Water quality plays a pivotal role in human health and environmental sustainability. However, traditional water quality prediction models are limited by high model complexity and long computation time, whereas AI models often struggle with high-dimensional time series and lack physical interpretability. This paper proposes a two-dimensional water quality surrogate model that couples physical numerical models and AI. The model employs physical simulation results as input, applies spectral proper orthogonal decomposition to reduce the dimensionality of the simulation results, utilizes a long short-term memory neural network for matrix forecasting, and reconstructs the two-dimensional concentration field. The simulation and predictive performance of the surrogate model were systematically evaluated through four design scenarios and three sampling dataset lengths, with a particular focus on the convection–diffusion zone and high-concentration zone. The results indicated that the model achieves high prediction accuracy for up to 7 h into the future, with sampling dataset lengths ranging from 20 to 80 h. Specifically, the model achieved an average R² of 0.92, a MAE of 0.38, and a MAPE of 1.77%, demonstrating its suitability for short-term water quality predictions. The methodology and findings of this study demonstrate the significant potential of integrating spectral proper orthogonal decomposition and deep learning for water quality prediction. By overcoming the limitations of traditional models, the proposed surrogate model provides high-accuracy predictions with enhanced physical interpretability, even in complex, dynamic environments. This work offers a practical tool for rapid responses to water pollution incidents and supports improved watershed water quality management by effectively capturing pollutant diffusion dynamics. Furthermore, the model’s scalability and adaptability make it a valuable resource for addressing intelligent management in environmental science.

Recurrent neural networks (RNNs) are widely used in reservoir operation because they can represent temporal dependencies, but their performance depends strongly on hyperparameter selection. This study compares eight evolutionary strategies—genetic algorithm (GA), particle swarm optimization (PSO), simulated annealing (SA), differential evolution (DE), ant colony optimization (ACO), artificial electric field algorithm (AEFA), slime mould algorithm (SMA), and white shark optimization (WSO)—for optimizing the hidden-node number and initial learning rate of long short-term memory (LSTM) and gated recurrent unit (GRU) models. Under common experimental settings, ACO required the shortest computation time in all reported dataset–architecture blocks, whereas the most accurate strategy depended on the dataset, architecture, and metric. For the total dataset, ACO–LSTM and AEFA–GRU produced the lowest reported MAEs (609.876 and 582.942 m³/s, respectively); WSO–LSTM was best for flood-season MAE (948.326 m³/s), and GA–GRU was best for non-flood-season MAE (516.088 m³/s). An independent chronological holdout experiment further verifies the search boundaries and validation-fitness procedure and shows that the fitness response also varies with hydrological period and RNN architecture. The results support objective- and dataset-specific optimizer selection, subject to the limitations of a single-reservoir case study and a restricted two-variable search.