Prévision des précipitations dans la zone de maga, Extrême-Nord Cameroun, par un modèle hybride LSTM-GRU.

Abstract

Precipitation forecasting is essential for agricultural management, water resource planning, and mitigating risks related to natural disasters, particularly in arid regions like Maga in the Extreme North of Cameroon. It is a complex issue due to climate change variations caused by numerous factors. Previous studies have employed numerical time approaches, statistical methods, and machine learning techniques to predict precipitation from other meteorological variables. Building on existing work, we proposed in this thesis to use a hybrid model combining LSTM and GRUneural networks, based on a dataset downloaded from the NASA website to improve the accuracy of precipitation forecasts in this region. The approach relies on leveraging the capabilities of LSTM and GRU neural networks to model complex time series and capture long-term dependencies from historical data such as temperature, humidity, and wind speed collected over several years. After several experiments and adjustments of hyperparameters, the optimized configuration successfully predicts monthly precipitation amounts, despite climate changes. It shows a MAE of 11.974, an R2 of 0.905, and an RMSE of 15.982 on the validation and test datasets. In future studies, integrating additional data and exploring new model architectures could further enhance forecasting accuracy, thereby strengthening community resilience to growing climate challenges