Daily inflow forecasting in Asomata reservoir, on Aliakmon River, using Long Short-Term Memory network

¹ National University of Science and Technology POLITEHNICA Bucharest, Faculty of Energy Engineering, Department of Hydraulics, Hydraulic Machinery and Environmental Engineering, 313 Spl. Independentei, sector 6, RO-060042, Romania
² National Directorate of Hydro Generation Operation, PPC S.A., Asomata Hydroelectric Power Plant, Aliakmonas Complex, 59150, Veroia, Imathia, Greece

^* Corresponding author: eliza.tica@upb.ro

Abstract

In this paper, the forecast of the daily inflow water volumes in the Asomata reservoir, on the Aliakmon River, Greece, based on long short-term memory was realized. A MATLAB program was developed for one day ahead prediction; the model was calibrated for the period 2011-2021 based on the historical values and used in a repetitive loop for 365 days corresponding to the testing year 2022. An improvement in the forecasting was observed when the daily index associated with the forecasted variable was used as input. This improvement was much more important than when the associated precipitation values were used in the absence of the day index. A possible explanation would be the operating schedule of the plants upstream, Asomata reservoir being the fourth in a five reservoirs cascade. More than that, the upstream hydropower plant, Sfikia, is a pumped storage plant, Asomata being lower reservoir for this one. The results are in good agreement with the measurements, confirming the fact that long-short term memory networks are widely used in the hydrological forecasting with good results.