TRAINING RECURRENT NEURAL NETWORKS FOR PARTICULATE MATTER CONCENTRATION PREDICTION
Keywords: recurrent neural networks, long short-term memory, gated recurrent unit, optimizers, Particulate Matter pollution, prediction, air quality monitoring, sensor network
Abstract. A high level of particulate matter in the atmosphere has an adverse long-term effect on human health. It has been associated with increased pulmonary tract and lung infections. It is more common in urban areas, especially megacities due to the confluence of industries and motorized machinery. Considering that most of the world’s population lives in urban areas, there is a need to monitor air pollution arising from particulate matter in order to ensure clean and safe air in cities in accordance with goal 11 of the Sustainable Development Goals. One way of doing this is through the use of Recurrent Neural Networks (RNN), which are suited for time varying data. Particulate matter concentration recorded by a network of low-cost sensors in Stuttgart is trained on three of the most popular RNN variants: Standard LSTM, Peephole LSTM and Gated Recurrent Unit. Two optimizers are used, Stochastic Gradient descent and Adam. Training is done on a single sensor and the optimum weights transferred and used in the prediction of other sensor values. This study concludes that Gated Recurrent Unit with Stochastic Gradient Descent is the most effective of the three variants in predicting particulate matter PM2.5 concentrations. In addition to this, weight transfer between sensors is not affected by temperature, wind direction, wind speed and geographic distance between sensors but rather by atmospheric pressure and the similarity of recorded Particulate matter levels.