Deep Temporal Models for AQI Forecasting Using Multivariate Time-Series Data: A Unified Evaluation of Attention, Recurrence and Convolution Methods at IOEGC

I co-authored the paper “Deep Temporal Models for AQI Forecasting Using Multivariate Time-Series Data: A Unified Evaluation of Attention, Recurrence and Convolution Methods,” presented at IOEGC. The study addresses the critical need for accurate one-hour-ahead Air Quality Index prediction for urban environmental monitoring and public health preparedness. Using 43,062 hourly samples from Kathmandu, we developed a unified evaluation framework comparing Transformer Encoder (attention), Stacked GRU (recurrence), and 1D-CNN (convolution) under a shared preprocessing and feature engineering pipeline. The Stacked GRU achieved the best performance with a test R² of 0.9487. This work was carried out collaboratively between Pulchowk Campus, IOE (time-series modeling and environmental analytics) and the Artificial Intelligence & Analytics Lab at Marichi Tech AI Pvt. Ltd. (AI architecture design and experimentation), providing practical insights for building robust real-time AQI forecasting systems.