AP-LSSVM modeling for water quality prediction

This paper addresses the problem of water quality predicting based on spectrometry. Spectrometry is a kind of novel, quickly, and green soft measurement technology for predicting water quality such as Total Organic Carbon (TOC) criterion. However the analysis accuracy and robustness of predicting model are greatly affected by training samples in modeling process. For solving such a problem, a suitable and effective clustering method is used to improve the model accuracy as well as the computing process time. Firstly, we propose affinity propagation (AP) clustering method with vector angle cosine similarity based on spectral data of water aiming to choose good training samplers. With the most suitable clusters after AP clustering process, a nonlinear modeling method based on a least squares support vector machine (LSSVM) is then given to evaluate TOCs of water samples. Finally, 100 water samples experiment is applied to the regression mode to assess the effectiveness of AP-LSSVM model. The results indicate that the effectiveness and robustness of our proposed model are better than the single LSSVM model and also superior to the model based on k-means clustering.