【作者】 耿光飞；

【导师】 杨仁刚；

【作者基本信息】 中国农业大学 ， 农业电气化与自动化， 2003， 博士

【摘要】 随着经济的发展和电网自动化水平的提高,人们对电力系统稳定性、经济性及电能质量提出了更高要求,局部的和静态的无功优化控制已不能满足实际需要,实现地区电网动态无功优化是当前电力系统发展的迫切需要和必然趋势。本文结合实际情况,从优化方法、负荷预测、动态优化等几个方面进行研究,提出了一系列解决方案,得到了比较满意的结果。在借鉴经典优化方法中梯度概念的基础上,本文首次提出进化方向的概念,提出了基于选向变异的十进制遗传算法,该算法编码长度短而且精度高,比较有效地解决了长期以来遗传算法收敛速度慢这个主要缺陷。该算法采用更适合无功优化特点的分组整数实数混合编码方式,并采用映射法计算适应度函数。映射法给群体施加定常的选择压力,能在迭代初期缩小个体间的差异避免早熟,迭代后期放大个体间的差异进行细致比较,而且简单易行。算法中对最优个体进行单独变异,由于每一代的最优个体相对于其它个体而言接近全局最优的可能性最大,从而使搜索效率进一步提高。复合形法全局收敛性能不如遗传算法,但其综合比较若干个点来确定寻优方向的办法,简单实用,成功的概率较高。遗传算法全局收敛性好,但是由于变异的随机性使得收敛速度慢。本文将两者的优点有机结合起来,利用复合形寻优方法确定遗传算法的进化方向,提出了基于复合形变异方法的混合遗传算法,从而加快遗传算法的收敛速度。针对无功优化问题中存在离散变量和不能保证可行域的非凸性这两个特点,对复合形法进行改进,增加了判断中心点和映射点是否可行的环节,并对映射点进行邻域搜索和归整处理,使之更适合无功优化计算的要求。要使动态无功优化问题得到在时间段上的全局最优结果,必须有比较准确的母线负荷预测。人工神经网络算法通过将影响负荷的诸因素量化并计入预测模型中,从而得到较高的预测精度,但其训练工作量大。本文采用自适应权值调整方法,使人工神经网络(ANN)权值按最近的负荷行为动态地变化,既减少了训练工作量,又提高了预测精度。在确定ANN负荷预测模型输入量时,首先进行相关分析,剔除相关系数较小的影响因素,以提高预测精度。本文提出在预测负荷与实际负荷发生较大偏离的情况下,充分利用当日已知实际负荷的新信息,采用组合预测法来修正当日剩余时间的预测负荷值,从而进一步提高无功优化调度的准确性。文中将各种预测方法有机结合起来,设计了负荷预测模块的结构,提高了日负荷曲线预测的精度。文中建立了动态无功优化数学模型,并研究了动态无功优化约束条件的处理方法。为了满足无功补偿设备对动作次数的限制条件以及减少计算量,本文提出了用遗传算法对负荷曲线进行自动分段的新方法,为实现动态无功优化奠定了一个基础。通过合理的分段能显著降低计算量,同时兼顾电能损耗和控制设备动作次数两个目标。本文的无功优化控制与现有方法的不同之处在于其优化结果指导VQC 设备建立合理更多还原

【Abstract】 With the development of economy and power system automation, power system requires higher stability, economy and quality. Local and static reactive power optimization cannot meet the practical needs. It is an impendence need and an inevitable trend for power system currently to realize dynamic reactive power optimization in district power grid. In this paper, new resolutions have been put forward from several aspects, such as optimization method, load forecasting and dynamic optimization, and satisfied results are obtained. On the basis of utilizing the gradient concept in classical optimization method, the concept of evolution-direction is introduced into this field for the first time. An evolution-direction based mutation GA is put forward. This algorithm with short code and high precision overcomes the main shortcoming of slow convergence speed in GA. A grouped, integer and real number mixed coding method is applied in this algorithm. The mapped-method is adopted to calculate the fitness function. Compared with the other methods, mapped-method, which is very simple and easy, has no negative and provides the group a constant elective pressure. The optimal individual in every generation is more possible closer to global optimization than other individual. Based on this idea, the optimal individual in every generation is carried out particular mutation, thus the search efficiency is improved greatly. The global convergence character of Box algorithm is not better than GA. But it’s simple and practical for Box algorithm to find optimum direction by comparing several points. It’s worthy and convenience for GA to use the method. The Box algorithm is used to calculate evolution-direction of GA. The algorithm that combined GA and Box algorithm has the advantages of both GA and Box algorithm, and increased the convergence speed. Box algorithm is improved in accordance with two features of discrete variables in reactive power optimization and concave in feasible region. The center point and mapped point are judged whether they are feasible or not. The mapped point is turned into integer and its adjacent region is searched for better point. Such improvements make box algorithm more fit for the requirement of reactive power optimization. It should have precise bus load forecasting if we want to obtain dynamic reactive global optimum result in time axis. The weights of ANN are self-adapted according to the change of recent load so as to get higher forecasting precision and lower calculation quantity. It should do correlation analysis first and eliminate the input variable with small correlation coefficient. When the forecasting load is departure from the real load, it’s necessary to utilize the known information of the past load to reforecast the future load in the remaining time of the day. Combination methods are utilized to reforecast the future load of the day so as to improve the real time reactive power optimization control. Dynamic reactive power optimization model is established in the paper. It is studied how to deal with the restriction of dynamic reactive power optimization. A new approach utilizing genetic algorithm to auto-separate load curve in segments is offered. This method lays a foundation for realization of dynamic reactive power optimization. Calculating quantity is reduced a lot by appropriate separations of load curve giving consideration to both power loss and action times of controlling machine. The difference between optimization control method in this paper and other methods is that the optimization results do not control the equipment directly. It is studied how to build appropriate VQ limits according to the calculation of reactive power optimization. When deciding VQ limits of VQC, the forecasting load of every period of time is defined as a fuzzy number, which turned exact by a certain believe level to calculate the range of reactive control. The soft structure of the reactive power control system is designed in this paper. The reactive power optimization program has been tested in IEEE 6 nodes system, IEEE 30 nodes and a district power net. The results are fairly satisfactory.更多还原