LSTM-SVM-Weibull modeling for decommissioning amount prediction of power batteries based on attention mechanism and ISPBO algorithm

Abstract

Taking Shanghai as the research area, fully considering the randomness and timeliness of power battery recycling, a combined prediction model of Long Short-term Memory network—Support Vector Machine—Weibull (LSTM-SVM-Weibull) that integrates attention mechanism and hyperparameters optimized by improved student psychology based optimization (ISPBO) algorithm is proposed to predict the retired amount of power batteries, to further improve the prediction accuracy. In the first stage, the grey relational analysis (GRA) method is used to screen out the strong related influence factors of power battery installed amount. In the second stage, a two-stage predictive model of power battery installed amount based on LSTM network with attention mechanism (Attention-LSTM) and SVM optimized by ISPBO algorithm (ISPBO-SVM) is constructed. Firstly, the attention mechanism is fused in the hidden layer of the LSTM network to accurately predict the various indicators selected by GRA, reduce the impact of indicator value errors on the target value prediction, and highlight the contribution degree of the input sequence at different time prediction points; Then, the Lévy flight strategy is introduced and combined with the Metropolis criterion of the simulated annealing algorithm to accept inferior solutions, the ISPBO algorithm is designed to optimize the hyperparameters of SVM model, so as to predict the target value (power battery installed amount) on the basis of future indicator values by training the ISPBO-SVM target prediction layer. By collecting the historical data of power battery installed amount in Shanghai for simulation experiments, the comparison results show that the designed Attention-LSTM-ISPBO-SVM two-stage power battery installed amount prediction model is significantly better than other comparison models in terms of error and accuracy on different data sets, and it has high accuracy and generalization ability for the prediction of power battery retirement amount. In the third stage, based on the predictive model of power battery installed amount, the Weibull life distribution model is combined to predict the retired amount of power batteries in Shanghai. Through the goodness-of-fit test and evaluating the retirement amount prediction results of different models, it is proved that the predictive model for power battery retired amount based on Weibull life distribution has high stability and practical application value, which effectively reflects the overall change trend of the future power battery retirement amount in Shanghai, and can provide data reference for improving the recovery rate of waste batteries.