Wavelet Neural Networks Model Used for Runoff Forecast Based on Fuzzy C-Means Clustering

Abstract

Considering various seasons differs greatly in runoff distribution, a new runoff forecasting method based on fuzzy clustering analysis on forecasting factor set is presented in this paper. Firstly, the historical runoff data are classified as four categories by fuzzy C-means clustering. Then partial forecasting models between the factor set and measured data are respectively established by using wavelet neural network model. A network model categorized recognizer is adopted, which can automatically search a compatible partial forecasting model. Comparison between simple wavelet neural model and integrated forecasting model proposed in this paper is made by illustration. The results demonstrate that the proposed integrated model is of higher forecasting accuracy than the simple one. Through statistical analysis on different regions' runoff over years, it is commonly discovered that various seasons differs greatly in runoff distribution. The traditional numerical methods for runoff forecasting generally applies a simple model to runoff series as a whole. The simple model always takes means as variables and parameters, which is not conducive to simulating the reservoir's great influx differences caused by different conditions of various seasons. Therefore, taking the runoff process as a whole that could be averaged by different phases is obviously inappropriate. In this paper, fuzzy C-means clustering(1), wavelet neural network(2,3,4) and pattern recognition technology are integrately applied in runoff forecasting. Firstly, make use of FCM method to classify the historical runoff data, and then respectively establish partial forecasting models between the forecasting factors set and measured data with wavelet neural network model. By a categorized recognizer of network model to automatically search a compatible partial forecasting model. The established model has been applied to a reservoir's daily runoff forecasting. The result indicates that the forecasting model established by the proposed method is provided with high speed and high accuracy.