Environ Monit Assess. 2026 Jan 3;198(1):78. doi: 10.1007/s10661-025-14929-z.
ABSTRACT
Artificial neural networks (ANNs) are widely applied in air quality modelling because they can capture nonlinear interactions among pollutants and support reliable air pollutant index (API) forecasting. This study aims to identify the pollutants that most strongly influence API variability and to evaluate the performance of two merged hybrid ANN models for forecasting API in the northern region of Peninsular Malaysia. Two hybrid frameworks were developed: an ANN integrated with sensitivity analysis (ANN-SAM) to identify influential pollutants and an ANN combined with principal component analysis (ANN-PCA) to reduce dimensionality while retaining key information. Sensitivity analysis identified O3, SO2, PM10, and PM2.5 as the most influential pollutants, whereas PCA retained all variables except SO2. These selected inputs were used to develop the MLP-FF-ANN-SAM and MLP-FF-ANN-PCA models. Both models achieved strong predictive performance, with R2 values ranging from 0.821 to 0.826 and RMSE values between 5.922 and 5.982. The slight improvement after removing NO2 indicates that it contributes limited independent predictive value due to its collinearity with particulate matter. Seasonal increases in PM10 and PM2.5 during haze periods further highlight the influence of regional transboundary pollution. Using 5 years of multi-station data, this study demonstrates that merged ANN-SAM and ANN-PCA frameworks can provide accurate, efficient, and interpretable API forecasts. These findings support the development of simplified and computationally efficient tools for operational air quality assessment and early-warning applications in Malaysia.
PMID:41483234 | DOI:10.1007/s10661-025-14929-z
