APPLICATION OF RECURRENT NEURAL NETWORK FOR DAM INFLOW PREDICTION

Abstract

We applied Recurrent Neural Network (RNN) and Long-Short Term Memory (LSTM) for flood forecasting problems and compared those models with Multi-Layer Perceptron (MLP) for the prediction of inflow to Yabakei Dam in the Yamakuni River system. We investigated RMSE value for the predictions of inflow against five validation floods. It was found that RNN is relatively reproducible model, although the variation due to the initial values is larger than that of MLP. We took the average for the predictions against the model parameters with different random seeds in ensemble model. Investigating the characters of these ensemble models, we got the three results: (1) The RMSE value was around or less than the median for all of the three models. (2) The average of the RMSE values against five validation floods for RNN was slightly smaller than that for MLP. (3) RNN was likely to be robust against each prediction period than MLP.