A novel error correction modelling framework for reducing prediction uncertainty in reservoir levels for operational control

Accurate reservoir level prediction is vital for effective reservoir operation. While an overprediction of the reservoir levels results in the release of excess water than required, an underestimation leads to insufficient water supply. This affects the multiple purposes served by the reservoir, such as domestic and municipal water supply, irrigation, hydropower generation, and flood control. However, predicting reservoir levels accurately is complex and challenging owing to the errors arising from the hydrological and routing models. This affects the accuracy of the predicted reservoir levels and incorporates uncertainty. Thus, it is vital to explore measures to reduce the error in the predicted reservoir levels to improve their reliability. The study thus proposes a novel error correction modelling framework for reducing the prediction uncertainty in the reservoir levels. For this endeavor, the state of the art of machine learning models is exploited. The proposed framework integrates an optimization technique with machine learning models to reduce the error in the predicted reservoir levels. The framework was tested on the Pong reservoir, India, and evaluated using several performance indices including the normalized root mean square error (NRMSE), Nash Sutcliffe efficiency (NSE), and percentage of coverage (POC). The evaluation revealed improvements in accuracy and a reduction in uncertainty of predicted reservoir levels. For example, the NRMSE of the predicted reservoir levels improved from 0.132% to 0.002% and 0.416% to 0.397% during calibration and validation respectively, while the percentage of coverage improved from 45% to 77.5% (calibration) and from 27.27% to 36.36% (validation). The framework thus has the potential to improve reservoir operational control and associated decision-making.