Low Carbon Distributed Generation Using Intelligent Optimal Allocation Algorithm For Microgrids Island

The development and innovation of micro-grid has been listed as one of the basicelements of smart grid construction. The modern power system had been many unstablefactors, actual operating conditions changed at all times, power system had a complexnon-linear characteristics, the application effects often limited to physical environments.Obviously, this does not meet the requirements of the building of smart grid. In order to solvethis problem, many scholars at home and abroad have been seeking a new self-learningoptimization algorithm.Therefore, this paper introduced RL(reinforcement learning)algorithm, which is an independent branch of Markov decision process.The maincharacteristics of RL algorithm is just to respond to the evaluation of the current controleffect,which had a higher real-time control and robustness.Therefore, reinforcement learningalgorithm had broad application prospects in Smart Grid.Firstly, the paper described the background, significance and research status ofmicro-grid. Load frequency model contained wind power, photovoltaic power generation,small hydro, gas turbines, energy storage, electric vehicles was built based onMATLAB/SIMULINK platform. Clean and low-carbon power in micro-grid model could beviewed as MAS(multi-agent system),therefore Q-learning based on stochastic optimal controland Q (λ) learning based on multi-step could be introduced to the field of micro-grid.Secondly, the concept of AGC(Automatic Generation Control)in regional interconnectedpower system was introduced into micro-grid,and AGC was divided into upper AGCcontroller and lower AGC controller.Upper AGC controller was a scheduling controller,whichautomatically send instructions by sampling on system frequency. Lower AGC was tooptimize the allocation factor controller based on RL algorithm.Lower AGC controller wasreplaced by using different algorithms. Results showed that,compared with PROP of capacityproportional distribution and Q-Learning algorithm based on discounted reward model, Q (λ)-Learning controller had a faster convergence characteristics and dynamic performance,whichcould effectively achieve the allocation strategy through on-line self-learning and dynamicoptimize decisions,which else could enhance the robustness and adaptability of the powersystem.Finally,real-time control of energy saving generation dispatch was introducted into themicro-grid model.Energy saving generation dispatch has been an important part of energysaving in the power system,real-time control of energy saving generation dispatch has beenthe difficulty in current engineering and scientific fields.Dynamic optimize allocation by AGC power instruction was presented to realize energy saving generation dispatch.Multi-step Q (λ)learning algorithm was introducted on the basis of the classic Q-learning in order to study theoptimal control strategy,which gave priority to low-carbon clean energy as much as possibleon the condition of ensuring micro-grid frequency and voltage stability.The simulation resultsshowed that,compared to classic PROP controller,optimization allocation controller based onreinforcement learning algorithm saved energy more than40%, which has been remarkableenergy-saving and economic benefits.The thesis is supported by the National Natural Science Foundation of China (50807016,51177051), the Fundamental Research Funds for the Central Universities(2012ZZ0020)，State Key Laboratory project of Tsinghua University (SKLD10KM01)..