Improving Sub-daily Runoff Forecast Based on the Multi-objective Optimized Extreme Learning Machine for Reservoir Operation

Data-driven models have shown remarkable achievements in runoff prediction, but their simulation results can be overly homogenized due to the distillation of all simulation aspects into the loss function. This can make the models unreliable for predicting extreme events, leading to issues subsequent reservoir operations. This paper proposes a novel data-driven hybrid machine-learning model called the multi-objective optimized extreme learning machine (MOELM) to provide an accurate runoff forecasting for reservoirs. The objective is to minimize simulation error, with an additional focus on flood deviation. The results show that: (1) MOELM can improve flood events prediction, reducing the root mean square error (RMSE) for flood series by 5.27% without increasing the overall prediction error at Longtan reservoir. Compared to hydrological models, MOELM can reduce operational risk with lower reservoir maximum outflow and water level during typical flood events, and it can potentially increase hydropower generation at Longtan reservoir by 130 million kW·h. (2) MOELM can be transferred to other cross-sections with excellent performances, demonstrating hydrological transferability from fluctuation to flatness in regime. (3) Partial mutual information is introduced for input variable selection, with discharge at lag times t-4, t-1, t-8, and t-2 being vital to the prediction model. Our model is practical, requiring no additional input, fitting the hydrological runoff holistically, and capable of providing accurate flood forecasts.