A comparative study of statistical learning methods to predict eutriphication tendency in a reservoir, northeast China

In order to early warn the alga bloom in a reservoir (drinking water source) in northeast China, we carried out a comparative investigation on three statistical machine learning methods to construct suitable one-step weekly Chlorophyll-a (Chla) prediction models: multiple linear regression (MLR), back propagation artificial neural network (BPANN) and support vector regression machine (SVR). Previously, five major environmental factors were selected as potential input variables of prediction models by correlation analysis and principal component analysis. The model training and validation point to that a low number of input variables are able to predict the Chla trends well by BPANN and SVR methods while MLR responses unacceptably. Additionally, the SVR machine presents the best performance in light of root mean square error and generalization ability. The sensitivity analysis indicates that the Chla in the next week is sensitive to the changes of Chla, water temperature. In conclusion, we choose SVR machine to be the most suitable model in this case.