A New Hyperparameters Identification Method of Support Vector Regression Model Based on Back Propagation Neural Network

Support Vector Regression (SVR) is extensively used in various engineering fields. Hyperparameters directly affect the prediction accuracy of SVR model. Traditional hyperparameters identification method simply relies on data of the current sample, which is prone to overfitting with a low prediction accuracy. Therefore, a new hyperparameters identification method is put forward in this paper based on Back Propagation (BP) neural network. Firstly, the optimum hyperparameters of SVR model for each sample in historical database are obtained by a new optimization model proposed. Secondly, the mapping relationship between input data and optimum hyperparameters is built on the basis of BP neural network using multiple historical samples. Furthermore, the optimum hyperparameters for a new sample can be directly acquired from the neural network established. Finally, a case study on state of health estimation for lithium-ion batteries applying SVR is conducted to verify the effectiveness of the presented method.