近年来，随着电网规模的日益增大，电力系统变得日趋复杂。交直流系统中引入的大量电力电子元件为电力系统带来了明显的非线性特性，再加上新能源出力的随机性与控制参数的分散性，这些都直接导致仿真对象的建模误差很难被准确估计。为此，本文提出一种参数校正算法，旨在根据实测波形校准仿真模型参数，使模型产生的波形曲线最符合客观现实。参数校正问题可以建模为优化问题，即在符合约束条件的各个参数组合中找到仿真结果与实际波形误差最小的组合。本文从不同模型的工程需求出发，基于不同的精度要求及仿真效率要求，建立与其相适应的参数校正算法，之后根据在工程实际中遇到的问题，进一步改进方法。本文围绕建模方法、仿真场景生成、参数校正算法、平台实现四个层面展开研究，具体的科研成果如下：第一，针对建筑物热负荷模型预测精度较低的工程问题，建立了基于热力学理论的供热系统模型，然后基于轨迹灵敏度判别系统的关键参数，再利用粒子群算法对参数进行优化，优化后的模型大幅提高了热负荷模型的预测精度，可应用于后续控制研究。第二，考虑到变压器饱和模型的研究现状与目前研究的局限性，设计了一种基于J-A磁滞理论的全新变压器电磁暂态模型，并通过插值法与回溯法改进了该模型的计算效率，从而最大程度上满足工程实际需求。然后基于贝叶斯分析中参数随机变量化的思想，提出了一种基于GMM的仿真模型参数校正方法，该方法适用于以变压器饱和模型为代表的设备级参数校正。第三，对于含暂态过程的复杂电力系统，由于仿真模型难以完全复现电网的真实动态行为，基于GMM的仿真模型参数校正算法很容易落入参数-特征量联合概率密度的低概率区间。为此在原算法的基础上，加入基于粒子群算法的闭环校验，通过对特征量进行小范围内的局部搜索，提高了算法对复杂系统的适用性。第四，考虑到实测波形可能受到扰动影响，导致无法反映系统的真实动态行为，原本的评价体系失去客观性的情况，本章将扰动因素按照成因分为两种特殊情况并提出相应的应对方法。仅实测波形受到扰动影响时采用基于特征的评价指标，而实测与仿真波形均含扰动源的情况以KL散度衡量波形之间的相似性。
In recent years, with the increase of the power grid, the power system has become increasingly complex. The large number of power electronic components introduced into the AC/DC system bring obvious nonlinear characteristics to the power system directly makes the modeling error of the simulation model difficult to be calculated. It is also difficult for the simulation model to fully reproduce the real dynamic behavior of the power grid even if it is explicitly estimated. Therefore, this paper proposes a parameter correction algorithm, which aims to calibrate the parameters of the simulation model according to the measured waveform, so that the curve generated by the model is most in line with the objective reality. The parameter calibration problem can be modeled as an optimization problem, that is, finding the combination with the smallest estimation error among the parameter combinations that meet the constraints. Starting from the actual engineering needs of different models, based on different accuracy requirements and simulation efficiency requirements, this paper establishes a parameter correction algorithm suitable for it, and then improves the method according to the problems encountered in engineering practice.This paper focusing on four research levels: modeling method, simulation scene generation, parameter correction algorithm, and platform implementation. The specific scientific research results are as follows:First, for the engineering problem of low prediction accuracy of the building heat load model, a heat load model of the heating system based on the thermodynamic system is established, and then the key parameters of the system are judged based on the trajectory sensitivity, and they are input into the particle swarm algorithm for optimization. The latter model greatly improves the fitting accuracy of the thermal load model, and can be applied to subsequent studies such as prediction and control.Second, considering the research status of transformer saturation model and the limitations of current research, a new electromagnetic transient model based on J-A hysteresis theory is designed, and the calculation efficiency of the model is improved by interpolation method and backtracking method. To the greatest extent possible to meet the actual needs of the project. Then, based on the idea of parameter random variation in Bayesian analysis, a GMM-based simulation model parameter calibration method is proposed to calibrate the parameters of the equipment-level system represented by the transformer saturation model.Third, for complex power systems with transient processes, since it is difficult for the simulation model to fully reproduce the real dynamic behavior of the power grid, the GMM-based simulation model parameter correction algorithm can easily fall into the low probability interval of the parameter-feature joint probability density. Therefore, on the basis of the original algorithm, the closed-loop verification link based on particle swarm algorithm is added, and the applicability of the algorithm to complex systems is improved by local search in a small range of feature quantities.Fourth, for the case where the measured waveform contains volatility, since the original evaluation system loses its objectivity, the fluctuation is divided into two special cases according to the causes, and feature extraction is used as the evaluation for the condition where only the measured waveform contains fluctuation. The KL divergence is used to measure the similarity between the distributions when both the measured waveform and the simulated waveform contain uncertainties.
