Weekly and Monthly Forecasting Rainfall Model based on LSTM

The climate of Iraq has become increasingly variable in recent years, characterized by high temperatures and low rainfall. Rainfall plays a crucial role in agriculture in Iraq and thus affects the economy. Rainfall prediction has become essential for the favorable management of rainfall in various aspects of life. In this research, weather data were collected from Hilla station of the Climate Department of the General Authority of Meteorology and Seismology in Iraq for the period from 2012 to 2022. The data consist of several columns: date, wind speed, maximum temperature, minimum temperature, relative humidity, sea pressure, normal temperature, and rainfall. The time series data used with the long short-term memory method represents one of the most effective applications of deep learning techniques. Two LSTMs were trained the first time using all available features, which are 6 features, in addition to training the LSTM and the inputs were the influential features that gave high values in the correlation matrix (wind speed, sea pressure, and relative humidity) to achieve accuracy and reduce the prediction error of rainfall. The weekly and monthly forecasts made with the influential features outperformed the forecasts made with all features. The evaluation metric (root mean square error) showed lower error when using all data columns (RMSE = 0.05 and RMSE = 0.025) for weekly and monthly forecasts, respectively, and less errors when using only a limited number of columns (RMSE = 0.04 and RMSE = 0.01) for weekly and monthly forecasts, respectively.