Deep Learning Models Based on CNN, RNN, and LSTM for Rainfall Forecasting: Jordan as a Case Study

Abstract

This study is the first to compare deep learning models for rainfall prediction across several Jordanian cities representing diverse climates using 11 years of recorded climate data, something that previous studies have not addressed in the Jordanian context. The climate records for four Jordanian cities (Amman, Irbid, Karak, and Ajloun) were recorded hourly. The data was divided into training sets (80%) and test sets (20%), with and without the application of correlation analysis, feature selection, and data standardization steps applied. Three neural network models, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) were used to evaluate the performance of rainfall prediction in the four cities using three sets of features: all variables (13 features), precipitation only, and a selection of eight correlated features. The results showed that the RNN model outperformed the others overall, especially when using correlated features, recording the lowest error values, Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) in most cities, with the exception of Amman, where the model performed best when using all features. Whereas in Irbid, the MSE was 0.0802×10⁻3 and RMSE = 0.009, while in Karak, the MSE was 0.118×10⁻3 and RMSE = 0.0109. In Amman, the RNN using all features achieved MSE = 0.0167×10⁻3 and RMSE = 0.0041.