Research and application of local perceptron neural network in highway rectifier for time series forecasting

Time series forecasting is one of the most effective ways to settle the elusive problems of increased penetration of energy industry and financial field. In recent years, neural networks are utilized in time series forecasting owing to the rationality and practicability. However, neural networks always overrate the performance of training data, which results in underestimating the test error. In this work, four important training tactics are proposed for the training and modeling of the networks, and the proposed model has a better forecasting result and a better extrapolation performance. The numerical simulation shows that the proposed methods have broader application in time series forecasting, it is not only effective for over-fitting problems but also has promoted the model accuracy considerably. (C) 2017 Elsevier B.V. All rights reserved.