Energy-Efficient PM2.5-Based Air-Quality Forecasting Using Bayesian-Optimized CNNs on Embedded Devices
Abstract
Real-time PM<sub>2.5</sub> forecasting on microcontroller-class devices is challenging because conventional learning models often require high memory, computation, and power. This article proposes a Bayesian-optimized convolutional neural network (BOCNN) for lightweight PM<sub>2.5</sub>-based air-quality forecasting on embedded devices. The proposed approach uses Bayesian optimization to tune key CNN hyperparameters, including convolutional filters, kernel sizes, dense units, dropout rates, and learning rates, and then converts the optimized model to TensorFlow Lite for TinyML deployment. The model was evaluated on the Delhi 2024 CPCB air-quality dataset and compared with MLP, FFNN, CNN, LSTM, GRU, BiLSTM, Random forest (RF), Support Vector Regression (SVR), and XGBoost baselines. The proposed BOCNN achieved MAE, MSE, RMSE, and <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> values of 0.011093, 0.000317, 0.017815, and 0.985727, respectively. On the original PM<sub>2.5</sub> scale, it obtained an MAE of 9.012691 and an RMSE of 14.473820. Compared with the standard CNN baseline, the BOCNN reduced RMSE by approximately 40.73% and improved <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> from 0.959377 to 0.985727. For embedded deployment, the dynamically quantized TensorFlow Lite model achieved a compact size of 12.19 kB, 4344 bytes of runtime memory, 11.40 ms MCU inference latency, and a 40.45% lower power consumption than the CNN baseline. These results show that BOCNN provides an effective accuracy-efficiency tradeoff for low-power PM<sub>2.5</sub> forecasting in IoT-enabled smart-city environments.