Forecasting Lake Chad’s Water Levels: Harnessing Transformer Model for Accurate Predictions

This article presents a Transformer model for forecasting water levels in Lake Chad. This lake currently provides services to more than 45 million inhabitants spread across Cameroon, Niger, Nigeria and Chad. The monthly data, covering the period from 1960 to 2024, were recorded at the Bol station, located in the lake region, by the Agence Nationale de la Météorologie du Tchad (ANAM-Tchad). An imputation technique for missing data was applied. An automatic learning method, in this case random forest, was used to impute missing data. Monthly water level data from Lake Chad were used as input data for the Transformer model. The results show that the Transfomer model performs at an intermediate level for all measurements, compared with LSTM and GRU. The values obtained for the LSTM model are : MAE = 0.1654, RMSE = 0.2169, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.8453 and NSE = 0.8453. For the GRU model, the values are: MAE = 0.1643, RMSE = 0.2551, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.7861 and NSE = 0.7861. The values obtained by the Transformer model are: MAE = 0.1649, RMSE = 0.2250, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.8335 and NSE = 0.8335.