A Generative Model for Rainfall Prediction based on Variational Autoencoder (VAE) Using Time-Series Weather parameters

¹ Associate Professor, Department of Computer Science and Applications, Sri Chandrasekharendra Saraswathi Viswa Mahavidyalaya, Kanchipuram, Tamil Nadu, 631561 – India
² Research Scholar, Department of Computer Science and Applications, Sri Chandrasekharendra Saraswathi Viswa Mahavidyalaya, Kanchipuram, Tamil Nadu, 631561 – India
³ Associate Professor, Department of Computer Science and Engineering, Sri Chandrasekharendra Saraswathi Viswa Mahavidyalaya, Kanchipuram, Tamil Nadu, 631561 – India

This email address is being protected from spambots. You need JavaScript enabled to view it.
This email address is being protected from spambots. You need JavaScript enabled to view it.
This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Rainfall prediction has always been elusive and challenging due to the stochastic nature of the atmospheric processes involved in the rainfall occurrence, especially under extreme weather conditions. Traditionally, various deterministic and statistical methods have been used for rainfall prediction; however, they have failed to capture the uncertainty involved in the process. In this paper, a novel generative probabilistic rainfall prediction framework based on Variational Auto Encoders (VAE) is proposed. The method encapsulates weather information in a compact dimension mapped to a latent space that governs rainfall generation. The Encoder learns to transform the daily weather parameters into a probabilistic cloud in a low-dimensional latent space characterized by mean and variance, capturing both the prominent atmospheric configuration and its uncertainty. The Decoder samples the clusters that emerge in the latent space due to recurring atmospheric situations and outputs the probability of rainfall occurrence for a given day’s weather parameters. The experimental analysis on weather data shows that the VAE-based approach improves the probabilistic accuracy and uncertainty calibration in comparison to deterministic methods. The proposed generative framework provides an interpretable latent representation of atmospheric states for reliable rainfall prediction.