Air pollution prediction with clustering-based ensemble of evolving spiking neural networks and a case study for London area

Abstract

In this article, we propose a novel Clustering-based Ensemble model (CEeSNN) for air pollution prediction based on evolving Spiking Neural Networks (eSNN), where each eSNN network is trained on a separate set of time series. In our approach, we generate training sets by clustering an initial set of time series with respect to pollution values. Each obtained cluster of time series is used to build a single eSNN network. In the experiments, we forecasted ozone and PM10 pollution for Greater London Area for 1, 3, and 6 h ahead based on data from three monitoring sites located there. The prediction quality of the proposed CEeSNN model, as well as the singleton NeuCube model, an MLP network and the ARIMA model was assessed by means of several quality measures. The experimental results show that the proposed ensemble model is able to give significantly better forecasting results than the other three models.