An efficiency dynamic seasonal regression forecasting technique for high variation of water level in Yom River Basin of Thailand

The Yom River Basin is one of 25 river basins in Thailand. The Yom River Basin experiences perennial droughts and floods that heavily impact the agricultural sector. In order to reduce the impact, water management, including water level estimation, must be applied to critical basins like the Yom River Basin. An important task of management is the quantitative prediction of water levels. Four different modeling approaches were applied to forecast the average monthly water level (AMWL) data from four water level measurement stations for the wet season (May–October) and dry season (November–April) from 2007 to 2020. The forecast patterns obtained from the four approaches were similar to the observed historical values, except the upstream in wet season and downstream in dry season. Furthermore, the artificial neural network (ANN) approach overestimated forecasts for almost every station in both seasons. All four approaches were more accurate in the dry season than the wet season. This study proposed a forecasting method called dynamic seasonal regression (DSR), which was obtained by combining multiple linear regression (MLR) and the autoregressive integrated moving average (ARIMA) model of the random error from MLR. DSR was more efficient than ANN, seasonal-ARIMA (SARIMA) and a hybridized SARIMA and ANN approach (SARIMANN). On average, for all stations in wet and dry seasons, DSR reduced RMSE by over 40.86%, 9.10% and 23.07% with respect to ANN, SARIMA and SARIMANN, and MAPE by over 35.01%, 13.02% and 15.96% with respect to ANN, SARIMA and SARIMANN. The RMSE of upstream was higher than the RMSE downstream in the wet season for all methods, and the MAPE of upstream was lower than the downstream in both seasons for all methods. Moreover, the RMSE of upstream was lower than the downstream in the dry season for all methods except the ANN method.