Forecasting riverine total nitrogen loads using wavelet analysis and support vector regression combination model in...

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Reliable nutrient forecasting is crucial for managing non-point source pollution and controlling algal blooms.
• Traditional models struggle with the stochastic, non-linear, and non-stationary nature of riverine total nitrogen (TN) load data.

Purpose of the Study:

• To propose and evaluate a combined Wavelet Analysis-Support Vector Regression (WA-SVR) model for forecasting riverine TN loads.
• To assess the impact of different mother wavelets on the model's forecasting accuracy.

Main Methods:

• Wavelet Analysis (WA) was used to decompose TN load time series data.
• Support Vector Regression (SVR) was applied to the wavelet components to build the WA-SVR model.
• The Load Estimator model calibrated TN loads for the ChangLe River watershed (2004-2012).
• Performance was evaluated using R², NS, and MSE, comparing WA-SVR with single SVR.

Main Results:

• The WA-SVR model demonstrated superior accuracy in forecasting daily and monthly TN loads compared to the single SVR model.
• The choice of mother wavelet significantly influenced the model's efficiency, with dmey wavelet performing best.
• For daily TN loads, WA-SVR achieved R²=0.9699, NS=0.9658, and MSE=0.4885×10⁷ kg/day.
• For monthly TN loads, WA-SVR achieved R²=0.9163, NS=0.9159, and MSE=0.3237×10¹⁰ kg/month.

Conclusions:

• The combined WA-SVR method is a promising and accurate approach for forecasting riverine TN loads in agricultural watersheds.
• This hybrid model effectively addresses the complexities of riverine TN load time series data.
• The findings support the use of WA-SVR for improved water quality management and pollution control strategies.