Short-term water level prediction using different artificial intelligent models

Artificial intelligent models (AIMs) have been successfully adopted in hydrological forecasting in a plenty of literatures. However, the comprehensive comparison of their applicability in particular short-term (i.e. hourly) water level prediction under heavy rainfall events was rarely discussed. Therefore, in this study, the artificial neural networks (ANN), support vector machine (SVM) and adaptive neuro fuzzy inference system (ANFIS) were selected for comparison. A study case in Yaojiang watershed, southeast China was selected for demonstration. The forecasting models were compared in terms of (i) using different lags as modeling inputs; (ii) using training datasets of heavy rainfall events only and all years; (iii) performance in multistep ahead (i.e. from 1 hour to 6 hours) forecasting; and (iv) performance in peak values and all values. The results showed that ANN model trained with datasets sampling from the duration of heavy rainfall performed better than that in using all datasets (including baseflow). The 1- to 4-hour ahead forecasting based on previous 2-hour inputs is the suggested modeling scenario. ANFIS showed better ability of avoiding information noise with different lags of inputs. SVM is relatively more robust during peak values under extreme typhoon events. Overall, this study shed some lights on the best practices in using various artificial intelligent methods for prediction of hourly water level with different model settings. Further studies were desired to issue the early warning with the availability of short-term flash flood forecasting.