A Hybrid Approach for Predicting River Runoff

Abstract

Time series prediction has attracted attention of many researchers as well as practitioners from different fields and many approaches have been proposed. Traditionally, sliding window technique was employed to transform data first and then some learning models such as fuzzy neural networks were exploited for prediction. In order to improve the prediction performance, we propose an approach that combines chaotic theory, recurrent fuzzy neural network (RFNN), and K-means. In the past few decades, fuzzy neural networks have been proven to be a great method for modeling, characterizing and predicting many kinds of nonlinear hydrology time series data such as rainfall, water quality, and river runoff. Chaotic theory is a field of physics and mathematics, and having been used to solve many practical problems emerging from industrial practices. In our proposed approach, chaotic theory is firstly exploited to transform original data to a new kind of data called phase space. Then, a novel hybrid model namely RFNN-KM including several RFNNs that are mixed together by K-means algorithm is used to perform prediction. We conduct experiments to evaluate our approach using runoff data of Srepok River in the Central Highland of Vietnam. The experiment results show that the proposed approach outperforms the one combining RFNN and sliding window technique on the same experiment data.