Online Evolving Spiking Neural Networks for Incremental Air Pollution Prediction

In this article, we offer a novel model, Online evolving Spiking Neural Network for Incremental Prediction (OeSNN-IP), of forecasting from data streams, which we employ to air pollution prediction. OeSNN-IP makes predictions of pollution values and learns from a given number of recent values in streams of pollution and weather data rather than only from the most recent value from each data stream. In the proposed prediction method, older stream values have less influence on predicted values than newer ones. Moreover, we contribute to the theory of evolving spiking neural networks by offering a new fast and effective technique for encoding input values into order values of input neurons and weights of synapses linking input and output neurons. Also, we formulated the tight upper bound on the Euclidean distance between vectors of synapses weights of an output neuron and of a candidate output neuron, which simplifies the selection of a similarity threshold used in the learning phase of OeSNN-IP when a candidate output neuron is compared with output neurons. The experiments conducted on Warsaw-Ursynow pollution data show highly competitive results of the OeSNN-IP prediction as compared to the results obtained using state-of-the-art methods and algorithms.