Predictive fuzzy clustering model for natural streamflow forecasting

Abstract

Planning of hydroelectric systems is a complex and difficult task once it involves non-linear production characteristics and depends on numerous variables. A key variable is the natural streamflow. The streamflow values covering the entire planning period must be accurately forecasted because they strongly influence energy production. Currently, streamflow prediction using Box and Jenkins methodology prevails in the electric power industry. This paper suggests a fuzzy prediction model based on fuzzy clustering as an alternative for streamflow forecast. The model uses fuzzy c-means clustering to explore past data structure, and a median and pattern recognition procedures to capture similarities between streamflow history and data used for prediction. Computational experiments with actual data suggest that the predictive clustering approach performs globally better than periodic auto regressive moving average models, the current streamflow forecasting methodology adopted by many hydroelectric systems worldwide.