Sensor-Based Air Pollution Prediction Using Deep CNN-LSTM

The devastating impacts of air pollution have be-come more and more evident in recent years. As our measurement technologies improve, we gain better insight into the true impact of this deadly, yet often ignored, threat. The first step in reducing the damages caused by this problem is being able to analyze and predict its patterns. The problem of predicting air quality and the presence of particulate matter lies in the nature of the data needed to create an accurate system. The sheer number of factors affecting air quality mean that previously proposed approaches often utilize a great many sources of data, aiming to incorporate images, wind graphs, traffic information, and more. Yet in truth, most areas outside large metropolises lack ready access to high-quality data, preventing them from ever implementing an effective system. We propose a system utilizing a 1-D deep convolutional neural network to analyze past sensor readings and predict air pollutant concentrations up to a day in the future at a 3-hour resolution. We specifically developed this model for predicting PM2.5 values. The system receives PM2.5 sensor values and discovers temporal pattern in the data, which will be later used for prediction. By removing the dependency on complex data inputs, the system becomes accesible and easily implementable for any region. Despite this simplified approach, the results are comparable to — and often better than — any current state-of-the-art predictive systems in this domain.