Optimized SVR model for predicting dissolved oxygen levels using wavelet denoising and variable reduction: Taking the Minjiang River estuary as an example
Peng Zhang, Xinyang Liu, Huiru Zhang, Chengchun Shi, Gangfu Song, Lei Tang, Ruihua Li
Abstract
Adequate dissolved oxygen (DO) is critical for the maintenance of aquatic ecosystems. However, predicting DO levels in regions with complex hydrological variations remains challenging. This study presents a novel DO prediction model using the Minjiang River estuary as an example by integrating advanced machine learning techniques. Key influencing factors were identified using the Maximum Information Coefficient (MIC) and noise was reduced using Wavelet Denoising (WD). Support Vector Regression (SVR) parameters were optimized using Particle Swarm Optimization (PSO), culminating in an optimized WD-MIC-PSO-SVR model for DO prediction. The results showed that the MIC effectively identified the key influencing factors of DO. Compared with the unoptimized SVR model, the proposed model achieved higher accuracy, R 2 and NSE reached 0.91 and 0.83, respectively, while the MAE and RMSE values were reduced by 67 % and 44 %, respectively, affirming its applicability for real-time DO prediction. This study contributes to water environment protection by providing an effective solution for DO modeling in regions with substantial hydrological changes. The integrated WD-MIC data processing method shows promising potential in reducing model errors and lowering water monitoring costs by focusing on highly correlated variables. • The study effectively identified the driving variables of DO such as WT, NH 3 , and PG through the application of MIC. • The model's predictive accuracy was significantly enhanced by employing a combined approach using WD and MIC. • The composite machine learning model, WD-MIC-PSO-SVR, successfully predicts DO in water bodies characterized by diverse hydrological conditions and varying levels of pollution. • The WD-MIC-PSO-SVR, can provide early warnings of hypoxic conditions with exceptional precision at 1-h, 12-h, and 24-h lag times.