WATER LEVEL PREDICTIONS THAT VISUALIZE UNCERTAINTY USING LSTM COUPLED WITH ALTERNATIVE METHODS IN BAYESIAN INFERENCE AT A RESERVOIR NEAR A DRAINAGE PUMPING STATION

Abstract

 This study shows that we implemented two alternative methods, instead of a normal Bayesian inference method, into Long, Short-Term Memory (LSTM) to visualize model-related uncertainties. This model is called Bayesian neural networks (BNN), which was applied to predict water levels in a reservoir used for drainage management in a lowland area. In BNN, we used two alternative methods: Monte Carlo dropout (MC Dropout), which is a method of randomly switching nodes in the network, and Stochastic Gradient Langevin Dynamics (SGLD), which is a sampling method. The BNN predictions of MC Dropout with the best dropout rate (=0.3) were performed by 10% improvement of accuracy when compared to the conventional model in short lead time for the drainage period during the largest flood. SGLD-based BNN had equivalent results to the BNN-MC Dropout. Comparing among the predictions of both methods, including credible intervals, MC Dropout-based BNN showed wider and smoother temporal distributions, especially near the peak water level.