Applying Deep Learning to Sub-seasonal to Seasonal (S2S) prediction of rainfall over Burkina Faso

Abstract

This study proposes to predict rainfall on a Sub-seasonal to seasonal (S2S) time scale over six (6) locations (Dori, Ouahigouya, Ouagadougou, Fada N’Gourma, Bobo-Dioulasso and Gaoua) in Burkina Faso, using a specific architecture of Deep Learning called LSTM. Historical monthly and daily climate parameters from different sources are used to calibrate the LSTM model. After data preprocessing, the model is set and run for each location. Afterwards, the model is evaluated using some statistical metrics such as R2, NSE, RMSE and MAE. The performance evaluation of the model using these metrics shows that LSTM model is effective and performs well in predicting rainfall at monthly timescales. For instance, forecasting at monthly timescale exibits a R2 ranging from 0.66 to 0.83, NSE ranging from 0.62 to 0.80, RMSE ranging from 32.9 to 59.9mm and MAE ranging from 21.1 to 39.7mm. Regarding the bimonthly rainfall prediction, R2 ranges from 0.63 to 0.83, NSE ranges from 0.6 to 0.82, RMSE ranges from 34.0 to 62.8mm and MAE ranges from 21.8 to 42.8mm. These results allow to highlight the impact of climatic zones and topography. Indeed, the models have better results on slightly humid plateaus than on very rainy and elevated areas of the country. However, trying to bring these monthly forecasts down to daily scales, the models struggle to capture daily rainfall for all locations. This requires more investigations to be done as part of future studies.