WATER LEVEL PREDICTION AT DRAINNAGE PUMP STATIONS IN LOW LANDS USING LSTM MODEL

Abstract

 We have developed a data-driven Artificial Neural Network (ANN) model, capable of predicting water levels only with observed data sets, to support the drainage operations at the pump stations in low landareas. We tested two ANN models: a conventional model (Multiple Layer Perceptron, MLP) and an updated model (Long Short-Term Memory, LSTM), which effectively works on time-series predictions, in two lowland areas where different drainage systems exist. As to the simple darainage system in a small area, the LSTM is superior to the MLP with approximately 10% improvement of water level predictions within 2 hours. In addition, the LSTM predicted upto 3-hour water level, approximately 6% better than the MLP during the heavry rainfall event even in a larger, complex drainage system.