[Objective] The time series of reservoir water quality indices,especially dissolved oxygen content, exhibit strong nonlinearity,high complexity,and uncertainty,which lead to insufficient accuracy of single prediction models.This study aims to construct a high-precision hybrid prediction model that integrates time series decomposition,intelligent optimization,and residual correction,thereby significantly improving the prediction accuracy of dissolved oxygen (DO) content and providing reliable support for water environment management and pollution early warning. [Methods] The core procedures of the proposed hybrid prediction model are as follows. 1) Data decomposition and reconstruction. Singular spectrum analysis (SSA) was applied to decompose the dissolved oxygen time series, and the series was reconstructed into trend components, periodic components, and residual components to reduce sequence complexity and highlight features at different frequencies. 2) An improved dung beetle optimizer (IDBO) which integrates piecewise chaotic mapping and opposition-based learning strategies was designed to enhance population diversity and initialization quality. The improved IDBO was used to optimize key hyperparameters of the GRU network, including the number of hidden layer neurons and the initial learning rate. 3) Component prediction and residual correction. The GRU model optimized by IDBO was used to predict the trend component and periodic components separately. A residual series prediction difference correction method (DCM) was proposed. The residual component was first predicted using GRU, and the difference sequence between the predicted values and the observed values was calculated. Variational mode decomposition (VMD) was then applied to the difference sequence to fully extract high-frequency detail information. Each decomposed component was predicted using GRU and aggregated to obtain the predicted difference values. Finally, the predicted differences were compensated into the initial residual prediction to obtain the corrected residual prediction results. 4) Model integration and validation. The prediction results of the three components were aggregated to obtain the final DO prediction values. Measured dissolved oxygen data from Daheiting Reservoir in Tangshan, Hebei Province were used for experiments. The dataset contained 2352 records with a sampling interval of four hours. Root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), and the coefficient of determination (R²) were used as evaluation metrics. The proposed model was compared with GRU, SSA-GRU, SSA-DBO-GRU, SSA-IDBO-GRU, and models reported in the literature such as LSTM and PSO-GRU. [Results] The proposed SSA-IDBO-GRU-DCM hybrid model achieved the best performance among all comparative models. The prediction errors were significantly reduced, with an RMSE of 0.580 2 mg/L, an MAE of 0.329 2 mg/L, an MRE of 0.0269, and an R² of 0.918 8. Ablation experiments confirmed that the proposed IDBO improvement strategies effectively enhanced hyperparameter optimization accuracy. The residual difference correction method (DCM) significantly improved the prediction performance of the residual component and was the key factor contributing to the overall accuracy improvement. These results fully demonstrated the effectiveness and superiority of the “decomposition-optimization-correction” framework. [Conclusion] SSA effectively decouples the complex characteristics of water quality time series. IDBO efficiently and accurately optimizes GRU hyperparameters. The proposed VMD-GRU-based residual difference correction method (DCM) is the key innovation for improving overall prediction accuracy. The proposed model significantly improves the prediction accuracy of dissolved oxygen content and provides an efficient and reliable new approach for reservoir dissolved oxygen prediction. Future work can extend this framework to the prediction of other key water quality parameters such as ammonia nitrogen and total phosphorus, and further explore the integration of natural evolutionary strategies to improve computational efficiency and generalization ability.