MODELLING AND EVALUATION OF ENCODER ATTACHED RNN FOR THE PREDICTION OF DAM INFLOW DURING SMALL AND MEDIUM-SIZED RAINFALL EVENTS

Abstract

 Prediction of dam inflow for small and medium-sized rainfall events is strongly affected by the noise of dam inflow, caused by the fluctuations of the reservoir’s water level. In addition, recurrent neural net-work (RNN), which is one of the prevailing deep learning models, have difficulty in adapting the condition of the basin preceding the target rainfall events. From these viewpoints, firstly, we applied Wavelet transforms to reduce the noise of dam inflow during the flood period. Secondly, we proposed encoder attached RNN to adapt the prior condition of the basin. Finally we compared it with multi-layer perceptron (MLP) to verify the accuracy of the predictions up to 24 hours.