Research On Optimal Scheduling With Distributed Generations On Consumer Side

With the consumption of fossil energy and global climate change,the development and utilization of new energy has become a consensus.Among them,wind energy has become a kind of renewable energy with great development potential.Within a certain range,wind energy is also the new energy with the greatest potential for development and utilization.However,the uncertainty of wind power makes it difficult to absorb wind power.When wind farms are established in areas with abundant wind energy,the long-term operation of wind farms requires continuous profitability of wind farms.Among them,the grid-connected operation of small wind farms has become the most scientific and economical way to configure energy storage.This is to focus on small wind farms,and explore effective energy storage system control strategies in small wind farms equipped with energy storage systems.And combined with the purchase of spare capacity to improve the self-discipline of wind farms,so as to ensure the power quality of large power grids and the revenue of wind farms.Based on the wind farm equipped with a real-time data collection system and forecasts of future wind power,future electricity prices and reserve prices,this paper uses a reinforcement learning method to solve the scheduling control problem of the energy storage system.The main research contents are as follows:This paper considers a wind farm system equipped with an energy storage system.This cooperation method is suitable for grid-connected operation of small wind farms.The mechanism of wind storage cooperation is introduced.The optimization problem model is established with the objective function of maximizing wind farm revenue,and considered In order to solve the various constraints of the energy storage system in the operation process,two classic reinforcement learning algorithms(Q-learning and DQN algorithms)are used to solve the optimization problem.In the Q-learning algorithm,due to the need to discretize the state variables,the consideration of the state variables is relatively simple,while the deep reinforcement learning algorithm(rainbow algorithm)allows the state variables to be continuous,so all the state variables can be truly considered As an improved version of the DQN algorithm,the rainbow algorithm allows the state space to be continuous,so the control amount is more precise.For the determination of the immediate reward of reinforcement learning,we use hypothetical methods to first obtain the charging and discharging power of the energy storage system required to achieve the control target,and then judge the ideal charging and discharging of the energy storage system according to the constraints of the energy storage system The feasibility of power,so as to gradually improve the penalty items that should be added to the immediate reward in various situations,and for the case where the charge and discharge power of the ideal energy storage system is feasible,the calculation of the immediate reward is consistent with the realization of the control target.