Integrating quantile regression with ARIMA and ANN for interpretable and accurate PM2.5 forecasting in Hat Yai, Thailand

Accurate forecasting of PM2.5 concentrations is crucial for effective air quality management and the protection of public health. This study proposes a novel hybrid model that integrates Autoregressive Integrated Moving Average (ARIMA), Artificial Neural Network (ANN), and quantile regression to enhance forecasting accuracy and robustness. The model is evaluated using daily PM2.5 concentration data from Station 44T in Hat Yai, Songkhla, Thailand, collected via the Air4Thai platform between January 1, 2020, and December 31, 2024. After handling missing values through data cleaning procedures, 1809 observations were retained and split into training (1447 days) and testing (362 days) sets. Several models were developed and evaluated, including a baseline ARIMA(1,1,2), a standalone ANN, and hybrid models integrating ARIMA and ANN, with an additional model incorporating quantile regression. Results showed that, while the ARIMA model demonstrated strong interpretability and the ANN struggled with linear dependencies, the hybrid ARIMA-ANN models showed marked improvements in predictive performance. The proposed ARIMA-ANN-QREG model achieved the best results, with the lowest MAE (1.704), MAPE (11.782%), and minimal bias (MFB = -0.0004), even under extreme PM2.5 conditions. By combining linear, nonlinear, and distributional modeling, the proposed approach offers a computationally efficient yet highly interpretable alternative to deep learning architectures such as ConvLSTM. These findings demonstrate that ARIMA-ANN-QREG is a robust and practical forecasting tool for real-world air quality management, with direct relevance for policy-making and early-warning systems.