EcoCast: Multi-Task Global Temperature Forecasting via Autoregressive Transformers

Rapid changes in global temperatures and greenhouse gas emissions make accurate forecasting critically essential to saving peoples lives from natural disasters. The urgency of advanced predictive models in environmental science cannot be emphasized more by the vital need for adaptation and mitigation efforts. This study uses a customized transformer-based model, exploring its effectiveness and efficiency in forecasting environmental factors in the format of time series data. This project compares single-shot and autoregressive models in single-task and multi-task learning processes using the NOAAGlobalTemp Dataset by the National Centers for Environmental Information and the CO2 Emissions Dataset curated by Our World in Data. The results show that autoregressive models significantly outperform single-shot models via iterative prediction. In addition, multi-task learning further improves prediction accuracy and reliability by capturing the correlation and causality between global temperature anomalies and greenhouse gas emissions. Compared to a baseline model, the models superiority is evident. This research explored the optimal transformer-based model architecture for multi-task environmental forecasting, which provides a benchmark for future research in the community. Furthermore, this projects outcome (the model) will be deployed publicly as a web application for users to access.