Enhancing spatiotemporal weather forecasting accuracy with 3D convolutional kernel through the sequence to sequence model

Abstract

Accurate weather forecasting is important when dealing with various sectors, such as retail, agriculture, and aviation, especially during extreme weather events like heat waves, droughts, and storms to prevent disaster impact. Traditional methods rely on complex, physics-based models to predict the Earth's stochastic systems. However, some technological advancements and the availability of extensive satellite data from beyond Earth have enhanced meteorological predictions and sent them to Earth's antennae. Deep learning models using this historical data show promise in improving forecast accuracy to enhance how models learn the data pattern. This study introduces a novel architecture, convolutional sequence to sequence (ConvSeq2Seq) network, which employs 3D convolutional neural networks (CNN) to address the challenges of spatiotemporal forecasting. Unlike recurrent neural network (RNN)--based models, which are time-consuming due to sequential processing, 3D CNNs capture spatial context more efficiently. ConvSeq2Seq overcomes the limitations of traditional CNN models by ensuring causal constraints and generating flexible length output sequences. Our experimental results demonstrate that ConvSeq2Seq outperforms traditional and modern RNN-based architectures in both prediction accuracy and time efficiency, leveraging historical meteorological data to provide a robust solution for weather forecasting applications. The proposed architecture outperforms the previous method, giving new insight when dealing with spatiotemporal with high density.