Indoor Air Quality Prediction Using Artificial Neural Networks

To investigate variations in indoor NO₂ concentrations under a subtropical climate, their relationships with indoor and outdoor environment parameters, and ways to improve prediction accuracy, this study monitored indoor and outdoor air quality in a teaching building from May 18 to May 31, 2025. Measurements included NO₂, PM2.5, PM10, O₃, temperature, and relative humidity. Pearson correlation coefficients were calculated, and variables with |r| ≥ 0.3 were selected as model inputs. Multiple linear regression (MLR) and artificial neural network (ANN) models were developed and evaluated using R² and RMSE. Based on validation results, the preferred ANN used two hidden layers, with 150 and 75 neurons, and ReLU activation. On the test set, MLR achieved an R² of 0.5975 and an RMSE of 4.8597, whereas ANN achieved an R² of 0.8528 and an RMSE of 2.9389. Compared with MLR, ANN increased R² by approximately 25.53 percentage points and reduced RMSE by approximately 39.52%. ANN also reduced the underestimation observed at high concentrations and improved accuracy across concentration ranges. These findings show that indoor NO₂ depends on multiple indoor and outdoor factors and that ANN captures their complex relationships more effectively than MLR. The results support ANN-based indoor NO₂ prediction under subtropical conditions and may inform air-quality monitoring and forecasting in similar buildings. However, the study was limited by its short monitoring period, incomplete coverage of relevant factors, and limited neural-network optimization. Future studies should address these limitations.