Table 7
Comparative analysis between this study and other studies in multi-step-ahead prediction
| Research study | Output variable | Methods | Data size | Highlights of results |
|---|---|---|---|---|
| Bhagwat & Maity (2012) | River flow; Narmada river (India) | LS-SVR and ANN | 2,556 samples for training | The best NSE = 0.49 for two-step-ahead prediction using LS-SVR; Reasonably good up to 5-day-ahead predictions (NSE = 0.3). |
| Chang et al. (2014) | Inundation level of flood; Yu–Cheng Pumping Station (Taipei City) | ANN, Elman Networks and NARX Neural networks | 1,985 samples for training and testing | NARX Networks perform the best, producing coefficients of efficiency within 0.9–0.7 (scenario I) and 0.7–0.5 (scenario II) in the testing stages for 10–60-min-ahead forecasts. |
| Guo et al. (2021) | River stage; Lan-Yang river basin (Lanyan, Simon and Kavalan stations at Taiwan) | Optimized four ML techniques, namely, SVR, RFR, ANN and LBGM | ≈7,500 samples for Simon and Lanyan; ≈4,000 samples for Kavalan | All models demonstrate favourable performance in terms of  of about 0.72 for 1–6 step-ahead forecasting at all stations. The LGBM model achieves more favourable prediction than SVR, RFR and ANN. |
| Yu et al. (2011) | Water level; Heshui catchment, China | Using eight different types of ANN training algorithms | 4,749 samples for training and testing | BFGS- and LM-trained ANN models gave the best performance among all of the prediction scenarios. Obtained a coefficient of determination of around  for two-step-ahead forecasting and  for five-step-ahead forecasting. |
| This paper | River flow; Kelantan river (Malaysia) | NARX Neural networks and LSTM | 348 samples for training and testing | LSTM with a direct sequence-to-sequence produces NSE = 0.75 and NSE = 0.39 for two-step-ahead and five-step-ahead forecasting, respectively. |
| Research study | Output variable | Methods | Data size | Highlights of results |
|---|---|---|---|---|
| Bhagwat & Maity (2012) | River flow; Narmada river (India) | LS-SVR and ANN | 2,556 samples for training | The best NSE = 0.49 for two-step-ahead prediction using LS-SVR; Reasonably good up to 5-day-ahead predictions (NSE = 0.3). |
| Chang et al. (2014) | Inundation level of flood; Yu–Cheng Pumping Station (Taipei City) | ANN, Elman Networks and NARX Neural networks | 1,985 samples for training and testing | NARX Networks perform the best, producing coefficients of efficiency within 0.9–0.7 (scenario I) and 0.7–0.5 (scenario II) in the testing stages for 10–60-min-ahead forecasts. |
| Guo et al. (2021) | River stage; Lan-Yang river basin (Lanyan, Simon and Kavalan stations at Taiwan) | Optimized four ML techniques, namely, SVR, RFR, ANN and LBGM | ≈7,500 samples for Simon and Lanyan; ≈4,000 samples for Kavalan | All models demonstrate favourable performance in terms of of about 0.72 for 1–6 step-ahead forecasting at all stations. The LGBM model achieves more favourable prediction than SVR, RFR and ANN. |
| Yu et al. (2011) | Water level; Heshui catchment, China | Using eight different types of ANN training algorithms | 4,749 samples for training and testing | BFGS- and LM-trained ANN models gave the best performance among all of the prediction scenarios. Obtained a coefficient of determination of around for two-step-ahead forecasting and for five-step-ahead forecasting. |
| This paper | River flow; Kelantan river (Malaysia) | NARX Neural networks and LSTM | 348 samples for training and testing | LSTM with a direct sequence-to-sequence produces NSE = 0.75 and NSE = 0.39 for two-step-ahead and five-step-ahead forecasting, respectively. |
LS-SVR, least-square-support vector regression; BFGS, Broyden–Fletcher–Goldfarb–Shanno; LM, Levenberg–Marquardt; LGBM, light gradient boosting machine; RFR, random forest regressor.