Abstract: Machine learning and other related technologies find increasing applications to extracting manual operation experiences from massive data in the practice of reservoir regulation. However, reservoir operation schemes solely based on machine learning fail to describe reservoir operation with enough accuracy, resulting in outliers in calculation results and a lack of operational experience. This paper constructs a deep learning model guided by the physical mechanism for reservoir operation, taking the water balance constraint, monotonicity constraint, and boundary constraint as the penalty terms of loss functions; data enhancement is used to include the factor of rare flood operations in the data sets of model training and verification. Results show this model is effective in simulating reservoir decisions for conventional operations and rare flood operations. It better satisfies the water balance equation, reduces negative flows effectively, and improves high flow simulation accuracies in comparison with the benchmark model, thus promoting the realization of intelligent reservoir operation.

Key words: reservoir operation, intelligent operation rules, physical mechanism, deep learning, rare flood