Abstract: To enhance the scientific rigor and intelligence of operation for complex water resource system, this study proposes an integrated framework combining multi-objective optimization with deep learning for extracting reservoir operation rules, using the Longyangxia Reservoir in the upper reaches of Yellow River as a case study. Based on monthly-scale data from 1956 to 2023, a dual-objective optimization operation model was first constructed to maximize power generation while minimizing water supply deficits, with Pareto-optimal solutions obtained using the NSGA-Ⅱ algorithm. Subsequently, LSTM, CNN, CNN-LSTM, and RF models were employed to extract and fit operation rules for optimal discharge sequences. The study evaluates rule extraction performance from operation objectives and model accuracy two dimensions. Results demonstrate that CNN-LSTM consistently outperforms other models across all three operation objectives, achieving fitting accuracy of 0.94 (power generation), 0.92 (water supply), and 0.95 (trade-off), respectively. This approach provides an effective method for extracting multi-objective reservoir operation rules while offering reference for model selection.