Capability of artificial intelligence models in inflows forecase : a case study in eastern river basin in thailand

Abstract

Efficient reservoir operation scheme depends on accuracy of forecasted inflow to reservoir. The study aims to investigate whether artificial intelligent model is capable to improve accuracy of forecast or not. Bang Phra and Nong Pla Lai reservoirs, situated in eastern river basin of Thailand, were selected as case study. Monthly rainfall above both reservoirs and monthly inflow to reservoir during 1994-2023 were collected for training of model. The Long Short-Term Memory (LSTM) model was compared with ARIMA, a widely used statistical model. Results during training phase of LSTM Model revealed that present inflow and one-month antecedent inflow together with monthly rainfall above reservoir provided better accuracy of forecast than using only inflow to reservoir as input to model. Testing of accuracy LSTM model in time of forecast showed that time of forecast 3 months ahead provided better accuracy of forecast than 6 and 12 months ahead, with 85%, 80% and 52% accuracy, respectively. Comparison of accuracy of forecast of LSTM Model with traditional ARIMA model revealed that averagely LSTM Model provided better accuracy than ARIMA model. At Bang Phra reservoir, LSTM Model provided accuracy of forecast from 53% to 74%, while ARIMA Model provided accuracy of forecast from 38% to 79%. At Nong Pla Lai reservoir, LSTM Model provided accuracy of forecast from 53% to 84%, while ARIMA Model provided accuracy of forecast from 48% to 70%. Moreover, during validation period of model using data between 2016-2022, it has been found that LSTM Model was able to forecast inflow during rainy season better than ARIMA model, which will consequently improve efficiency of reservoir operation scheme.