Real Time Air Pollutant Forecast Using Ensemble Learning

Indoor air pollution poses a hidden threat affecting our health and well-being. Evaluating the indoor environment's air quality is necessary to minimize the risk caused by air pollution. Previous research has explored machine learning (ML) and deep learning (DL) techniques like ARIMA, Prophet, and Multi-Layer Perceptron (MLP) for pollution prediction using pre-recorded data, leaving real-time forecasting unaddressed. This paper compares ARIMA, Prophet, MLP, and Staked models, combining all three models to evaluate the best model for forecasting air quality with collected datasets and real-time data and integrating the Internet of Things device with the Raspberry Pi Pico and MQ 135 sensor. Using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R2 Score, and time taken to train and forecast, evaluate models to determine whether they are suitable for live predictions. The result shows that the MLP model managed to get the best evaluation score on the available dataset at 2.5 MAE, 8.7 MSE, 2.95 RMSE, and 0.98 R2 Score. Despite the Prophet model shows an inferior metrics score compared to the MLP, it is considered suitable for live predictions because of its significantly faster training time of 0.08 seconds and forecasting time of 0.05 seconds. The Prophet model also predicts its lower bound and upper bound, enhancing its generalization capability. The prophet model generally predicts within its prediction bounds with 75% accuracy on the available dataset and 85% on the live dataset.