Recurrent Neural Network Predictions for Water Levels at Drainage Pumping Stations in an Agricultural Lowland

Abstract

Drainage management in a complicated system in an agricultural lowland must operate pumps flexibly and quickly, based on the water level at the pumping station. A data-driven model without any physical-based information was implemented in a complicated drainage management system to predict the water level of a lagoon near a main drainage pumping station. We employed a long shortterm memory (LSTM) model as an advanced neural network model to utilize the field datasets obtained from water-related facilities and sensors over about eight years as model input data. We performed sensitivity tests for model accuracy with different types of data and locations of data using cross-validation with an error quantity between observed and predicted water levels at the main drainage pumping station. The results showed that the LSTM model with the input of all available datasets predicted better than the models using several parts of datasets or it was roughly equivalent to those for water levels over the entire observed period in 3-h and 6-h lead times. In addition, the LSTM with only inputs of the water level and rainfall observed by drainage pumping stations performed better for the observed subperiod, including the severest flood event.