Exploring Deep Learning Architectures for Localised Hourly Air Quality Prediction

Abstract

Air pollution is a global environmental and public health issue, but it is at the local scale that many mitigation measures are implemented. In a human context, we propose that as a decision support tool it is more valuable to provide hourly forecasts at local scales with the following considerations: (1) the system should be designed for rapid and simple human-tuning of different trade-offs; (2) the chosen model and hyper-parameters should maximise consistency of learning given the likelihood of regular retraining with new data; (3) reducing errors when predicting low pollutant values is far less important than accurate prediction of spikes. Target users include local officials deciding whether to enact short-term plans for meeting regulatory objectives or citizens, deciding to change their behaviour or travel patterns to avoid likely exposure. Both groups will also wish to reduce inconvenience and disruption, but the relative importance they will place on these two conflicting factors cannot be pre-determined and hence there is desirability for rapid exploration and tuning of false vs m.sised alarm trade-offs.

Through a series of experiments, we show how Deep Neural architectures can be developed to create an ‘early warning’ decision support tool, with the ability to personalise the accuracy trade-offs at different time-steps from predicting the possibility of a spike 24 h in advance, to increasingly accurate confirmations that the spike will take place. The results also show that we can significantly improve the prediction accuracy if we could include meteorological prediction values as additional input to the models.