Abstract:

In this study, a state of charge (SOC) estimation method is proposed based on a combination of data preprocessing and algorithm optimization in order to improve the efficiency and accuracy of data-driven method in predicting SOC. The open data was selected as the training set, and the random forest (RF) algorithm was used to determine the influence degree of each feature of the training set on SOC. For this, optimal training samples were selected, and the rationality of the optimal samples was verified. Limit learning machine (ELM) was used to predict SOC, aiming at the instability of random weights and thresholds generated by ELM in the prediction process. This leads to the unsatisfactory accuracy of SOC estimation. Sparrow search algorithm (SSA) was selected for optimize parameters and improve estimation accuracy. Subsequently, the effectiveness of SSA parameter optimization was verified using the BJDST simulation test. Under the constant current discharge and DST working conditions, the improved extreme learning machine (SSA-ELM), ELM, support vector machine, and back-propagation neural network were used to predict the SOC. The results show that the SSA-ELM algorithm has the best prediction effect and a prediction error \ within 1.5%, thus achieving high-precision SOC prediction.

Key words: random forest, sparrow search algorithm, extreme learning machine, feature optimization, state of charge