Prediction of water quality time series based on the dynamic sliding window BP neural network model

In order to improve the prediction precision of water quality having time series properties by BP neural network (BPNN), principal component analysis (PCA) was used for characteristic extraction and dimension reduction of the original data. Concentrations of dissolved organic compound (DOC) and total nitrogen (TN), and turbidity were selected as the water quality prediction indices, a three-layer BPNN model was established for prediction, and the prediction performance was analyzed. The results showed that the optimal training-set sizes of concentrations of DOC and TN, and turbidity were 60, 60, and 90 days, while the best BPNN topological structures were 9-12-1, 8-6-1 and 7-13-1, respectively. The optimized BPNN model exhibited favorable prediction performance on the variation trends of concentrations of DOC and TN, and turbidity. In contrast, the prediction performance of DOC by BPNN model was significantly better than that of TN and turbidity, with RMSE, MAPE, and R values of 0.040, 0.66% and 0.867, respectively. This model had a good applicability and precision for prediction of surface water qualities having non-linear characteristics.