Comparative Surrogate Modeling of Reservoir Levels in the Drin Hydropower Cascade Using Hybrid and Data-Driven Approaches for Short-term Predictions

This paper proposes a data-driven surrogate modeling approach for short-term prediction of the reservoir level in the Drin hydroelectric cascade of the Fierza, Koman and Vau Dejes reservoirs in Albania. The suggested methodology combines local operational factors such as inflow, energy production and historical reservoir levels with the influence of upstream reservoir operations to reflect the interconnected nature of the cascade. Three alternative modelling approaches are adapted and compared. These include linear regression methods (Naive and Ridge regression), nonlinear Radial Basis Function (RBF) networks and a hybrid RBF-Kalman Filter model. The RBF models are trained offline to learn the nonlinear relationships in the system, and the Kalman Filter makes the parameters adaptable with time to enhance the system response to the change of conditions. The model’s performance is evaluated based on daily operating data collected from 2020 to 2022. The evaluation performance is based by error measurements: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R2). Model implementation and performance is dependent on the variability of reservoir dynamics. The hybrid system between dams reaches forecast accuracy of 93% for all three hydropower plants. The recursive hybrid model improves forecasting accuracy for each dam. The proposed methodology provides a flexible and physically relevant surrogate modelling solution for hydropower cascades, applicable for short-term forecasting and operational decision support.