Supplementary file 1_Enhancing dissolved oxygen prediction in lake-reservoirs via a hybrid BO+SSA-driven backpropagation neural network.docx

With the self-purification ability of lake-reservoir water body gradually weakened and the oscillation of dissolved oxygen (DO) concentration intensifying, the high-precision prediction of lake-reservoir DO is important to the aquatic ecological safety. Aiming at the key problem that the prediction precision is low, the model structure and hyperparameters of back propagation neural network (BPNN) are highly sensitive, and the global convergence is poor with high tendency to fall into local optima in traditional DO prediction. In this study, a new hybrid optimization technology called Bayesian Optimization (BO) + improved Sparrow Search Algorithm (SSA), named BO+SSA, is employed to optimize the hyperparameters of BPNN and search initial weights and thresholds to overcome such a problem. Chaotic initialization, adaptive weight adjustment, and dynamic search strategies are integrated to enhance global optimization capability and accelerate convergence of BPNN. Four representative monitoring sections (including Baiheshan and Luojiang) from lakes and reservoirs in the eastern Sichuan Basin, China, were selected for analysis. Based on correlation analysis and feature importance assessment, pH, water temperature (WT), air temperature (AT), and atmospheric pressure (AP) were identified as input variables for testing the predictive performance of the BO+SSA-BPNN model. The coefficient of determination (R2) for the test set ranged from 0.861 to 0.939. Furthermore, the improved BPNN model demonstrated a reduction of 30%−61% in Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) compared to the original BPNN model. The result proves that the method of hybrid optimization of BO+SSA can better solve the problems of complex nonlinear relationship modeling and provide an efficient BPNN-based DO prediction model that can be applied to lake-reservoir dynamic monitoring and management.