| With the rapid development of new energy power generation technologies such as wind energy and solar energy,the power generation of new energy accounts for an increasing proportion of the power grid.The flexibility and deep peak regulation transformation of thermal power plants are the most effective way to absorb renewable energy.However,when the thermal power plants are operating at low load,the coupling,nonlinearity and large delay of the unit coordination system are more prominent and controllable.The performance is significantly reduced.Conventional control technology often has problems such as low unit variable load rate and large fluctuations in key parameters,which are difficult to effectively control.It is necessary to study a coordinated control system based on advanced control algorithms to effectively control the unit.This paper takes an ultra-supercritical unit coordination system as the research object.First,the system is modeled and the global dynamic model of the unit within the range of deep peak load regulation is established.On this basis,advanced control strategies based on deep reinforcement learning are proposed.In order to solve the control problems of the unit coordination system under deep peak regulation conditions,a new idea and new method are provided.The main research content of this paper includes the following aspects:(1)By analyzing the structure and operating characteristics of the ultra-supercritical unit,the controlled object model of the coordinated system within the full-load range of the ultrasupercritical unit is established by using the piecewise lumped method.A teaching and learning algorithm based on hierarchical learning is proposed,which is used to model the controlled object of the coordination system,and the model is further compared and verified with field test data.The results show that the established model can accurately reflect the operating characteristics of the unit under dry and wet conditions provide basic conditions for subsequent deep reinforcement learning research.(2)The principle and characteristics of deep reinforcement learning are studied.In order to reduce the number of samples obtained by the agent from the actual environment before the learning is completed,the Asynchronous Advantage Actor-Critic(A3C)algorithm is improved,Proposed an A3 C algorithm based on virtual environment-assisted learning,by adding a virtual environment model to improve the efficiency of the algorithm and the actual environment learning,and improve the push mechanism of each thread.The algorithm is used in the car climbing experiment to compare and test with various algorithms.The experiment shows that the improved deep reinforcement learning algorithm has improved learning convergence speed and stability after convergence.It also shows that deep reinforcement learning has the advantages of Ability to be applied in actual thermal process control.(3)In order to achieve the control goal of the coordinated control system under deep peak regulation conditions,the state design and reward design in the deep reinforcement learning algorithm are improved,and a reward value guidance method based on the human-simulated intelligent control idea is proposed.The new method is In the reward design,control quantity constraint items,continuous reward items and auxiliary task items are added.By analyzing the change of the reward value in the learning process,it is verified that the reward value guidance method in this paper can accelerate the learning process of the agent in the initial stage.On this basis,the learning and control framework of the ultra-supercritical unit coordinated control system is further proposed.Simulation tests show that the unit coordinated control system based on the improved deep reinforcement learning algorithm can obtain better setpoint tracking performance under dry and wet conditions,so that the unit has good load regulation under deep peaking conditions ability. |
PDF Full Text Request