Technology Sharing

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Tutorial Articles

Model Description

A joint method of Transformer and gated recurrent unit (GRU) based on variational mode decomposition (VMD) and particle swarm optimization (PSO) is proposed. Firstly, the capacity information of lithium battery is decomposed by variational mode decomposition algorithm. In order to avoid the unreasonable decomposition degree affecting the prediction ability, the center frequency method is used to judge the decomposition state as the basis for effective interpretation of the original data information; then the particle swarm optimization algorithm is used to optimize the hyperparameters of the adjusted Transformer Neural Network and Gate Recurrent Unit structure. The Transformer Neural Network uses linear layers instead of decoders to better adapt to time series data, and retains the encoder to capture the global characteristics and internal correlation of the data, which improves the prediction accuracy of a single Transformer and its joint model; finally, the Transformer and GRU predict the main trend subsequence and high frequency subsequence respectively, and the predictions of the two models are fused to complete the estimation of SOH of lithium-ion batteries. The prediction effect of the model was verified using the NASA lithium battery dataset, and compared with single models such as multi-layer perception (MLP), recurrent neural network (RNN) and joint models such as Gaussian function-GRU, Transformer-MLP. The results show that the prediction model in this paper is better than other single models or joint models in terms of accuracy and the degree of fit of the regeneration phenomenon. The mean absolute error and root mean square error of the prediction results are maintained within 0.62% and 1.19%, and the determination coefficient is above 87.08%, which verifies the effectiveness of the proposed research method.

Prediction effect