MODÉLISATION ET PRÉVISION DES PERTES D’EAU PAR ÉVAPORATION A L’AIDE DE MODELES D’APPRENTISSAGE AUTOMATIQUE

Abstract

Evaporation plays a significant role in water loss in arid and semi-arid zones, which presents a major obstacle for managing water resources and maintaining dam stocks. This study aims to model and predict water losses resulting from evaporation at the "Koudia Asardoun" Dam in Bouira Province, which is the second largest dam in Algeria with a storage capacity of about 640 million cubic metres, and controls a casting basin with an area of 2,787 square kilometers.To establish the physical and dynamic framework of the basin, Geographic Information Systems (SIG) and remote sensing were relied upon to determine morphometric, topographic, and hydrographic characteristics; in addition to land cover classification using Sentinel-2 satellite visuals, which showed a pasture land dominance of 56.84%. In the digital modeling segment, a comparative study was conducted between a reference linear model represented by multiple linear regression (MLR) and an intelligent nonlinear model based on artificial neural 6 networks represented by the multilayer perceptron (MLPNN). Hydrometeorological factors such as peak temperature, lowest temperature, humidity levels, and wind velocity were employed as key inputs for forecasting evaporation. The models were tested under three data egmentation scenarios (60% - 40%, 70% - 30%, 80% - 20%), and performance was evaluated based on four statistical parameters: coefficient of determination (R), Nash-Sutcliffe coefficient (NSE), root mean square error (RMSE), and mean absolute error (MAE). The results showed the ability of the adopted models to extract complex relationships between climatic factors with high accuracy, as the multiple linear regression (MLR) model achieved a coefficient of determination (R) of 0.868 in the training phase and 0.851 in the validation phase in the distribution scenario (80% - 20%). The study concludes that integrating artificial intelligence and machine learning technologies represents an advanced and effective tool to support decision-making and develop proactive strategies to manage dams and conserve surface water resources in light of current climate change.