| Reservoir flood control operation is a complex multi-objective optimization problem,which has the characteristics of high dimensionality of decision variables and chain correlation between different decision dimensions.Therefore,there are many difficults such as excessive search space and strong correlation of decision variables when using evolutionary multiobjective optimization algorithm to solve such problems,which leads to slow convergence,low solution efficiency,and even evolution failure.On the other hand,because of the optimization algorithm requires the complete process of the flood to solve optimization problem,it is often used as an offline solution.The online scheduling mainly uses the scheduling rules for decision-making,but the scheduling effect is not satisfactory.In order to overcome the problem of the efficiency of the evolutionary algorithm,the idea of using the solutions produced by similar tasks or the same kind of tasks in history is proposed.In order to use the optimized algorithm for online scheduling and improve the scheduling effect.Starting from a data-driven approach,this paper attempts to dig out effective scheduling knowledge from historical scheduling data and migrate it to new problems to speed up the solution process.The main research and innovation points of this article are as follows:(1)Propose a method of structural flood based on stratified sampling.This method adds layered sampling of typical floods on the basis of the same ratio method,and adjusts the different statistical characteristics of structural floods by sorting and changing the sampling points.So as to overcome the lack of diversity in the characteristics of traditional design flood methods.A total of 180 flood samples were generated to provide data support for follow-up research.(2)Propose a data-driven real-time population initialization method.For the same reservoir,its flood control and dispatching strategies have certain similarities.Based on this hypothesis,the historical flood dispatch results are used as the initial solution for flood optimization in a typical scenario to realize the display and migration of historical dispatching knowledge.For the problem of how to select target floods,a target selection strategy based on clustering and feature similarity ranking is proposed.The verification on typical floods in Ankang shows that the proposed algorithm improves the convergence efficiency by 10 times.(3)Propose an EMO-LSTM online scheduling algorithm based on optimal solution learning.Using the LSTM network as the migration medium,the scheduling strategy in the pareto optimal solution of the historical flood is extracted,and online decision-making can be made according to the current state when facing a new flood.Furthermore,in order to enhance the robustness and generalization ability of the algorithm,an EMO-LSTM algorithm based on sample partitioning is proposed.The historical flood samples are grouped according to the decision variable interval,and multiple groups of LSTM learners adapted to different scheduling modes are constructed to solve new problems together,and finally the Decision makers will select the appropriate solution according to the preference.It was verified on the structural floods with different frequencies of 6typical floods in Ankang.Experiments show that the solution of the EMO-LSTM algorithm is closer to the PF frontier than the rule scheduling.In the test flood with a larger flood volume,the EMO-LSTM algorithm based on sample partition is more effective than EMO-LSTM... |
PDF Full Text Request