Accurate and dynamic prediction of water quality indicators is increasingly critical due to rising pollution and water resource insecurity, particularly when dealing with high-dimensional, nonlinear time series data. Dissolved oxygen (DO), a key indicator of aquatic ecosystem health and pollution, requires high prediction accuracy for effective environmental management. This study aims to enhance the accuracy and adaptability of DO prediction by addressing the limitations of traditional deep learning methods, such as slow convergence and local optima. We propose a novel hybrid optimization framework that combines Nesterov-accelerated Adaptive Moment Estimation (Nadam) with the differential evolution algorithm. A dual-population cooperation strategy and an information exchange mechanism were incorporated during the training of a long short-term memory (LSTM) network to achieve a dynamic balance between global exploration and local exploitation. This improves the model’s optimization efficiency and generalization. The research utilized a multivariate water quality time series dataset from Kaggle based on official data monitoring. Correlation analysis was conducted to ensure the scientific validity and effectiveness of the selected input variables. Experimental results demonstrated that the proposed method significantly outperforms traditional optimization strategies for DO prediction. Compared to the original Nadam optimizer, it reduced the mean squared prediction error by 47.8%, exhibiting enhanced adaptability and robustness in complex pollution scenarios. This study presents an effective optimization strategy to improve LSTM performance in water quality forecasting, along with a scalable and interpretable intelligent analysis framework. It provides both theoretical and practical support for water quality forecasting, early warning systems, and intelligent environmental monitoring.