An Attention-based Rainfall Prediction System using LSTM Model and Convolutional Block Module

For practical purposes including managing water supplies, building infrastructure, and growing food, rainwater prediction is crucial. Since precipitation time series inherently includes a non-stationary component, the model’s effectiveness is diminished for both practicing hydrologists and prospective drought investigators. A model that predicts the likelihood of rain, based on Deep Learning (DL), was proposed by many researchers as a solution to the weather prediction issue. The need for study into the challenges of short-term rainfall quantity prediction has grown in response to climate change. In particular, the high computing needs and error proneness of the current numerical weather prediction models make them unsuitable for precipitation forecasting. Predicting the quantity of precipitation is difficult because it requires long-term information on many environmental phenomena, including temperature, humidity, wind direction, and more. Three attention models namely Squeeze and Excitation (SE), Efficient Channel Attention (ECA), and Convolutional Block Attention (CBA) are utilized with Long Short-Term Memory (LSTM) in this work. CBA with LSTM performed best in terms of Root Mean Square Error (RMSE). CBA with LSTM achieved a very low RMSE value of 0.28 when compared to other attention models.