Research On Multi-regional Collaborative Ensemble Learning Control Method For The Modern Complex Grid

The "dual carbon" goal has accelerated the rapid development of a new type of power system based on new energy,but the intermittent and uncertain large-scale integration of new energy sources poses enormous challenges to the stable operation of the power grid and may cause strong random disturbances.Additionally,as more plug-in electric vehicles are connected to the grid through charging stations,their dynamic changes can affect the control performance standard of the grid,leading to instability in grid control.In order to enable the grid to accommodate these strong stochastic fluctuations as much as possible,this paper explores a class of distributed automatic generation control algorithms for new energy from the perspective of automatic generation control,transforming the multi-area coordinated control of distributed AGC into a multi-agent optimal decision problem with Markov decision processes as the pivot and solving the problem using integrated learning methods.Firstly,the paper summarizes and reviews the current research status of the control performance evaluation criteria and control strategies of automatic generation control at home and abroad,elaborates on the control principles and structures of automatic generation control,and the relevant theoretical basis of machine learning.Finally,the automatic generation control process is transformed into a Markov decision process from three aspects: state,action,and reward value,and the specific mechanism of machine learning in automatic generation control is briefly described.The reasons for the transition from centralized automatic generation control to distributed automatic generation control are also analyzed based on reality.Secondly,a novel ensemble learning-based EBQ(σ,λ)algorithm is proposed from the perspective of automatic generation control to improve the control performance standard of the power system.The proposed algorithm not only solves the problem of value estimation bias in traditional reinforcement learning by reducing the mean squared error of the next state Q value but also balances "algorithm efficiency" and "training samples" by introducing the sampling parameter σ,and handles the problem of temporal credit assignment by incorporating eligibility trace policy.Simulation results of the improved IEEE standard two-area load frequency control model and Guangdong power grid model show that compared with traditional algorithms,the EBQ(σ,λ)algorithm can improve control performance indicators while reducing carbon emissions.Finally,an EBDQN algorithm based on ensemble learning and deep reinforcement learning is proposed.Although the traditional DQN algorithm can solve the problem of the curse of dimensionality,it can lead to the problem of value function overestimation.By introducing the integrated learning method and using the idea of "group decision-making to improve decision accuracy," the problem of value function overestimation is solved,and the performance of the algorithm is improved.By comparing the simulation results of multiple algorithms using the IEEE standard two-area load frequency control model and multi-area multi-source system model,the EBDQN algorithm is shown to achieve multi-area coordinated control and have superior performance.