| Because of the developments in renewable energy power generation technology and policy supporting, distributed generators (DGs) are playing significant roles in power systems. However, as plenty DGs be integrated into power grid, the disadvantages among management, randomness and controllability emerged gradually. Virtual power plant (VPP) is a mid-layer between DGs and power grid dispatch, it blocks the complexity of DG characteristics in bottom layer, and performing as a conceptual power plant, which could participating in power dispatching. The concept of VPP provides a feasible theoretical framework for further development of DGs, and has important practical significance in energy conservation, emissions reduction, and environmental protection. However, the diversity and complexity of DGs within VPP challenge the energy management and operation optimization. This paper proposed an optimization dispatch model for VPP, and investigate the variable reduction strategy for large-scale optimization problem.Firstly, the external dispatching cost characteristic is analyzed to realize optimization variable reduction for controllable plants set. A margin cost based fast algorithm which only has linear time complexity is designed for convex optimization problem. The form of fitting curve of external characteristic is proposed for convex quadratic functions which is often used in engineering application. The optimization variable reduction divides and conquers the whole optimization problem, and decreases the amount of optimization variables in VPP dispatch center.Secondly, the concept of clustering is used to realize variable reduction for uncontrollable plants. The distance matrix which indicates the dissimilarity between DGs is construct based on Pearson correlation coefficient. The clustering can be realized by k-means algorithm based on low dimensional coordinates extracted from distance matrix by Multi Dimensional Scaling (MDS).Thirdly, to improve the prediction accuracy for uncontrollable plant power output, several prediction model is compared, in which the BP neural network considering phase space and time have the best performance. Interval prediction model is proposed based on an assumption for prediction error distribution, and the correlation of prediction errors is further analyzed. When the prediction model is combined with random variable clustering, it seems that the clustering play roles in model simplifying, maintenance cost reduction and accuracy promotion.Lastly, quantum genetic algorithm (QGA) is used for non-convex VPP dispatching problem. Special coding and decoding strategy is designed for this kind of problems, and is verified by functional example. |
