Investigation of the power production forecasting problem from hydroelectic reservoirs across different scales

I. Vatsikouridis, Investigation of the power production forecasting problem from hydroelectic reservoirs across different scales, Diploma thesis, 139 pages, Department of Water Resources and Environmental Engineering – National Technical University of Athens, March 2022.

[doc_id=2180]

[Greek]

Hydroelectric power from large-scale dams is time-regulated, which makes it a reliable source of energy production, especially in peak periods. In the present study, inflow forecasts are conducted for the Kremasta, Mesochora and Evinos reservoirs with the use of three different deterministic models and for two forecast time horizons on monthly scale (4 and 12 months ahead). Firstly, the Autoregressive Model of first order AR(1) is used, with the advantage of its small computational load. Then, artificial neural networks are used, and specifically the LSTM (Long-Short Term Memory) model, which generally has good performance in serial data but big computing time as a drawback, and finally, we try the KNN algorithm (K-Nearest Neighbors), a nonparameterized machine learning method based on the similarity of the historical data with the predicted ones. After the inflow forecasts, the latter is converted into a forecast of energy production through the simulation of the water balance of each reservoir and its respective operating rules. Finally, it turns out that the simple AR (1) model has very good performance in medium and low flows but cannot adequately predict their variability during the winter months. In contrast, the two machine learning models are significantly more capable. The LSTM model has the best performance in terms of 4-month forecasts, which is reflected on the financial evaluation that has been studied. After the point forecasts were made, the best-performing model (LSTM) was used to make probabilistic predictions using copulas. As far as the operational level of the reservoirs is concerned, the quantification of predictive uncertainty is as much, if not more important, as it prepares the project manager for both the optimistic and the pessimistic scenario. Our analyzes focus on the different performance of the three deterministic models for each reservoir due to their different scale, operation and capacity, which gives rise to the idea of creating a hybrid forecast model that takes advantage of the strengths of each model.

PDF Full text: