This paper presents a comparative study aimed at evaluating the performance of four deep learning models for forecasting water levels in the Lake Chad basin over the period 2026–2050. The models studied are based on two distinct neural network architectures. The first is based on recurrent networks, namely LSTM and GRU. The first is based on recurrent neural networks, specifically LSTM and GRU. The second relies on attention mechanisms, specifically Transformer and Informer. The experimental results show that models based on the Transformer architecture and in particular its improved version, Informer—outperform other models according to the statistical metrics used. The performance metrics obtained with the Informer model are as follows~: MAE~$= 0{,}149$, RMSE~$= 0{,}202$, MAPE~$= 0{,}053$, $R^2 = 0{,}862$, and NSE~$= 0{,}862$. This model has demonstrated strong predictive power, particularly in capturing long-term dependencies in time series.

Water level, Lake Chad basin, Deep learning, Forecasting