Multiple Machine Learning Methods for Runoff Prediction: Contrast and Improvement

Abstract

Abstract Machine learning methods provide new alternative methods and ideas for runoff prediction. In order to improve the application of machine learning methods in the field of runoff prediction, we selected five rivers with different conditions from north to south in Japan as the research objects, and compared the six watersheds and different types methods of time series prediction in machine learning methods, to evaluate the accuracy and applicability of these machine learning methods for daily runoff prediction in different watersheds, and improve the commonality problem found in the prediction process. The results show that before the improvement, the prediction results of the six methods in Kushiro river, Yodogawa river and Shinano Gawa river are good. After the improvement, the runoff prediction errors of the six methods in the five watersheds are greatly reduced, and the prediction accuracy and applicability are greatly improved. Among them, the improved deep temporal convolutional network (DeepTCN) has the best prediction effect and applicability. Of all prediction results in the five watersheds, the NSE coefficients are above 0.94. In general, the improved DeepTCN has the best comprehensive prediction effect, and has the potential to be widely recommended for runoff prediction