Investigation of Multimodel Ensemble Performance Using Machine Learning Method for Operational Dam Safety

Abstract

The efficient and effective management of hydropower reservoirs is vital for hydroelectric power plant operation. Therefore, accurate and reliable flow forecasting forms an important basis for efficient real-time hydropower reservoir operation. The inflow forecast modeling process involves various computations and modeling techniques, which results in uncertainties various factors. The inflow forecasting techniques also may vary with the purpose of the system, physical characteristics and availability of data. Over the decades, there is an increasing awareness of the risk of relying on a single model among researchers and practitioners. There is clearly a potential danger in relying entirely on one rainfall-runoff model in such systems, as each model, provides, through its forecast, an important source of information that may be different in some detail from those of the other models calibrated with the same data set. Moreover, the failure of the model to yield consistent and reasonably accurate forecasts may undermine its credibility and results in poor reservoir planning and operation. Hence, consideration of the development of more flexible inflow forecasting systems is needed. These will not be based on a single substantive rainfall-runoff model, but efficient utilization of the synchronous flow forecasts from several substantive rainfall-runoff models, each having different strengths and weaknesses, to produce improved forecasts. In this study, multi-model ensemble strategies will be developed by considering outputs from a committee of models. Multi-model ensemble methods using a machine-learning algorithm combining different model outputs to the water level forecasts will be applied. The performance of the different machine-learning algorithms against the component models and accuracy improvement will be investigated.