High-Flow Performance of Machine Learning-Based Models for 30-Minute River Discharge Forecasting

Reliable short-term flood forecasting and warning systems remain challenging due to the uncertainties during extreme events. This investigation aims to develop Machine Learning (ML) algorithms and establish a river discharge forecasting model to assist decision-makers in an urbanised catchment. In this study, the Pinxton Catchment, located within the River Erewash basin in Derbyshire, East Midlands, United Kingdom, evaluates the capability of standalone models to effectively predict river discharge in a small and fast-response catchment using ROOT Mean Square Error (RMSE) and Nash-Sutcliffe Efficiency (NSE), and hydrograph comparison. Therefore, three Machine Learning algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Extreme Learning Machine (ELM) are explored using the same 80% training, 20% testing, and evaluation for each model's performance with 15-minute time resolution observed discharge and rainfall data. The findings indicated that at the 30-minute forecasting horizon, ELM demonstrated reliable RMSE- and NSE-based performance for river discharge forecasting. RF and XGBoost, as tree-based algorithms, generated comparatively smooth predictions and tended to underestimate the extreme events. Their smoothing behaviour reduced responsiveness to abrupt changes in discharge and peaks. Although all three models reproduced low- and moderate-flow conditions reasonably well, it remains challenging to capture the nonlinear relationship in real-time time-series data during flood forecasting. ELM followed rapid temporal variations more closely, which represents the potential for further investigation with an RMSE of 0.1891 m³/s and an NSE of 0.9150. The analysis evaluates the high-flow reliability of three standalone ML models under a 30-minute-ahead forecasting condition, whereas peak magnitude predictions remained challenging. In conclusion, this approach was expected to be useful to support river discharge forecasting, and future work will investigate whether model optimisation and ensemble techniques can improve the strategies while introducing the outperformed model.