A Time Series Prediction Method Based On Fitting Error Compensation

In the field of time series prediction, it has been widely used in several aspects aiming at our real life. With the development of information technology, the demand of higher prediction accuracy is required than ever before.Since the prediction accuracy is highly related to the dynamic characteristics of nonlinear time series, in this study, a novel prediction model based on error compensation is proposed to improve the accuracy for nonlinear time series prediction, where an echo state network (ESN) is firstly modeled to describe the complexity of dynamic system, and then an error compensation model is constructed based on least square support vector machine (LSSVM). In the modeling of LSSVM, a genetic algorithm is designed to select the training samples of LSSVM model so as to reduce the negative influence by the noise.To verify the effectiveness of the proposed model, a benchmark task and a class of industrial problem of byproduct gas flow prediction are employed. The preliminary prediction models are constructed by BP based neural network, ESN and classical SVM, respectively, and the error compensation model is established by LSSVM. From the results, we discover that the performances of preliminary prediction models are all improved by the error compensation model. And the ESN model based fitting error compensation show the best performance among all methods, and the prediction accuracy has improved compared to the prediction without error compensation. From this point, the error compensation model proposed in this paper is a feasible to improve the accuracy of time series prediction.