Research On Ultra-short Term Wind Power Prediction And AGC Dynamic Optimization Control Based On Machine Learning

The global warming caused by carbon dioxide emissions has brought irreversible damage to the ecological environment,and achieving the goal of “carbon peak and carbon neutrality” has become the direction of joint efforts of all countries in the world.Increasing the proportion of renewable energy power generation is an important means and decisive factors to carbon emission reduction and carbon neutrality.As the main form of renewable energy power generation,wind power is bound to usher in large-scale development.However,due to its intermittence,randomness and fluctuation,large-scale grid connection of wind power will bring great random fluctuating power to power system and adversely affect the system frequency.As an important part of the power management system of power grid dispatching center,Automatic Generation Control(AGC)is the main means to maintain the active power balance of the system and ensure the frequency quality of the system.How to adapt to the impact of the large-scale grid connection of wind power and improve the frequency control capability of system is a new research focus of AGC.This thesis studied two key issues in the AGC dynamic optimization control problem with machine learning technology.One was the minute-level wind power ultra-short term prediction,and the other was the AGC dynamic optimization control strategy.The main research results were as follows:(1)A one-minute ultra-short-term wind power prediction method based on Long Short-term Memory(LSTM)neural network was proposed.Firstly,the historical data of wind farms was preprocessed to locate and correct abnormal data and improve the overall quality of data.Then,Spearman rank correlation coefficient method was used to screen meteorological factors affecting wind power to improve the calculation efficiency of the model.Next,taking the short-term continuity characteristic of wind power into consideration,Spearman rank correlation coefficient was used to analyze the autocorrelation of wind power time series and determine the time step of the model,so as to improve the prediction accuracy of the model.Finally,the single-step and multi-step prediction models of wind power at one-minute level were built based on LSTM neural network.With the actual data of a wind field in western China,the performance of the proposed models was verified.The results showed that the proposed models in this thesis could effectively improve the accuracy of ultra-short term wind power prediction.(2)An AGC dynamic optimization control strategy based on Deep Reinforcement Learning(DRL)was proposed.Firstly,the AGC dynamic optimization control that takes wind power grid connection into consideration was transformed into a Markov decision process,whose control objective is to maximize cumulative rewards among multiple continuous time section.Secondly,the discrete reinforcement learning algorithm Dueling Deep Q Network(Dueling DQN)was used to build the dynamic optimization control model of AGC,and its solution strategy and working mode were given.Furthermore,because of the error caused by discretization and the dimensionality disaster problem in solving large-scale problems,an AGC dynamic optimization control strategy based on Proximal Policy Optimization(PPO)of continuous reinforcement learning algorithm was proposed.Finally,the improved IEEE-14 and IEEE-39 node system were used to verify the feasibility and effectiveness of the proposed AGC dynamic optimization control strategy based on deep reinforcement learning.