Jump to Main Content
Forecasting riverine total nitrogen loads using wavelet analysis and support vector regression combination model in an agricultural watershed
- Ji, Xiaoliang, Lu, Jun
- Environmental science and pollution research international 2018 v.25 no.26 pp. 26405-26422
- agricultural watersheds, algal blooms, artificial intelligence, models, nitrogen, nitrogen content, nonpoint source pollution, pollution load, rivers, time series analysis, wavelet, China
- In the context of non-point source pollution management and algal blooms control, the reliable nutrient forecasting is of critical importance. Considering the highly stochastic, non-linear, and non-stationary natures involved in riverine total nitrogen (TN) load time series data, some traditional statistical and artificial intelligence models are inherently unable to give accurate nutrient forecasts due to their mechanism and structure characteristics. In this study, based on the wavelet analysis (WA) and support vector regression (SVR), a promising combined WA-SVR model was proposed for forecasting riverine TN loads. The data pro-processing tool WA was employed to decompose the time series data of riverine TN load for revealing its dominator. Subsequently, all wavelet components were used as inputs to SVR for WA-SVR model. The continuous riverine TN loads during 2004–2012 in the ChangLe River watershed of eastern China were estimated by using a calibrated Load Estimator model. Performance criteria, namely, determination coefficient (R²), Nash-Sutcliffe model efficiency (NS), and mean square error (MSE) were applied to assess the performance of the developed models. The effects of different mother wavelets on the efficiency of the conjunction model were investigated. The results demonstrated that the mother wavelet played a crucial role for the successful implementation of the WA-SVR model. Among the 23 selected mother wavelet functions, dmey wavelet performed best in forecasting the daily and monthly TN loads. Furthermore, the performance of the optimal WA-SVR model was compared with that of single SVR model without wavelet decomposition. The comparison indicated that the hybrid model provided better accuracy than that of single SVR model. For daily riverine TN loads, the R², NS, and MSE values of WA-SVR model during the test stage were 0.9699, 0.9658, and 0.4885 × 10⁷ kg/day, respectively. For monthly riverine TN loads, the R², NS, and MSE values of the model during the test stage were 0.9163, 0.9159, and 0.3237 × 10¹⁰ kg/month, respectively. The overall results strongly suggested that the combined WA-SVR method can successfully forecast riverine TN loads in agricultural watersheds.