Research on the State of Charge Estimation of Electric Forklift Batteries Based on an Improved Transformer Model

The state of charge (SoC) is one of the critical parameters in battery management systems, as it directly determines the operational safety and reliability of batteries. To accurately predict the SoC of an electric forklift under varying operating conditions, two surrogate models, an improved Transformer and an improved Transformer 2, are developed. The experimental data obtained through real-vehicle tests are multi-dimensional and contain multiple sources of noise, resulting in poor prediction accuracy when only a single preprocessing algorithm is used. Therefore, this paper first discusses the effect of the preprocessing algorithms on SoC estimation. Compared with the original experimental data and the Kalman filter algorithm, the Kalman filter–principal component analysis (PCA) method is more suitable for preprocessing the original electric forklift data. The mean absolute error (MAE) and root mean square error (RMSE) of the improved Transformer model obtained using the Kalman filter – PCA method are reduced by 26.32% and 27.73% respectively, compared to the single Kalman method. Then, this study investigates the impact of data with different dimensions on the prediction performance of the improved Transformer mode. The results show that five-dimensional data can more effectively train the improved Transformer model, since the MAE decreases by 14.63% and 19.54%, and the RMSE decreases by 14.85% and 20.37% compared to three-dimensional and seven-dimensional data. Through the analysis of the improved Transformer model, an improved Transformer 2 model with higher prediction accuracy is obtained. Then, the improved Transformer 2 model is compared with the LSTM and CNN algorithms. The results indicate that the improved Transformer 2 model can predict SoC more stably and accurately than the single LSTM and CNN algorithms. Specifically, compared with the LSTM model, the proposed Transformer 2 model reduces the MAE by 77.16% and the RMSE by 91.75%. In comparison with the CNN model, the MAE is reduced by 71.81% and the RMSE by 80%.