Air Quality Forecasting using LSTM RNN and Wireless Sensor Networks

In the past few decades, many urban areas around the world have suffered from severe air pollution and the health hazards that come with it, making gathering real-time air quality and air quality forecasting very important to take preventive and corrective measures. This paper proposes a scalable architecture to monitor and gather real-time air pollutant concentration data from various places and to use this data to forecast future air pollutant concentrations. Two sources are used to collect air quality data. The first being a wireless sensor network that gathers and sends pollutant concentrations to a server, with its sensor nodes deployed in various locations in Bengaluru city in South India. The second source is the real-time air quality data gathered and made available by the Government of India as a part of its Open Data initiative. Both sources provide average concentrations of various air pollutants on an hourly basis. Due to its proven track record of success with time-series data, a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) model was chosen to perform the task of air quality forecasting. This paper critically analyses the performance of the model in two regions that exhibit a significant difference in temporal variations in air quality. As these variations increase, the model suffers performance degradation necessitating adaptive modelling.