Real-Time Prediction of Indoor Particulate Matter (PM2.5) using IoT Multi-Sensor Data and Neural Networks

Fine particulate matter (PM2.5) is one of the top causes of heart and lung issues, making indoor air quality very important for human health. If you are to handle IAQ correctly you need effective real-time prediction models that give you the insights you need in advance. To address the difficulties of indoor PM2.5 concentration forecasting, the temporal information measured across a multi-sensor Internet of Things (IoT) network has been utilized in this study. The dataset does not only contain PM2.5 levels measured in this way, but also information about the environment, such as: temperature, humidity, carbon dioxide (CO2), total volatile organic compounds (TVOC), carbon monoxide (CO) and light intensity. Several machine learning methodologies—Gradient Boosting, Ensemble, Decision Tree, Forest, and Neural Networks (NN)—were tested to determine their performance in prediction. From the above testing results, we can see that the neural network (NN) architecture is more useful than other architectures. The Average Squared Error (ASE) for independent test partition is 407.4532. The analysis by the champion model shows that Imputed CO (ppm) and Imputed Light Intensity (lux) are the most relevant predictors. That implies a high dependence on PM2.5 generation, localized human activity and pollutants from combustion. The methodology provided (established) and the tested NN model represent an effective and deployable solution for predicting indoor PM2.5 levels.