Quantitative Chemical Oxygen Demand Prediction Model based on Improved African Vulture Optimization Algorithm

Abstract

UV-Vis spectroscopy is crucial for online COD detection in surface water monitoring. But in actual water detection, turbidity and other interfering factors create a complex non - linear relationship between UV-Vis data and COD concentration, leading to unstable and inaccurate detection. To improve prediction accuracy and robustness, this study proposes the African Vulture Optimization with Adaptive Cooperation and Opposition learning (AACO) algorithm to optimize back-propagation (BP) neural network parameters for more accurate and stable COD prediction. AVOA enhances BP network structural parameter optimization, while dynamic opposition learning (DOL), adaptive search strategy (ASS), and fitness distance balance selection (FDB) further improve robustness and detection accuracy. Partial least squares (PLS) is used to analyze the characteristic wavelengths of COD and turbidity , followed by construction of a multi-wavelength AACO-BP quantitative prediction model. Experimental results show that the AACO-BP model outperforms AO, AVOA, PSO, and traditional BP neural networks, achieving R² = 0.9960, RMSE = 0.7216 mg/L. In addition, twelve real surface water samples collected from representative water bodies in Changchun City were used for preliminary validation. The AACO-BP model achieved R² = 0.9170, RMSE = 0.1438 mg/L. These results demonstrate that the proposed model has good prediction accuracy and stability under turbidity interference and shows potential for COD prediction in real surface water monitoring.