Author(s): Nobuaki Kimura Ikuo Yoshinaga; Kenji Sekijima Issaku Azechi; Daichi Baba
Linked Author(s):
Keywords: ANN model; Long short-term memory; Drainage system; Water level
Abstract: Automatic and adaptive pumping operations on drainage management in a lowland have been required to reduce run costs on efficient, regular pumping and promote effective water-supply for paddy fields and proper controls for flood events. To satisfy these requirements in controlling the entire water volume in the lowland, an effective and efficient, real-time prediction model is required. The model is usually run using observed data from the fields. For a sustainable operation with reduction of run cost and manpower, this study focused on investigating minimum observed instruments based on locations and data items to maintain accurate model predictions. We employed the long short-term memory (LSTM) model as a data-driven model, which can predict long-term, time-series data accurately. The LSTM model that predicts water level and discharge was implemented to a mid-size agricultural lowland with a complicated drainage system. In the area, several stations for intake (four irrigation pumping stations) and for drainage (five drainage pumping stations), numerous canals and a regulating pond exist. Continuous, long-term observed data are available. Water level predictions were conducted with a variety of cases with different inputs based on the number of stations and the combinations of data items (e.g., water level & rainfall, and water level & drain discharge) during April to August (irrigation season). The error evaluation was conducted by K-fold cross-validation. The results showed that the information of water level as input data was greatly crucial for accurate outputs and that four irrigation pumping stations and a main drainage pumping station were proper to minimize observed locations during irrigation season.
Year: 2020