Enhancing Deep Neural Network Architecture in Spatio- Temporal Forecasting Through Neural Architecture Search

Spatiotemporal forecasting is indispensable for pre-dicting dynamic patterns across space and time, providing valu-able insights into phenomena such as climate changes and disease outbreaks like COVID-19. Despite its importance, modeling spatiotemporal data is challenging due to the intricate patterns influenced by both spatial and temporal factors. While deep learning offers a potent solution for spatiotemporal forecasting, manually designed architectures often struggle to capture these complexities, resulting in suboptimal accuracy. To overcome the constraints of current deep learning architectures in spatiotempo-ral forecasting, this research employs Neural Architecture Search (NAS) models. NAS facilitates the automatic discovery and refinement of effective neural network structures, addressing the shortcomings of manually designed architectures. The proposed work extends the application of NAS to existing spatiotemporal deep learning architectures, including CNN-LSTM, CNN-GRU, and CNN-BiLSTM. This enables automatic customization of deep learning structures for spatiotemporal forecasting, enhancing their effectiveness in spatiotemporal forecasting. NAS played a pivotal role in automatically exploring and optimizing the archi-tecture, enabling the network to adapt and excel in capturing complex spatio-temporal patterns for more accurate forecasting. The proposed methodology is implemented and extensively eval-uated on three diverse datasets: Traffic, Corona, and Climate datasets to highlight the versatility and efficacy of NAS-optimized models across various domains. The evaluation results show that the NAS optimized models show 7%, 3% and 17% improvement in having reduced RMSE values compared to the manually designed models. This reduced RMSE values across different domains, demonstrates the robustness of the NAS-optimized methodology in addressing the complexities of spatiotemporal forecasting.