Deep Learning for PM2.5 Forecasting: An Empirical Comparison of LSTM, GRU, and N-BEATS

Accurate predictions of fine particulate matter (PM2.5) levels will significantly contribute toward enhancing public health and provide valuable information to policymakers regarding how to create and implement policies that protect the environment. The application of deep-learning techniques has produced promising results in the area of PM2.5 forecasting; however, few comparative studies using the same dataset and evaluation criteria have been conducted. This study evaluates the performance of three of the most commonly used sequence-oriented architectures (i.e., Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Neural Basis Expansion Analysis for Time Series (N-BEATS)) to predict short-term PM2.5 levels. Each model was trained and tested using the same 70/15/15 data-split ratio and the same four input window sizes (i.e., 5, 10, 15, and 20 hours). Generally, the use of longer input windows improves accuracy; however, the rate of improvement is much reduced beyond 15–20 time steps. N-BEATS produced higher accuracy than both recurrent models over each of the evaluation metrics and input window sizes; specifically, at the 20-hour time step, N-BEATS achieved a root mean squared error (RMSE) of 55.28 µg/m³ and an R² value of 0.5116, which represents a reduction in RMSE of approximately 14 % compared to the GRU while providing improved tracking of hazardous pollution events. N-BEATS also trained at a significantly faster pace and experienced very little increase in computational costs as input window lengths increased; whereas the training times of both LSTM and GRU were essentially tripled. N-BEATS’ feed-forward architecture and block-based design lend themselves to be executed in parallel, and the inherent basis-function decomposition within N-BEATS is more conducive to modelling the sharp, non-linear fluctuations typically present in pollution datasets. Overall, the findings presented in this study demonstrate that N-BEATS is a viable and effective tool for near-real-time air-quality forecasting applications and emphasize the need for standardization of benchmarking in the field of atmospheric time-series analysis.