Contribution à l'optimisation de la planification intelligente de la charge dans un réseau électrique

Abstract

In this dissertation, an approach based on Long Short-Term Memory (LSTM) recurrent neural networks was developed for the long-term forecasting of the electrical load of the 60/30 kV Guerrara substation, operated by the Ghardaïa Distribution Directorate (SONELGAZ). A chronological database covering a period of 21 years was established using historical apparent power data from the IEEE 33-Bus network and enriched with representative synthetic load data to reproduce the evolution of electricity demand over a long-term horizon. The data were preprocessed, normalized, and organized as time series to enable effective training of the proposed model. The obtained results demonstrated the ability of the LSTM model to capture temporal dependencies and nonlinear characteristics of electrical load data. The use of synthetic load data consistent with demand growth trends improved the representativeness of the training dataset and enabled more realistic simulations. The forecasting results for the real case of the Guerrara substation showed good agreement with reference values and highlighted a progressive increase in electrical load over a 15-year forecasting horizon. Furthermore, the complementary application carried out on the IEEE 33- Bus network successfully evaluated the performance of the model over a more extended time horizon.