【作者】 陆宁；

【导师】 周建中；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2010， 博士

【摘要】 电力负荷预测是电力系统进行规划设计的基础,是电力系统运行的经济性与可靠性的保证。由于电力负荷本身具有一定的不确定性、非线性、随机性等内在特点,负荷预测一直是学术研究的前沿与热点问题。随着电力市场的发展,负荷变化规律的复杂程度增加,而传统的单一预测方法自适应能力较差,问题的复杂性与求解方法的局限性之间的矛盾更加突出,预测不一定能得到满意的结果,因此,组合预测的探讨成为当今负荷预测的研究重点之一。组合预测能综合更多信息,增强单一的预测方法之间的互补,能够更完整的描述负荷的发展规律,达到提高预测精度的目的。组合预测思路主要有两方面：其一,从模型自身的优化与完善的角度,提出将学习算法和单一预测技术相结合的组合预测,用学习算法优化单一预测模型的相关参数以提高预测模型的预测精度。其二,将多种单一预测方法得到结果按一定方式结合起来,综合利用各种预测方法所提供的相关信息,并在综合这些信息的基础之上进行最优组合。本文针对电力负荷自身特点,引入群集智能优化计算方法以及组合预测技术,对电力负荷预测的理论与方法进行研究,取得了一定具有理论意义和实用价值的成果,主要研究工作与成果如下：(1)针对负荷预测中对单一预测模型进行优化的问题,研究群集智能优化计算方法及其改进,提出一种新的自适应粒子群算法(FAPSO),该算法根据粒子适应值的变化动态调整惯性权重ω的数值,平衡粒子的探索和开发两种搜索行为,提高粒子搜索能力的自适应性。同时在算法中加入极值扰动策略,防止算法陷入局部最优。对FAPSO算法进行了收敛性分析与收敛性测试,由仿真结果可知,与经典粒子群算法PSO比较,FAPSO具有更快的收敛速度与更可靠的全局收敛性。(2)提出一种由惯性权重递减型粒子群优化算法(LWPSO)与径向基函数神经网络(RBF)相结合的组合预测模型,应用粒子群算法的全局搜索能力搜索RBF神经网络的最优网络输出权值,从而达到优化RBF神经网络预测模型的目的。并利用虚拟仪器开发平台(LabVIEW)的强大数组处理能力和直观的编程方式实现了基于LWPSO优化的RBF神经网络,将训练好的网络对某实际电网进行了日整点负荷预测。仿真结果表明,该方法具有良好的预测精度和稳定性。(3)针对有限样本信息下的负荷预测问题,提出基于自适应粒子群算法FAPSO与支持向量机SVM相结合的组合预测模型。用FAPSO算法来优化SVM模型的相关参数,通过实例仿真与分析表明,与传统的通过交叉验证试算或遗传算法等确定SVM参数的方法相比,该方法不仅有较快的运算收敛速度,而且在预测精度和稳定性等方面都有一定程度提高。(4)研究粒子群算法PSO与模拟退火算法SA的自学习融合,提出新的优化算法NSAPSO,即在粒子群算法的寻优过程中加入模拟退火思想,结合SA算法的随机接受准则有效弥补经典粒子群算法在求解复杂问题较易陷入局部最优的缺点,使得算法在具有高效搜索的同时,能选择接受非最优解进而有能力跳出局部最优。同时在算法中加入极值扰动,增加粒子的多样性,进一步改善算法性能。由仿真测试结果,NSAPSO在对复杂多峰问题的寻优过程中体现出很好的收敛性能。(5)针对多单一预测模型的组合问题,提出基于自学习融合算法NSAPSO与SVM模型结合的组合方式,对单一模型实现分时段变权重非线性组合,用SVM的非线性回归学习能力描述多个单一预测模型的非线性组合关系,其优点在于不用确定描述输入输出关系的具体函数表达式,避免了传统组合方式中加权系数的复杂的求取问题,同时尽可能地满足了组合方法中的非线性、变权值的需求。仿真结果证明,该组合预测模型降低了传统单一预测模型的预测风险,同时有较高的预测精度和可靠性。更多还原

【Abstract】 Load forecasting is important for power system planning and design, system operation and management. It is the guarantee for the reliability and economic operation of electric power system. The intrinsic features of the electricity load are nonlinear, uncertainty, and randomness. For a long period scholars dedicated themselves to the research work of load forecasting technique and many effective methods have been proposed. However, the rules of load forecasting have become more complex with the development of electricity market and the traditional prediction way establish single model and has poor adaptive abilities. Traditional load forecasting method which using a single model to predict in complex conditions has obviously become less capable for the prediction task and may not be able to get the satisfactory results. Hence, the conflict between the complexity of the problem and the limitations of solving methods is more prominent. The combination Forecasting has become a focus for the research which can be integrated more information and enhanced the adaptive capability. The ideas for the combination can be summed up in two main aspects:Firstly, improve the performance of prediction model by the combination with the optimal algorithm. The algorithm is used to optimize the prediction models for better performance in forecasting. Secondly, the combination of different single prediction models by given each model certain weight. The combination model can integrate relevant information given by each model and improve the accuracy of load forecasting.In this paper, we deeply study the theories and methods of the load forecasting by adopting the virtual forecasting technique, swarm intelligence algorithm and combination forecasting theory, and obtain a series of conclusions which have theoretical and practical value. The main research and innovative results are as follows:(1) According to the optimization problem of the prediction model, describes the process of the standard particle swarm algorithm optimization and then proposed the improved particle swarm optimization(FAPSO). In particle swarm optimization, the search process must include explore and develop. The inertia weight needs larger step for global search at first, then the inertia weight requires gradually reduce for the local search. Take the changes of weight followed with the particles’fitness change. Map the changes in inertia weight by the exponential function. The disturbance for extreme value also added in the algorithm. The convergence analysis and the corresponding function tests for this self-adaptive particle swarm algorithm showed that the performance of FAPSO is better than that of the traditional particle swarm algorithm. (2) This paper proposes particle swarm optimization algorithm with weight linearity reduced (LWPSO) combined with Radial Basis Function (RBF) neural network model for short term load forecasting. The global search capability of particle swarm is used to optimize the weight of Radial Basis Function neural network. Implement the process by using LabVIEW for its powerful array processing capabilities and intuitionist way of programming. Simulation results show that the prediction accuracy and stability are better than that of the traditional one.(3) In view of the sample limited problem of load forecasting, the paper discussed support vector machine load forecasting model based on the improved particle swarm (FAPSO) optimization. The impact of the parameters for support vector machine is analyzed and then proposed FAPSO algorithm to optimize the relevant parameters of support vector machine which is found to overcome the shortcoming of the traditional way. The simulation shows that this method has faster convergence speed, and better performance compared with the traditional support vector machine parameter determination method.(4) The global convergence capability of simulated annealing (SA) and the efficiency of particle swarm search are combined as fusion algorithm in the way of self-study. Proposed the fusion algorithm NSAPSO with the disturbance for extreme value added in. SA has the ability of choosing to receive the bad solution in order to have the ability to leap out the local optimization. The simulation shows that this algorithm has better performance for global convergence.(5) In view of the problem of the combination of different single prediction models, support vector machines based on NSAPSO is proposed for nonlinear combined model. The model avoid the complex way of solving the weighting coefficient in the traditional combination way and it can fit with the requirement of non linear, variable weight requirement. It has a good performance and practice value for short term load forecasting.更多还原