Incorporating Novel Input Variable Selection for Improved Precipitation Forecasting in the Different Water Basins of Thailand

Precipitation forecasting is essential in water resource planning and management, particularly in tropical countries like Thailand. Selecting appropriate input variables for developing prediction models for rainfall is a significant difficulty. Recent studies in various disciplines have highlighted the utility of artificial intelligence-based techniques for determining explanatory variables for use in non-linear scenarios, which remain largely unexplored in rainfall forecasting. The present study was carried out to fill this knowledge gap. Two river basins in the northern region of Thailand were selected as a study area. Monthly observation and large-scale climatic variables (LCVs) at both River basins from 1993 to 2022 were used for model development. This study proposed a novel hybrid bootstrapped long short-term recurrent neural network (BTSP-LSTM-RNN) for input variables selection (IVS) for monthly precipitation forecasting. A novel BTSP-LSTMRNN model was compared with the support vector regression with recursive feature elimination (SVR-RFE) and gradient boosting (GB). For the evaluation of these models, statistical metrics such as coefficient of determination (R2), mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) are used. Remarkably, the proposed BTSP-LSTM-RNN hybrid model outperformed other models, achieving a notably higher R2 value of 0.85 compared to SVR-RFE (0.68) and GB (0.72). BTSP-LSTM-RNN exhibited a remarkably low MAE of 0.0094, underscoring its accuracy in IVS, whereas SVR-RFE and GB achieved 120.92 and 123.7, respectively. The RMSE value of 116.62, although higher than the MAE, still indicates a relatively low error compared to SVR-RFE and GB. Most notably, the BTSP-LSTM-RNN model demonstrated the lowest MAPE among the three models, standing at 27.88