network, which has strong temporal characteristics, for prediction. Compared with the BP neural network, it improves the stability of prediction accuracy and reduces prediction error. The Elman neural network is consistent with the BP neural network in terms of weight and threshold update method, namely gradient descent method, which is prone to getting trapped in local optima. The AGA algorithm is introduced to optimize the initial weights and thresholds of the Elman neural network to avoid getting trapped in local optima. In addition, the Elman neural network can achieve fast and stable operation by utilizing the parallel search characteristics implicit in the AGA algorithm.

[0039] 3. In terms of prediction performance, the AGA-Elman algorithm model can control various errors of the prediction results to within 3% with 50% training set. Attached Figure Description

[0040] Figure 1 A schematic diagram of the process for extracting incremental capacity data for the 18650 battery.

[0041] Figure 2 A schematic diagram of the AGA-Elman algorithm model.

[0042] Figure 3 Schematic diagram of the Elman neural network structure.

[0043] Figure 4 Comparison chart of predicted and actual values ​​from the AGA-Elman algorithm model. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] Combined with appendix Figure 1-3 This invention provides a battery health prediction method based on the AGA-Elman algorithm, comprising the following steps:

[0046] Step 1: Data Acquisition. The specific steps for data acquisition are as follows:

[0047] S11. Obtain battery test data for 168 cycles of charge and discharge of 18650 battery and perform preliminary extraction. Select the extraction to include the battery's nominal capacity data, voltage data, current data, capacity data and discharge time data for all cycles of the discharge phase.

[0048] S12. Based on the obtained nominal capacity data and capacity data, the battery health data is obtained using the first calculation formula, which is:

[0049]

[0050] Where Q0 is the nominal capacity of the battery, Q t Let St be the battery capacity at time t, and SOH be the battery health status.

[0051] S13. Based on the obtained voltage data, current data, and discharge time data, the battery incremental capacity data is obtained using the second calculation formula, and the obtained battery incremental capacity data is then processed to reduce noise. The second calculation formula is as follows:

[0052]

[0053] Where Q is the discharge capacity, I is the current, t is the time, and V is the voltage;

[0054] S14. Perform Pearson correlation analysis on the incremental capacity data and battery health data after noise reduction processing. The correlation analysis results are obtained using the third calculation formula, which includes:

[0055]

[0056] Where, ρ X,Y X is the Pearson correlation coefficient. i , Corresponding to the characteristics and their average values, Y i , It represents battery health and its average value, ρ X,Y The value range of is [-1, 1]. When the value of this coefficient is closer to 1, it is considered to be more correlated. Based on the correlation analysis results, the incremental capacity data with higher correlation are selected.

[0057] Step 2: Data preprocessing; Divide the incremental capacity data with high relevance obtained in step S14 and the battery health data obtained in step S12 into training set and test set, with 50% of the data set as training set and 50% of the data set as test set.

[0058] Step 3: AGA-Elman prediction model training. Import the training data from Step 2 into the AGA-Elman model for training. The specific training steps are as follows:

[0059] S31. Determine the network structure; based on the training data, construct an Elman neural network, including an input layer, hidden layers, an output layer, and a relay layer. The relay layer is used to store and transmit time data. Set the initial training iterations to 1000, the target error to 0.000001, the learning rate to 0.01, the number of nodes in the input layer to 168, the number of nodes in the output layer to 1, and the number of nodes in the hidden layer to be determined according to the fourth calculation formula. The fourth calculation formula includes:

[0060]

[0061] Where h is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is generally a constant between 1 and 10;

[0062] S32. Initialize the Elman neural network and obtain the initial weights and thresholds of the Elman neural network;

[0063] S33. AGA optimizes Elman weights and thresholds; AGA obtains and encodes the initial weights and thresholds of the Elman neural network, sets the initial population size to 100, the maximum number of iterations to 100, constructs the corresponding fitness value function, and writes the adaptive crossover and mutation probabilities determined by the individual fitness value and the number of algorithm iterations. The adaptive crossover probability is determined according to the fifth calculation formula, which includes:

[0064]

[0065] Where k1 and k2 are constants with a range of [0,1], i is the current iteration number, fitnessgbest is the optimal individual fitness value, fitnessbest is the globally optimal adaptive value, maxgen is the total number of iterations of the algorithm, and P cj Let be the crossover probability of the j-th individual at the i-th time;

[0066] The adaptive mutation probability is determined according to the sixth calculation formula, which includes:

[0067]

[0068] Where k3 is a constant with a value range of [0, 0.95], P mj Let be the mutation probability of the j-th individual in the i-th iteration;

[0069] S34. Perform crossover, mutation, and selection operations;

[0070] S35. Calculate the fitness value and determine whether the optimization goal has been achieved. If the optimization goal has been achieved, determine the optimal weight and threshold. If the optimization goal has not been achieved, proceed to step S34.

[0071] S36. After determining the optimal weights and thresholds, assign the optimal weights and thresholds to the Elman neural network. Train the Elman neural network with training data to obtain an accurate data prediction model and complete the accurate prediction of battery health.

[0072] Finally, the final simulation prediction results can be compared with the actual battery health change curve to evaluate the prediction model. Figure 4 The image shows a comparison between the predicted and actual values ​​of the AGA-Elman algorithm model. As can be seen from the image, the prediction model has a smaller error and better prediction performance.

[0073] Calculate the mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and absolute coefficient (R²) based on the output prediction results. 2 The calculation results of the AGA-Elman algorithm model and the comparison results with those using the BP algorithm model, Elman algorithm model, and GA-Elman algorithm model are shown in Table 1:

[0074] Table 1 Prediction Error Results of the Algorithm Model

[0075]

[0076] Table 1 shows that the mean absolute error, mean absolute percentage error, root mean square error, and absolute coefficient of the AGA-Elman algorithm model are 0.020141, 0.028873, 0.028904, and 0.90728, respectively. Compared with the BP model, these indicators have improved by 0.018189, 0.026211, 0.025686, and 0.23802, respectively. Compared with the Elman model, these indicators have improved by 0.014049, 0.019282, 0.018364, and 0.15525, respectively. Compared with the GA-Elman model, these indicators have improved by 0.008434, 0.007366, 0.008983, and 0.06659, respectively. Therefore, the AGA-Elman algorithm model has the smallest errors across all categories compared to other algorithm models, and R0 is the lowest. 2 When the value is maximized, the prediction effect is better and the error is lower.

[0077] Furthermore, since the Elman neural network has an additional layer compared to the BP neural network, it can remember and transmit time information. Therefore, the Elman neural network has better performance in predicting battery health based on incremental battery capacity data with a certain time sequence. While the Genetic Algorithm (GA) optimizes the weights and thresholds of the Elman neural network through global search, enabling prediction with less training data, traditional GA, which uses fixed crossover and mutation probabilities, is prone to local optima and slow convergence. In contrast, this application uses an Adaptive Genetic Algorithm (AGA) that uses adaptive crossover and mutation probabilities determined by individual fitness values ​​and the number of algorithm iterations, enhancing global optimization capabilities and improving convergence speed, thereby further improving prediction accuracy.

[0078] In terms of algorithm model construction, this application uses the Elman neural network, which has strong temporal characteristics, for prediction. Compared with the BP neural network, it improves the stability of prediction accuracy and reduces prediction error. The Elman neural network is consistent with the BP neural network in terms of weight and threshold update method, namely gradient descent method, which is prone to getting trapped in local optima. The AGA algorithm is introduced to optimize the initial weights and thresholds of the Elman neural network to avoid getting trapped in local optima. In addition, the Elman neural network can achieve fast and stable operation by utilizing the parallel search characteristics implicit in the AGA algorithm.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A battery health prediction method based on the AGA-Elman algorithm, characterized in that, Includes the following steps: Step 1: Data Acquisition. The specific steps for data acquisition are as follows: S11. Obtain battery test data from multiple cycles of charge and discharge and perform preliminary extraction. Select the data to be extracted, including the battery's nominal capacity data, voltage data, current data, capacity data, and discharge time data for all cycles during the discharge phase. S12. Based on the obtained nominal capacity data and capacity data, the battery health data is obtained using the first calculation formula; S13. Based on the obtained voltage data, current data, and discharge time data, the second calculation formula is used to obtain the battery incremental capacity data, and the obtained battery incremental capacity data is processed for noise reduction. S14. Perform Pearson correlation analysis on the incremental capacity data and battery health data after noise reduction processing, and select the incremental capacity data with high correlation based on the correlation analysis results. Step 2: Data preprocessing; The incremental capacity data with high relevance obtained in step S14 and the battery health data obtained in step S12 are divided into training set and test set, with 50% of the data set as training set and 50% of the data set as test set. Step 3: AGA-Elman prediction model training. Import the training data from Step 2 into the AGA-Elman model for training. The specific training steps are as follows: S31. Determine the network structure; Based on the training data, construct an Elman neural network, including an input layer, hidden layers, an output layer, and a relay layer. The relay layer is used to carry the memory and transmission of time data. Set the initial training times to 1000 times, the target error to 0.000001, the learning rate to 0.01, the number of nodes in the input layer to 168, and the number of nodes in the output layer to 1. S32. Initialize the Elman neural network and obtain the initial weights and thresholds of the Elman neural network; S33, AGA optimizes Elman weights and thresholds; AGA obtains the initial weights and thresholds of the Elman neural network and encodes them, sets the initial population size and the maximum number of iterations, constructs the corresponding fitness value function, and writes the adaptive crossover and mutation probabilities determined by the individual fitness value and the number of algorithm iterations; The adaptive crossover probability in S33 is determined according to the fifth calculation formula, which includes: wherein k1, k2 are constants and can be selected in the range [0, 1], i is the current iteration number, fitnessgbest is the best individual fitness value, fitnesszbest is the global best fitness value, maxgen is the total number of iterations of the algorithm, P cj is the crossover probability of the jth individual at the ith time; The adaptive mutation probability is determined according to the sixth calculation formula, which includes: Where k3 is a constant with a value range of [0, 0.95], P mj Let be the mutation probability of the j-th individual in the i-th iteration; S34. Perform crossover, mutation, and selection operations; S35. Calculate the fitness value and determine whether the optimization goal has been achieved. If the optimization goal has been achieved, determine the optimal weight and threshold. If the optimization goal has not been achieved, proceed to step S34. S36. After determining the optimal weights and thresholds, assign the optimal weights and thresholds to the Elman neural network. Train the Elman neural network with training data to obtain an accurate data prediction model and complete the accurate prediction of battery health.

2. The battery health prediction method based on the AGA-Elman algorithm according to claim 1, characterized in that: The first calculation formula in step S12 is: in, This refers to the battery's nominal capacity. Let t represent the battery capacity at time t, and SOH represent the battery health.

3. The battery health prediction method based on the AGA-Elman algorithm according to claim 1, characterized in that: The second calculation formula in step S12 is: Where Q is the discharge capacity, I is the current, t is the time, and V is the voltage.

4. The battery health prediction method based on the AGA-Elman algorithm according to claim 1, characterized in that: The correlation analysis in step S14 uses the third calculation formula to obtain the Pearson correlation analysis results. include: in, X is the Pearson correlation coefficient. i , Corresponding to the characteristics and their average values, Y i , It refers to battery health and its average value. The value of is in the range of [-1, 1]. The closer the value of this coefficient is to 1, the higher the correlation is considered to be.

5. The battery health prediction method based on the AGA-Elman algorithm according to claim 1, characterized in that: The number of hidden layer nodes in step S31 is determined according to the fourth calculation formula, which includes: Where h is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is a constant between 1 and 10.

Images