In the present study, the performance of three hybrid Support Vector Regression (SVR) models optimized with metaheuristic algorithms—Wavelet (W-SVR), the Innovative Gunner Algorithm (AIG-SVR), and Black Widow Optimization (BWO-SVR)—was evaluated for the simultaneous estimation of BOD and COD in the Khorramabad River. Monthly data on 14 qualitative and hydrological parameters were collected from hydrometric and water-quality monitoring stations over a 20-year period (2005–2025) and were divided into training and testing sets using an 80:20 ratio. The results indicated that the input scenario comprising all parameters yielded the best performance. Moreover, the evaluation criteria demonstrated that, during the testing phase, the W-SVR model with the Radial Basis Function (RBF) kernel outperformed the other models, achieving correlation coefficients (R) of 0.978 and 0.982, Nash–Sutcliffe efficiency (NSE) values of 0.980 and 0.985, root mean square errors (RMSE) of 0.681 and 0.617 mg/L, and mean absolute errors (MAE) of 0.506 and 0.402 mg/L for BOD and COD, respectively. The AIG-SVR and BWO-SVR models ranked second and third, respectively. Box plot and Taylor diagram analyses further confirmed the superiority of the W-SVR model in reproducing the statistical distribution of the observed data. The findings demonstrate that the wavelet-based mutation operator, by establishing an appropriate balance between exploration and exploitation in optimizing the SVR hyperparameters, significantly enhances prediction accuracy and provides an efficient, cost-effective tool for water quality monitoring in semi-arid regions.