Classifier based local artificial neural networks applied to watershed runoff forecasting

摘要

Hydrologists and engineers need effective methods to forecast watershed runoff for many researches and engineering applications. In the present integrated water management context, time efficiency and cost effectiveness are mainly concerned, and at the same time, forecasting techniques must be practical and accurate. This paper describes the application of artificial intelligence techniques to forecast the runoff of a waterbasin in the upper reach of Huai River, east of China. A hybrid approach to simulate the future behaviour of waterbasin runoff is presented which combines the clustering ability of decision tree to feed the instances of data into different branches of an integrated ANN model and the generalising ability of an ANN to establish rainfall-runoff relationship and further forecast the future system behaviour. The main focus of this paper is to train several classifier-based local ANN (LANN) models to deal with different part of a wide range of flow stages, namely dry period, average period and high flow (wet) period respectively. Then when new data coming, a classifier will send data into different LANN. The results obtained are presented and compared with outputs of other global ANN (GANN).