Water level forecasting using multi-model ensemble approach at Temengor hydropower reservoir

Thumbnail Image
Hidayah Basri, Dr.
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Reservoir water level forecasting systems serve as crucial element for optimum reservoir operation, especially during flood seasons where huge amounts of inflow will enter a reservoir. Failure in proper reservoir operation can have catastrophic impacts, such as dam failure. Consistent, accurate, and timely water level forecasting allows dam managers to release water gradually for flood control in the downstream areas. Every hydrological model has its own unique capability to simulate the catchment behaviour, however, the results produced from any hydrological model, regardless of its complexity, are also subjected to uncertainties that occur at every stage of the modelling process. The beginning of this study explores the challenges of developing forecast models, as they have the potential to lead to modelling uncertainties. This can result from insufficient rain gauge data to represent entire catchment areas, and the shortfall of streamflow stations for calibration processes. Calibration processes often rely on observed reservoir water levels, which are influenced by the operation of the turbine intake – due to its installed location. This study aims to evaluate the ability of four conceptual, models namely Génie Rural 4 Paramètres Horaire, Integrated Flood Analysis System, Probability Distributed Model and Nedbør Affstrømnings Model, to simulate the water level at Temengor reservoir, and to test the ability of an ensemble approach to improve modelling accuracy. In this study all four models were developed and calibrated to achieve satisfactory model performance. The outputs from the models are then combined using multi-model ensemble methods to produce a single output. There are three statistically based ensemble methods employed in this study, namely the Standard Average Method, the Weighted Average Method, and the Multi-model Super Ensemble. The machine learning based ensemble methods are Back Propagation Neural Networks. During the validation period, comparable performance was found for three out of four models developed in this study. The Probability Distributed Model, Génie Rural 4 Paramètres Horaire and Nedbør Affstrømnings Model models are classified as having very good performance, with Nash-Sutcliffe Coefficient of Efficiency values ranging from 0.83 to 0.88. Simulations using ensemble methods significantly improved the overall model accuracy, with Nash-Sutcliffe Coefficient of Efficiency values up to 0.91 using the Standard Average Method. The final output from this study is a model selection matrix for ensemble modelling approaches, based on seven rainfall and turbine outflow categories, to assist dam operators in choosing the best operational water level forecast model for the Temengor dam, for accurate decision making, in order to ensure timely responses to any flood events, as part of disaster preparedness and dam safety measures.
Water level forecasting using multi-model