神经网络径流预报模型中基于互信息的预报因子选择方法

Mutual information-based input variable selection method for runoff-forecasting neural network model

神经网络在径流预报中得到了广泛应用并取得了良好效果,其关键问题之一是输入变量(预报因子)选择,但这一问题通常没有受到重视。本研究基于互信息的概念探讨了如何选择径流预报输入变量,并结合三峡工程建成前长江干流宜昌水文站的日径流预报进行了研究。结果表明,基于互信息能够有效地判断待选预报因子(输入变量)与预报变量之间的相互关系,以帮助选择神经网络预报模型的输入变量,从而提高径流预报的精度。

Neural network is one of the popular and effective models used for hydrological forecasting.One of the major problems of this method is the selection of input variables.This study focuses on this problem and uses the mutual information（MI） to select neural network input variables.Application to the daily discharge forecasting at the Yichang hydrological station before the Three Gorges project shows that MI is an effective technique for selecting the input variables and thus for improving the runoff forecasting accuracy.