Prediction of monthly rainfall in selected African stations using deep learning algorithms

Abstract

Abstract Precipitation has a huge impact on human life and culture. Agriculture and public infrastructure is dependent on accurate rainfall forecasting, which can aid in water supply planning, reservoir management, and flood mitigation. In view of the climate change, it is very imperative to provide a reliable forecast of rainfall leveraging on Deep Learning algorithms. This study therefore compared the forecasting performance of five different Deep Learning (DL) algorithms in predicting rainfall in five African stations (Abuja, Cairo, Nairobi, Pretoria, and Yaoundé). Deep Learning algorithms considered in this study include: Long Short Term Memory (LSTM), Convolutional Neural Network CNN, K-Nearest Neighbor’s Regression (KNNR), and Multilayer Perceptron (MLP). These algorithms were trained using data on monthly rainfall in these stations between 31 st January, 1980 and 31 st December, 2022 as obtained from Helioclim-1 archives. Data were preprocessed and 80% of the data were used in training and 20% in validation. The forecasting performances of these algorithms were evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Result showed that although these DL algorithms were competitive in terms of their performance, MLP was found to be most suitable of these algorithms for Cairo (RMSE =5.0246), Pretoria (RMSE =47.4608) and Yaoundé (RMSE = 140.1368) while for Abuja (RMSE = 88.7868) and Nairobi (RMSE =95.3637), result favoured KNN and CNN respectively. As a result, the study concluded that MLP is superior for the majority of stations. KNN excelled in Abuja, CNN in Nairobi, and MLP in Cairo, Pretoria, and Yaoundé, indicating that each station benefits from a unique deep learning algorithm.