Research And Application On Intelligent Scheduling With Data-driven Learning Optimization Method For Large Power System

With the continuous increase in the scale of power grids,the physical characteristics of power generation and power consumption are gradually diversified.The large-scale application of new energy sources,such as wind and solar power,and massive dispatchable units make the scheduling optimization problem more and more complicated.The limitations of model-based methods have gradually become prominent.Data-driven optimization method has become an important support and development trend for the new generation of power systems.However,the existing data-driven learning methods suffer from a few problems,such as lack of adaptability,insufficient flexibility,and low learning efficiency when faced with power dispatch tasks that require refined output,strong logical correlation,and changeable application scenarios.In addition,the existing methods have insufficient representation of scheduling data,and it is difficult to construct the mapping between various types of scheduling data.When the operation mode of the power system changes,the lack of sufficient generalization ability for key data limits the practical application of data-driven intelligent dispatch optimization methods.Therefore,this article focuses on power generation scheduling for large-scale power system,relying on supervised learning and deep reinforcement learning methods.The main research works and results are shown as following four aspects:First,under the framework of supervised learning,an expert data set consisting of bus load data,topology data,maintenance data and power generation schedule is constructed.The specific meaning of the parameterized scheduling policy is given,which laid the foundation for the realization of the data-driven intelligent scheduling method.A scheduling policy learning algorithm based on convolutional neural network with mean squared error or mean absolute error(C-MSE/MAE)is proposed,which realizes the mining of the association relationship among the different type of expert data sets,and analyzes the performance of the method in accuracy and generalization ability.Furthermore,this paper introduces K-L distance to construct a mixed loss function,and proposes a scheduling strategy learning algorithm based on a generative confrontation mechanism(CGAN-MSE).The proposed method effectively improves the performance level of parameterized scheduling policy in the day-ahead power generation scheduling problem,reduces the computational time of the optimization process,and improves the decision-making efficiency.Secondly,we focus on how to optimize the scheduling policy through the self-learning method in the absence of expert data.Based on the sequential decision-making model for power dispatching task,a policy gradient algorithm considering continuous actions specifically for power dispatch optimization is proposed,and an autonomous learning optimization algorithm framework of offline learning-online decision-making is designed.Without the guidance of expert data,the method realizes that the dispatch strategy can quickly output a power generation plan with high economic benefits while ensuring the safety of the power grid.A parallel multi-scenario learning optimization method is proposed.The application of the method effectively improves the learning efficiency and performance effect of the parameterized scheduling strategy,and enhances the generalization and adaptability of the scheduling strategy in dealing with uncertain source-load scenarios.Then,after an analysis of the weight penalty cost method and the exploration space reduction,we find that the scheduling policy self-learning optimization method brings many drawbacks.It is proposed to embed the load balancing equation into the policy network action solution process.The process of ensuring a complete exploration of the space is a better alternative to the learning mode of penalizing the cost of constructing the equality constraint.Based on this,two algorithms are designed,namely Ray-shooting with Dynamic Interpterion(RDI)and Parameter Projection Method(PPM).The experiment results show that compared with the weight penalty cost and the reduction of exploration space,RDI and PPM present a greater improvement in learning efficiency and learning stability.Finally,relying on the New Generation Operation and Control System(NGOCS),the autonomous learning optimization method is applied to application.The interactive framework of offline learning and decision-making,the basis of realizing the independent update of the policy and the technical details in the application are given.The real-time power generation scheduling module based on self-learning optimization is developed on the NGOCS.The module maintains stable operation on the NGOCS,which provides a reference for dispatchers in the process of deciding real-time power generation schedules for a provincial power grid.