Zero-Shot Inference with TimeGPT for Air Quality Prediction in Bangladesh

Air quality represents an important environmental challenge, particularly in densely populated urban areas, where the presence of pollutants like Particulate Matter (PM) considerably threatens public health. This study aims to develop predictive models for PM2.5 concentrations utilizing data sourced from the U.S. Embassy air quality monitor spanning from January 1, 2018, to September 17, 2024. A robust methodology was employed, encompassing two Deep Learning (DL) models, Long Short-Term Memory (LSTM) and TimeGPT. The results demonstrate that TimeGPT consistently outperforms LSTM across all key performance metrics. The average Mean Absolute Error (MAE) for TimeGPT was determined to be 10.76, which is notably lower than the average MAE of 15.89 observed for LSTM. Furthermore, TimeGPT exhibited an average Mean Absolute Percentage Error (MAPE) of 4.68, in contrast to LSTM’s average MAPE of 9.84. The Root Mean Squared Error (RMSE) values corroborated this trend, with TimeGPT yielding an RMSE of 13.58, compared to 17.76 for LSTM. TimeGPT achieved its optimal MAE of 3.81 when it was trained using data from 2018 to 2024. These results emphasize the prospects of modern Machine Learning (ML) methodologies in the accurate forecasting of air quality.