Clustering approach to urban rainfall time series prediction with support vector regression model

Abstract

ABSTRACT The aim of rainfall prediction in urban areas is to provide information about precipitation, which may cause flooding resulting from the insufficient capacity of sewage systems. The research combined cluster analysis and support vector regression (CSVR) to create a simpler, and as effective as hybrid models, method predicting daily rainfall in Wroclaw, Poland. Results demonstrate that the clustering approach while modelling is reasonable, improves the quality of prediction and minimises error values. The best SVR performance, RMSE = 2.492 mm and r2 = 0.830 in the testing subset, was obtained for the Ward clustering method. Nevertheless, comparisons with MLP prediction, combined with k-means clustering, proved to be slightly more accurate and led to the creation of a model with RMSE = 1.974 and r2 = 0.899. The presented approach might constitute an alternative method to be used for creating rainfall patterns in stormwater discharge or management systems leading to an increase in flash flood safety in an urban catchment.