Compared with domestic and industrial water demands, which remain relatively constant within each operation period, agricultural water demand varies depending on meteorological conditions such as precipitation and air temperature. Consequently, irrigation water demand is a critical factor influencing reservoir operation, especially for rice. This study focuses on irrigated rice fields supplied by the Lam Pao Dam in northeastern Thailand, located in the Mekong River basin, where hydrological and irrigation data are limited. We utilized rice production data, which are relatively well documented, together with a crop growth model to inversely estimate the actual irrigation water use.
The study area was the Lam Pao Irrigation Project in Kalasin Province (approximately 49,000 ha), supplied by the Lam Pao Dam, which has an effective storage capacity of 1.4 billion m³. First, meteorological data for 2011–2020 were input into the crop growth model, and irrigation water use was inversely estimated by calibrating the model to reproduce the observed provincial rice yield statistics for Kalasin Province. The estimated daily irrigation water demand for 2010–2018 was then used to train a Long Short-Term Memory (LSTM) model, while data from 2019–2020 were used for model validation. Batch sizes ranging from 16 to 256 and past window lengths of 15, 20, 25, 30, and 35 days were evaluated.
The predicted irrigation water demand was compared with the actual irrigation releases from the Lam Pao Dam during 2019–2020. The LSTM model achieved an RMSE of 9% and a correlation coefficient of 0.94 for one-day-ahead prediction, while the three-day-ahead prediction yielded an RMSE of 16% and a correlation coefficient of 0.77. These results demonstrate that the proposed deep learning approach can accurately predict short-term irrigation water demand and has strong potential for supporting efficient reservoir operation in data-scarce irrigation systems.