Forecasting Air Quality Using Massive-Scale WSN Based on Convolutional LSTM Network

PM2.5 is ultra-light micro-grained particles and dangerous air pollution that threatens public health. A real-time wireless sensor network (WSN) that measure air pollution is a solution to increase public awareness about the long-term impact of PM2.5 exposure. However, in a massive-scale WSN-based air pollution monitoring system, there are numerous noisy and low-concentration periods in the raw PM2.5 dataset, which may lead to unreliable causality predictions. This paper addresses the problem of optimizing sensor acquisition of a wireless sensor network to reconstruct and predict spatiotemporal PM concentrations data using ConvLSTM network. This prediction model is built by combining convolution network and long short-term memory network. The dataset is gathered from air quality WSN s that are already installed across Taiwan. Using the last 48-hour records, the next hour PM2.5 concentration is predicted. RMSE is used to evaluate the prediction accuracy. The results reveal that the ConvLSTM network achieves better performance than those using the LSTM network and regression analysis with RMSE of 1.31, 2.59, and 16.34, respectively.