并联水库群调度规则提取的机理神经网络研究
Research on Mechanism Neural Network for Extracting Operation Rules of Parallel Reservoirs

本文提出了基于理论指导的深度学习方法，提取并联供水水库联合调度规则，该方法以损失函数惩罚项的形式，将成员水库蓄水期弃水概率相等且供水期达到死库容状态概率相等、蓄水量边界约束和供水量边界约束加入到机器学习模型中。结果表明，在纯机器学习模型中耦合适当的调度机理知识可以提高其模拟水库调度规则的精度。在三种约束中，成员水库蓄水期弃水概率相等且供水期达到死库容状态概率相等对模拟结果的改善效果最佳。此外，同时耦合三种约束的场景模拟结果最优。
In this paper, a theory-guided deep learning method is proposed for extracting the joint scheduling rules of parallel water supply reservoirs, which incorporates the equal probability of abandonment of the member reservoirs during the storage period, the equal probability of reaching the dead capacity state during the supply period, and the boundary constraints on the water supply into the machine learning model in the form of a penalty term of the loss function. The results show that coupling an appropriate amount of scheduling mechanism knowledge in the machine learning model can improve its accuracy in simulating reservoir scheduling rules. Among the three constraints coupled, equal probability of abandonment of member reservoirs during the impoundment period has the best improvement effect on the simulation results. In addition, the scenario that is coupled with the three constraints at the same time has the best simulation results.