Data-driven novel deep learning applications for the prediction of rainfall using meteorological data

Abstract

Rainfall plays an important role in maintaining the water cycle by replenishing aquifers, lakes, and rivers, supporting aquatic life, and sustaining terrestrial ecosystems. Accurate prediction is crucial given the intricate interplay of atmospheric and oceanic phenomena, especially amidst contemporary challenges. In this study, to predict rainfall, 12,852 data points from open-source global weather data for three cities in Indonesia were utilized, incorporating input variables such as maximum temperature (°C), minimum temperature (°C), wind speed (m/s), relative humidity (%), and solar radiation (MJ/m2). Three novel and robust Deep Learning models were used: Recurrent Neural Network (DRNN), Deep Gated Recurrent Unit (DGRU), and Deep Long Short-Term Memory (DLSTM). Evaluation of the results, including statistical metrics like Root-Mean-Square Errors and Correction Coefficient (R2), revealed that the Deep Long Short-Term Memory model outperformed DRNN and Deep Gated Recurrent Unit with values of 0.1289 and 0.9995, respectively. DLSTM networks offer several advantages for rainfall prediction, particularly in sequential data like time series prediction, excelling in handling long-term dependencies important for capturing weather patterns over extended periods. Equipped with memory cell architecture and forget gates, DLSTM networks effectively retain and retrieve relevant information. Furthermore, DLSTM networks enable parallelization, enhancing computational efficiency, and offer flexibility in model design and regularization techniques for improved generalization performance. Additionally, the results indicate that maximum temperature and solar radiation parameters exhibit an indirect influence on rainfall, while minimum temperature, wind speed, and relative humidity parameters have a direct relationship with rainfall.