A Hybrid Approach to Forecasting Water Quality in Urban Drainage Systems

Abstract

Abstract The pollutant monitoring in urban sewer systems is currently done by captors based on UV-Vis spectrometry. These have the purpose of determining the dynamics of water quality indicator concentrations. Analyzing time series and UV-Vis absorbance spectra is necessary to develop forecast methodologies as supporting tools for online and real-time control support. This work presents a hybrid approach based on seven methodologies combined with Principal Component Analysis (PCA) for UV-Vis absorbance time series forecasting, such as: Principal Component Analysis combined with Discrete Fourier Transform (PCA/DFT), PCA combined with Chebyshev polynomials, PCA combined with Legendre polynomials (PCA/L-Poly), PCA combined with Feed-forward Artificial Neural Networks (PCA/ANN), PCA combined with Polynomial (PCA/PolyReg), PCA combined with SVM (PCA/SVM) and Clustering process combined with Markov chains (kmMC). Three UV-Vis absorbances time series collected online were used at different experimental sites in Colombia, two in Bogotá and one in Medellin. The Mean Absolute Percentage Error (MAPE) values were obtained between 0% and 57% for all the study sites. Results show that it is impossible to determine the best forecasting methodology among the proposed ones because they would complement each other for different forecasting time steps and spectra ranges according to the target water quality.