文章快速检索     高级检索
  山东大学学报(工学版)  2016, Vol. 46 Issue (6): 62-68  DOI: 10.6040/j.issn.1672-3961.0.2016.095
0

引用本文 

孙一冰, 付敏跃, 王炳昌, 张焕水. 大规模动态系统的分布式状态估计算法[J]. 山东大学学报(工学版), 2016, 46(6): 62-68. DOI: 10.6040/j.issn.1672-3961.0.2016.095.
SUN Yibing, FU Minyue, WANG Bingchang, ZHANG Huanshui. Distributed state estimation algorithm for large-scale dynamic systems[J]. Journal of Shandong University (Engineering Science), 2016, 46(6): 62-68. DOI: 10.6040/j.issn.1672-3961.0.2016.095.

基金项目

国家自然科学基金资助项目(61120106011, 61573221, 61403233);国家科技支撑计划资助项目(2014BAF07B03)

作者简介

孙一冰(1987-), 男, 山东威海人, 博士研究生, 主要研究方向为分布式状态估计. E-mail:sun_yibing@126.com

通讯作者

付敏跃(1958-), 男, 浙江台州人, 教授, 博士, 主要研究方向为网络控制系统, 智能电网, 信号处理与通讯, 量化反馈控制. E-mail:minyue.fu@newcastle.edu.au

文章历史

收稿日期:2016-03-18
网络出版时间:2016-11-28 09:51:45
大规模动态系统的分布式状态估计算法
孙一冰1, 付敏跃2,3, 王炳昌1, 张焕水1     
1. 山东大学控制科学与工程学院, 山东 济南 250061;
2. 纽卡斯尔大学电气工程与计算机科学学院, 澳大利亚 新南威尔士州 纽卡斯尔 2308;
3. 广东工业大学自动化学院, 广东 广州 510006
摘要: 主要研究离散时间大规模动态系统的分布式状态估计问题。首先, 将系统划分为若干个子系统, 基于区域内部量测信息和邻居传递的信息, 各子系统利用该算法对本地状态进行估计, 降低状态变量的维数、算法的计算复杂度和通信压力。该算法独立运行, 并且平行运行该算法可以有效减少整体运行时间。通过减弱约束条件, 利用数学归纳法证明由该算法得到的估计误差协方差和预测误差协方差矩阵正定。根据系统能观测性秩判据和不等式技巧, 证明误差协方差矩阵有上界, 并且上界是有界的, 保证该算法在应用中的可行性。最后通过仿真研究, 验证主要结论。
关键词: 状态估计    电力系统    动态系统    分布式估计    最大后验估计    
Distributed state estimation algorithm for large-scale dynamic systems
SUN Yibing1, FU Minyue2,3, WANG Bingchang1, ZHANG Huanshui1     
1. School of Control Science and Engineering, Shandong University, Jinan 250061, Shandong, China;
2. School of Electrical Engineering and Computer Science, University of Newcastle, NSW 2308, Australia;
3. School of Automation, Guangdong University of Technology, Guangzhou 510006, Guangdong, China
Abstract: The problem of distributed state estimation over discrete-time large-scale dynamic systems was studied. The system was divided into some subsystem, and based on the local measurement and the information received from its neighbors, each subsystem utilized the proposed algorithm to estimate its local state, which reduced the dimension of the state vector, and enjoyed low computational complexity and communication load. This algorithm was run independently and in parallel to effectively reduce the overall execution time. By weakening the constraint condition, the mathematical induction was used to prove that the state estimation and prediction error covariance matrices obtained from this algorithm were positive definite. The rank criterion of system observability together with the inequality technique were utilized to prove that error covariance matrices had upper bounds and the upper bounds were also existence and bounded, which supported the feasibility of this algorithm in applications. At last, simulations of an example were provided to demonstrate the main results.
Key words: state estimation    power system    dynamic system    distributed state estimation    MAP estimation    
0 引言

为保证现代化的电力系统安全可靠地运行, 能量管理系统(energy management system, EMS)得到了广泛的应用并发挥了重要作用。电力系统的状态估计是EMS的重要组成部分[1], 主要功能是利用电力系统的实时冗余量测信息, 对系统状态进行估计, 得到尽可能准确的估计值。状态估计保证了EMS执行不同的控制和决策任务, 例如最优化功率流、电压的稳定性分析、异常数据的检测和分析等[2-3]

随着电力系统的飞速发展, 电力系统状态估计, 尤其是动态状态估计, 受到了不同领域学者的广泛关注。由于电力系统是典型的动态系统, 状态变量随时间变化, 但是静态状态估计不考虑状态随时间变化的特点, 所以动态状态估计比静态估计更适合电力系统。目前, 关于电力系统动态估计的研究已经有很多成果, 如基于扩展卡尔曼滤波(extended Kalman filter, EKF)的动态估计算法[4-8], 基于无迹卡尔曼滤波(unscented Kalman filter, UKF)的动态估计算法[9-11]和基于容积卡尔曼滤波(cubature Kalman filter, CKF)的动态估计算法[12-13]。随着区域电网的互联和电力系统规模的不断扩大, 传统的状态估计算法在估计精度和计算速度上遇到挑战, 并且不便于实施和管理, 这促进了分布式状态估计的发展[14-19]。文献[14-15]通过对电力系统分区, 在每个区域内分别进行状态估计的迭代运算, 从而达到分布式状态估计的目的, 这样做大大地减少了运行时间, 但是邻居子系统所包含的边界节点是重叠的, 必须满足边界节点电压幅值和相角分别相等的约束条件。文献[19]基于MAP估计技术, 提出了一种新的分布式状态估计算法。该算法提高了状态估计的运算速度, 降低了通讯负担, 适用于大规模的电力系统。

本研究继续研究文献[19]中提出的分布式动态状态估计算法, 但是减弱了文献[19]所需的约束条件, 即文献[19]假设每个子系统的内部量测矩阵列满秩, 而本研究假设每个子系统是可观测的。根据这个条件和文献[19]中提供的结论, 本研究证明了由该算法得到的估计误差协方差和预测误差协方差是正定并且有上界的, 保证了算法的可行性。最后通过IEEE 39节点系统的模拟仿真, 验证了本研究的主要结果。

1 问题描述

在介绍系统模型之前, 首先给出本研究需要的一些预备知识。

考虑含有N个子系统的网络G={V, E}, 其中V={1, …, N}表示子系统代表的点集, EV×V是边的集合, 假设图G是无向连通的。Ni={jV:(i, j)∈E}表示子系统i的邻居集合。Ni/{j}表示集合中不包含节点j。对于无环的图G, εi表示子系统i到图中其它任意子系统的最大路径长度。Rl表示l维实列向量的集合, Rl×q表示l×q的实矩阵的集合。为了方便, 记MnT=(Mn)T。diag{A1, …, An}表示分块对角矩阵, 对角线的元素是矩阵A1, …, An。rank (A)表示矩阵A的秩。⊗表示矩阵的克罗内克积。

一个大的系统可划分成N个不重叠的子系统。每个子系统内部含有一个控制中心, 用来收集子系统的量测, 并对子系统的内部状态进行估计。每个子系统独立地并行运行算法, 这样可以极大减少运行时间, 并且可以降低所需研究的矩阵的规模。对于子系统i, 考虑如下所示的离散时间线性状态方程:

${\mathit{\boldsymbol{x}}_i}\left( {k + 1} \right) = {\mathit{\boldsymbol{A}}_i}{\mathit{\boldsymbol{x}}_i}\left( k \right) + {\mathit{\boldsymbol{g}}_i} + {\mathit{\boldsymbol{\omega }}_i}\left( k \right),$ (1)

式中,xi(k)∈Rni为子系统ik时刻的状态; AiRni×ni为定常的状态转移矩阵, 假设Ai是可逆矩阵; gi是非零向量, 与状态轨迹的趋势相关; ωi(k)是高斯白噪声, 其均值为0, 方差Ri≥0。假设初始状态xi(0)是均值为xi(0), 方差为Pi(0)>0的高斯变量。采样时间k的取值是0, 1, 2, …。

注1    在时变系统中, 利用Holt两参数线性指数平滑法[20], 求得

$ {\boldsymbol{A}_i} = {\alpha _i}(1 + {\beta _i})\boldsymbol{I}, $

其中0 < αi, βi < 1。所以本文中假设Ai可逆是合理的。

子系统i的量测方程可表示成如下两类:

${\mathit{\boldsymbol{z}}_{i,i}}\left( k \right) = {\mathit{\boldsymbol{C}}_i}{\mathit{\boldsymbol{x}}_i}\left( k \right) + {\mathit{\boldsymbol{v}}_{i,i}}\left( k \right),$ (2)
${\mathit{\boldsymbol{z}}_{i,j}}\left( k \right) = {\mathit{\boldsymbol{B}}_{ij}}{\mathit{\boldsymbol{x}}_i}\left( k \right) + {\mathit{\boldsymbol{B}}_{ji}}{\mathit{\boldsymbol{x}}_j}\left( k \right) + {\mathit{\boldsymbol{v}}_{i,j}}\left( k \right),$ (3)

式中,zi, i(k)∈Rrii为子系统i的内部量测; zi, j(k)∈Rrij为连接子系统ij的边的量测; CiBijBji为具有适当维数的常矩阵; νi, i(k)和νi, j(k)是高斯白噪声, 其均值为0, 方差Ti, i>0和Ti, j>0。假设ωi(k)、νi, i(k)、νi, j(k)和xi(0)是互不相关的。

注2    在电力系统中, 电压(幅值和相角)可以看作是系统状态, 量测包括节点电压的幅值量测、有功功率和无功功率, 其中邻居节点都可以得到功率量测, 所以zi, j(k)由子系统之间的有功功率和无功功率组成; zi, i(k)由子系统内部节点的幅值量测、有功功率和无功功率组成。这也证实了本文研究的系统适用于电力系统。由于子系统i和j都可以得到边的量测(3), 所以假设zi, j(k)=zj, i(k)。

2 分布式状态估计算法

本文假设划分后的子系统组成的图G是无环的。

定义${{\mathit{\boldsymbol{\hat x}}}_i}\left( {k|k - 1} \right)$Σi(k|k-1)分别是子系统i基于0~k-1时刻量测的状态预测和误差协方差。定义局部信息向量和信息矩阵为

$\begin{array}{l} {{\mathit{\boldsymbol{\bar \alpha }}}_i}\left( k \right) = \mathit{\boldsymbol{C}}_i^T\mathit{\boldsymbol{T}}_{i,i}^{ - 1}{\mathit{\boldsymbol{z}}_{i,i}}\left( k \right) + \sum\nolimits_i^{ - 1} {\left( {k|k - 1} \right){{\mathit{\boldsymbol{\hat x}}}_i}\left( {k|k - 1} \right)} ,\\ \;\;\;\;\;\;\;\;\;\;\;\;{{\mathit{\boldsymbol{\bar Q}}}_i}\left( k \right) = \mathit{\boldsymbol{C}}_i^T\mathit{\boldsymbol{T}}_{i,i}^{ - 1}{\mathit{\boldsymbol{C}}_i} + \sum\nolimits_i^{ - 1} {\left( {k|k - 1} \right).} \end{array}$ (4)

文献[19]给出的分布式MAP估计算法如下:

(1) 初始迭代

在时刻k=0, 1, 2, …, 每个子系统i计算局部的状态估计和估计误差协方差为

$ \begin{array}{l} {{\mathit{\boldsymbol{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over x} }}}_{i,j}}\left( {k|k,0} \right) = \mathit{\boldsymbol{\bar Q}}_i^{ - 1}{\left( k \right)_i}{{\mathit{\boldsymbol{\bar \alpha }}}_i}\left( k \right),\\ {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }}_{i,j}}\left( {k|k,0} \right) = \mathit{\boldsymbol{\bar Q}}_i^{ - 1}\left( k \right); \end{array} $

同时, 子系统i计算降维的估计信息

$ \begin{array}{l} \mathit{\boldsymbol{\beta }}_j^j\left( {k,0} \right) = {\mathit{\boldsymbol{B}}_{ji}}{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over x} }_{j,i}}\left( {k|k,0} \right),\\ \mathit{\boldsymbol{ \boldsymbol{\varPhi} }}_j^j\left( {k,0} \right) = {\mathit{\boldsymbol{B}}_{ji}}{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }_{j,i}}\left( {k|k,0} \right)B_{ji}^T, \end{array} $ (5)

并将该信息传递给邻居j

(2) 循环

迭代次数h=1, 2, …, 对于子系统i,

① 根据邻居jNi传递的信息βji (k, h-1)和Φji (k, h-1), 子系统i更新边界量测, 得到

${y_{ji}}\left( {k,h} \right) = {z_{i,j}}\left( k \right) - \beta _j^i\left( {k,h - 1} \right),$ (6)
${S_{ji}}\left( {k,h} \right) - {T_{i,j}} + \Phi _j^i\left( {k,h - 1} \right).$ (7)

② 子系统i计算当前的状态估计和估计误差协方差

$\begin{array}{l} {{\mathit{\boldsymbol{\hat x}}}_i}\left( {k|k,h} \right) = \mathit{\boldsymbol{Q}}_i^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{\alpha }}_i}\left( {k,h} \right),\\ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i}\left( {k|k,h} \right) = \mathit{\boldsymbol{Q}}_i^{ - 1}\left( {k,h} \right) \end{array}$ (8)

式中:

${\mathit{\boldsymbol{\alpha }}_i}\left( {k,h} \right) = {{\mathit{\boldsymbol{\bar \alpha }}}_i}\left( k \right) + \sum\limits_{j \in {N_i}} {\mathit{\boldsymbol{B}}_{ij}^{\rm{T}}} \mathit{\boldsymbol{S}}_{ji}^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{y}}_{ji}}\left( {k,h} \right),$ (9)
${\mathit{\boldsymbol{Q}}_i}\left( {k,h} \right) = {{\mathit{\boldsymbol{\bar Q}}}_i}\left( k \right) + \sum\limits_{j \in {N_i}} {\mathit{\boldsymbol{B}}_{ij}^{\rm{T}}} \mathit{\boldsymbol{S}}_{ji}^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{B}}_{ji}}$ (10)

③ 同时, 子系统i得到

$ \begin{array}{l} {{\mathit{\boldsymbol{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over x} }}}_{i,j}}\left( {k,h} \right) = {\left[ {\mathit{\boldsymbol{Q}}_i^j\left( {k,h} \right)} \right]^{ - 1}}\mathit{\boldsymbol{\alpha }}_i^j\left( {k,h} \right),\\ {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }}_{i,j}}\left( {k,h} \right) = {\left[ {\mathit{\boldsymbol{Q}}_i^j\left( {k,h} \right)} \right]^{ - 1}}, \end{array} $

式中:$\mathit{\boldsymbol{\alpha }}_i^j\left( {k,h} \right) = {{\mathit{\boldsymbol{\bar \alpha }}}_i}\left( k \right) + \sum\limits_{m \in {N_i}\left\{ j \right\}} {\mathit{\boldsymbol{B}}_{mi}^{\rm{T}}} \mathit{\boldsymbol{S}}_{mi}^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{y}}_{mi}}\left( {k,h} \right);{\rm{ }}$; $\mathit{\boldsymbol{Q}}_i^j\left( {k,h} \right) = {{\mathit{\boldsymbol{\bar Q}}}_i}\left( k \right) + \sum\limits_{m \in {N_i}\left\{ j \right\}} {\mathit{\boldsymbol{B}}_{im}^{\rm{T}}} \mathit{\boldsymbol{S}}_{mi}^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{B}}_{im}}$。并将βij (k, h)和Φij (k, h)传递给邻居j

文献[19]中的定理1已经证明了由子系统组成的图无环时, 在每个时刻k, 子系统i的状态估计迭代有限次之后收敛, 即对于所有的l≥0,

$\begin{array}{l} {{\mathit{\boldsymbol{\hat x}}}_i}\left( {k|k,{\varepsilon _i} + l} \right) = \mathit{\boldsymbol{\hat x}}_i^*\left( {k|k} \right),\\ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i}\left( {k|k,{\varepsilon _i} + l} \right) = \mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right). \end{array}$ (11)

根据在时刻k得到的状态估计$\mathit{\boldsymbol{\hat x}}_i^*\left( {k|k} \right)$和估计误差协方差Σi* (k|k), 可以对k+1时刻的状态进行预测。子系统i的状态预测方程为

$\begin{array}{l} {{\mathit{\boldsymbol{\hat x}}}_i}\left( {k + 1|k} \right) = {\mathit{\boldsymbol{A}}_i}\mathit{\boldsymbol{\hat x}}_i^*\left( {k|k} \right) + {\mathit{\boldsymbol{g}}_i},\\ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i}\left( {k + 1|k} \right) = {\mathit{\boldsymbol{A}}_i}\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)\mathit{\boldsymbol{A}}_i^{\rm{T}} + {\mathit{\boldsymbol{R}}_i}. \end{array}$ (12)
3 算法的有界性

下面的两个定理证明了Σi* (k|k)和Σi(k+1|k)有界。

定理1    在每个k时刻, 子系统i的估计误差协方差和预测误差协方差是正定的, 即

$ \begin{array}{l} \mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right) > 0,\\ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i}\left( {k + 1|k} \right) > 0。 \end{array} $

证明    根据数学归纳法证明这个定理。当k=0并且h=0时, 根据每个子系统i的初始值Σi(0|-1)=Pi(0)>0和Ti, i>0, 得到Qi(0)>0和${{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }_{i,j}}$ (0|0, 0)>0。根据式(5)和(7)有Sji(0, 1)>0。由式(10)可得Qi(0, 1)>0, 从而得到Σi(0|0, 1)>0。假设该结论在h>1时成立, 即对于所有的子系统i, 有Σi(0|0, h)>0和 ${{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }_{i,j}}$ (0|0, h)>0成立。当迭代次数为h+1时, 由式(5)和(7)得到Sji(0, h+1)>0。再次根据式(10)可得Σi(0|0, h+1)>0。当hεi时, 根据式(11)得到Σi* (0|0)>0, 并且由式(12), 显然可得Σi(1|0)>0。

假设在k>0时刻, 该结论成立, 即Σi* (k|k)>0和Σi(k+1|k)>0。在k+1时刻并且h=0时, 根据式(4)每个子系统i得到Qi(k+1)>0从而有${{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over \Sigma } }_{i,j}}$ (k+1|k+1, 0)>0。根据式(5)和(7), 有Sji(k+1, 1)>0。由式(10)可得Qi(k+1, 1)>0, 并且Σi(k+1|k+1, 1)>0。根据此前用到的数学归纳法可证得当hεi时, 有Σi* (k+1|k+1)>0, 并且由式(12)得到Σi(k+2|k+1)>0。证毕。

在分布式状态估计算法中, 边量测的更新方程(6)可以写成:

${\mathit{\boldsymbol{y}}_{ji}}\left( {k,h + 1} \right) = {\mathit{\boldsymbol{B}}_{ij}}{\mathit{\boldsymbol{x}}_i}\left( k \right) + {\mathit{\boldsymbol{B}}_{ji}}\left( {{x_j}\left( k \right) - {{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\smile$}} \over x} }_{j,i}}\left( {k|k,h} \right)} \right) + {\mathit{\boldsymbol{v}}_{i,j}}\left( k \right)$ (13)

式(13)右边的最后两项可以看作是量测噪声, 方差是Φji (k, h)+Ti, j。式(9)和(10)可以写成

$\begin{array}{l} {\mathit{\boldsymbol{\alpha }}_i}\left( {k,h} \right) = \mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\bar R}}_i^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{z}}_i}\left( {k,h} \right) + \mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^{ - 1}\left( {k|k - 1} \right){{\mathit{\boldsymbol{\hat x}}}_i}\left( {k|k - 1} \right),\\ {\mathit{\boldsymbol{Q}}_i}\left( {k,h} \right) = \mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\bar R}}_i^{ - 1}\left( {k,h} \right){\mathit{\boldsymbol{H}}_i} + \mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^{ - 1}\left( {k|k - 1} \right) \end{array}$ (14)

即向量zi(k, h)由量测zi, i(k)和yji(k, h), (jNi)组成, 用来估计子系统i的状态, 矩阵HiCiBij组成, Ri(k, h)是分块对角矩阵, 对角线的分块矩阵分别是Ti, iSji(k, h)。特别地, 如果子系统i拥有邻居1, …, ai, 则zi(k, h), HiRi(k, h)具有如下形式:

$ \begin{array}{l} {\mathit{\boldsymbol{z}}_i}\left( {k,h} \right) = \left( {\begin{array}{*{20}{c}} {{\mathit{\boldsymbol{z}}_{i,i}}\left( k \right)}\\ {{\mathit{\boldsymbol{y}}_{1i}}\left( {k,h} \right)}\\ \vdots \\ {{\mathit{\boldsymbol{y}}_{{a_i}i}}\left( {k,h} \right)} \end{array}} \right),{H_i} = \left( {\begin{array}{*{20}{c}} {{\mathit{\boldsymbol{C}}_i}}\\ {{\mathit{\boldsymbol{B}}_{i1}}}\\ \vdots \\ {{\mathit{\boldsymbol{B}}_{i{a_i}}}} \end{array}} \right),\\ {{\mathit{\boldsymbol{\bar R}}}_i}\left( {k,h} \right) = \left[ {\begin{array}{*{20}{c}} {{\mathit{\boldsymbol{T}}_{i,i}}}&0& \cdots &0\\ 0&{{\mathit{\boldsymbol{S}}_{1i}}\left( {k,h} \right)}& \cdots &0\\ \vdots & \vdots & \ddots & \vdots \\ 0&0& \cdots &{{\mathit{\boldsymbol{S}}_{{a_i}i}}\left( {k,h} \right)} \end{array}} \right]. \end{array} $

事实上, 定理1已经证明了Sji(k, h)>0, 所以矩阵Ri(k, h)也是正定的。定义矩阵Wi=diag{Ti, i, Ti, 1, …, Ti, ai}>0, 则Ri(k, h)≥Wi

k时刻, 当估计误差协方差收敛到Σi* (k|k)时, Sji(k, h)也是收敛的。定义矩阵${{\mathit{\boldsymbol{\tilde R}}}_i}\left( k \right) = {{\mathit{\boldsymbol{\bar R}}}_i}\left( {k,h} \right)$, hεi。根据式(8) (11)和(14), 得到

${\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right)^{ - 1}} = \mathit{\boldsymbol{H}}_i^{\rm{T}}{{\mathit{\boldsymbol{\tilde R}}}_i}\left( k \right){\mathit{\boldsymbol{H}}_i} + \mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^{ - 1}\left( {k|k - 1} \right).$ (15)

定义矩阵

$ \begin{array}{l} \mathit{\boldsymbol{A}} = \left( {\begin{array}{*{20}{c}} {\mathit{\boldsymbol{A}}_i^{ - ({n_i} - 1)}}\\ {\mathit{\boldsymbol{A}}_i^{1 - ({n_i} - 1)}}\\ \vdots \\ \mathit{\boldsymbol{I}} \end{array}} \right),\mathit{\boldsymbol{\tilde A = }}\left( {\begin{array}{*{20}{c}} {\mathit{\boldsymbol{A}}_i^{ - {n_i}}}\\ {\mathit{\boldsymbol{A}}_i^{ - ({n_i} - 1)}}\\ \vdots \\ {\mathit{\boldsymbol{A}}_i^{ - 1}} \end{array}} \right),\\ \mathit{\boldsymbol{R}}\left( k \right) = \\ \left[ {\begin{array}{*{20}{c}} {{{\left( {1 + c} \right)}^{ - ({n_i} - 1)}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}(k - ({n_i} - 1))}&0& \cdots &0\\ 0&{{{\left( {1 + c} \right)}^{1 - ({n_i} - 1)}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}k + 1 - ({n_i} - 1))}& \cdots &0\\ \vdots & \vdots & \ddots & \vdots \\ 0&0& \cdots &{\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( k \right)} \end{array}} \right], \end{array} $

其中常数c>0; kni-1。定义Wi=diag{(1+c)-(ni-1)Wi-1, (1+c)1-(ni-1)Wi-1, …, Wi-1}, 则R(k)≤i。下面的定理将证明Σi* (k|k)和Σi(k+1|k)有上界。

定理2   假设矩阵列(Ai, Hi)完全能观测, 则存在常数c>0, 使得协方差矩阵Σi*(k|k)和Σi(k+1|k), (kni-1), 有上界, 即

$\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right) < {\left( {{{\left( {\left( {\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i}} \right)\mathit{\boldsymbol{A}}} \right)}^{\rm{T}}}\mathit{\boldsymbol{R}}\left( k \right)\left( {\left( {\mathit{\boldsymbol{I}} \otimes \mathit{\boldsymbol{H}}} \right)\mathit{\boldsymbol{A}}} \right)} \right)^{ - 1}},$ (16)
$\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^{ - 1}\left( {k + 1|k} \right) < {\left( {{{\left( {\left( {\mathit{\boldsymbol{I}} \otimes \mathit{\boldsymbol{H}}} \right)\mathit{\boldsymbol{\tilde A}}} \right)}^{\rm{T}}}\mathit{\boldsymbol{R}}\left( k \right)\left( {\left( {\mathit{\boldsymbol{I}} \otimes \mathit{\boldsymbol{H}}} \right)\mathit{\boldsymbol{\tilde A}}} \right)} \right)^{ - 1}} + {\mathit{\boldsymbol{R}}_i}.$ (17)

证明    定理1已证明在每个时刻k, 子系统i的估计误差协方差Σi* (k|k)>0。由于Ri≥0并且Ai可逆, 则存在常数c>0使得

${\mathit{\boldsymbol{R}}_i} \le c{\mathit{\boldsymbol{A}}_i}\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)\mathit{\boldsymbol{A}}_i^{\rm{T}}$ (18)

由式(15)(18), 得到

${\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right)^{ - 1}} \ge \mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( k \right){\mathit{\boldsymbol{H}}_i} + {\left( {\left( {1 + c} \right){\mathit{\boldsymbol{A}}_i}\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - 1|k - 1} \right)\mathit{\boldsymbol{A}}_i^{\rm{T}}} \right)^{ - 1}}.$ (19)

将式(19)两端同时乘以AiTAi, 并根据Ai可逆, 得到

$\mathit{\boldsymbol{A}}_i^{\rm{T}}{\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right)^{ - 1}}{\mathit{\boldsymbol{A}}_i} \ge \mathit{\boldsymbol{A}}_i^{\rm{T}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( k \right){\mathit{\boldsymbol{H}}_i}{\mathit{\boldsymbol{A}}_i} + {\left( {\left( {1 + c} \right){\mathit{\boldsymbol{A}}_i}\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - 1|k - 1} \right)} \right)^{ - 1}}.$ (20)

类似地, 在k-1时刻, 可得

$\mathit{\boldsymbol{A}}_i^{\rm{T}}{\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - 1|k - 1} \right)} \right)^{ - 1}}{\mathit{\boldsymbol{A}}_i} \ge \mathit{\boldsymbol{A}}_i^{\rm{T}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k - 1} \right) \\ {\mathit{\boldsymbol{H}}_i}{\mathit{\boldsymbol{A}}_i} + {\left( {\left( {1 + c} \right)\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - 2|k - 2} \right)} \right)^{ - 1}}.$ (21)

将式(20)和(21)结合, 得到

$ \begin{array}{l} \mathit{\boldsymbol{A}}_i^{2{\rm{T}}}{(\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right))^{ - 1}}\mathit{\boldsymbol{A}}_i^2 \ge \mathit{\boldsymbol{A}}_i^{2{\rm{T}}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( k \right){\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^2 + {\left( {1 + c} \right)^{ - 1}}\\ \mathit{\boldsymbol{A}}_i^{\rm{T}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k - 1} \right){\mathit{\boldsymbol{H}}_i}{\mathit{\boldsymbol{A}}_i} + {\left( {1 + c} \right)^{ - 2}}{(\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - 2|k - 2} \right))^{ - 1}}。 \end{array} $

继续进行, 得到

$ \begin{array}{l} \mathit{\boldsymbol{A}}_i^{({n_i} - 1){\rm{T}}}{\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right)^{ - 1}}\mathit{\boldsymbol{A}}_i^{{n_i} - 1} \ge \mathit{\boldsymbol{A}}_i^{({n_i} - 1){\rm{T}}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( k \right){\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^{{n_i} - 1} + {\left( {1 + c} \right)^{ - 1}}\\ \mathit{\boldsymbol{A}}_i^{({n_i} - 2){\rm{T}}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k - 1} \right){\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^{{n_i} - 2} + \cdots + {\left( {1 + c} \right)^{ - ({n_i} - 2)}}\mathit{\boldsymbol{A}}_i^{\rm{T}}\mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k - {n_i} + 2} \right){\mathit{\boldsymbol{H}}_i}{\mathit{\boldsymbol{A}}_i} + \\ {\left( {1 + c} \right)^{ - ({n_i} - 1)}}{\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - {n_i} + 1|k - {n_i} + 1} \right)} \right)^{ - 1}}。 \end{array} $

由于

$ \begin{array}{l} {\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k - {n_i} + 1|k - {n_i} + 1} \right)} \right)^{ - 1}} = \\ \mathit{\boldsymbol{H}}_i^{\rm{T}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k - {n_i} + 1} \right){\mathit{\boldsymbol{H}}_i} + \sum _i^{ - 1}\left( {k - {n_i} + 1|k - {n_i} + 1} \right), \end{array} $

并且定理1已经证明了Σi-1 (k-ni+1|k-ni)>0, kni-1, 得到

$\mathit{\boldsymbol{A}}_i^{\left( {{n_i} - 1} \right){\rm{T}}}{\left( {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right)^{ - 1}}\mathit{\boldsymbol{A}}_i^{{n_i} - 1} > \sum\limits_{l = 0}^{{n_i} - 1} {{{\left( {1 + c} \right)}^{l - \left( {{n_i} - 1} \right)}}} \\ {\left( {{\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^l} \right)^{\rm{T}}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}\left( {k + l - \left( {{n_i} - 1} \right)} \right){\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^l.$ (22)

将式(22)两端分别乘以Ai-(ni-1) TAi-(ni-1), 可得

$ \begin{array}{l} {(\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right))^{ - 1}} > \sum\limits_{l = 0}^{{n_i} - 1} {{{\left( {1 + c} \right)}^{l - ({n_i} - 1)}}} {({\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^{l - ({n_i} - 1)})^{\rm{T}}}\mathit{\boldsymbol{\tilde R}}_i^{ - 1}(k + l - ({n_i} - 1))\\ {\mathit{\boldsymbol{H}}_i}\mathit{\boldsymbol{A}}_i^{l - ({n_i} - 1)} = {((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}})^{\rm{T}}}\mathit{\boldsymbol{R}}\left( k \right)((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}})。 \end{array} $

因为(Ai, Hi)是可观测的, Ai可逆, 并且R(k)正定, 有

$ \begin{array}{c} {\rm{rank}}((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}}) = {n_i},\\ {((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}})^{\rm{T}}}\mathit{\boldsymbol{R}}\left( k \right)((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}}) > 0, \end{array} $

则得到式(16)。根据式(12)和(16), 得到

$ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }} _i}\left( {k + 1|k} \right) < {\mathit{\boldsymbol{A}}_i}{\left( {{{((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}})}^{\rm{T}}}\mathit{\boldsymbol{R}}\left( k \right)((\mathit{\boldsymbol{I}} \otimes {\mathit{\boldsymbol{H}}_i})\mathit{\boldsymbol{A}})} \right)^{ - 1}}\mathit{\boldsymbol{A}}_i^{\rm{T}} + {\mathit{\boldsymbol{R}}_i}, $

从而可得式(17)。证毕。

注3    文献[19]假设每个子系统的内部量测矩阵Ci列满秩, 而本文需要条件(Ai, Hi)完全能观测即可, 这表明本文的条件减弱了文献[19]所需的条件, 但是该条件仍然可以保证误差协方差矩阵正定并且有上界, 从而保证了算法的可行性。

4 仿真实例

图 1所示, IEEE 39节点系统被分为4个不重叠的子系统, 每个子系统分别包含10、10、10和9个节点, 并且划分后的图是无环的, 这满足本文的假设条件。划分后的图的直径是2, 文献[19]中的定理1已经证明该算法在每个时刻迭代2次后即可收敛, 所以为了得到Σi* (k|k), 在仿真时每个时刻迭代2次。

图 1 39节点系统的划分 Figure 1 Partition of the IEEE 39-bus system

系统中每个节点s的系统参数如下:

$ {\mathit{\boldsymbol{A}}_s} = \left[ {\begin{array}{*{20}{c}} {0.84}&0\\ 0&{0.87} \end{array}} \right],{\mathit{\boldsymbol{C}}_s} = \left[ {\begin{array}{*{20}{c}} 1&0 \end{array}} \right],{\mathit{\boldsymbol{B}}_{ss'}} = {\mathit{\boldsymbol{B}}_{s's}} = \left[ {\begin{array}{*{20}{c}} {2.35}&{1.17}\\ { - 0.83}&{1.64} \end{array}} \right], $

式中s′∈Ns。并假设Rs=0.032I, Ts, s=0.012I, Ts, s=0.022I。将每个子系统所包含节点的参数结合起来即可得到子系统i的系统(1)(2)和(3)中的参数。

仿真的目的是根据该算法得到的误差协方差的迹来验证定理1和定理2的结论。误差协方差的迹分别表示为$\sum\limits_{i = 1}^4 {{\rm{Tr}}} \left\{ {\mathit{\boldsymbol{ \boldsymbol{\varSigma} }}_i^*\left( {k|k} \right)} \right\}({\rm{TEC}})$$\sum\limits_{i = 1}^4 {{\rm{Tr}}} \left\{ {{\Sigma _i}\left( {k + 1|k} \right)} \right\}({\rm{TPC}})$, 其上界分别表示为BEC和BPC。TEC和TPC随时间的变化曲线如图 2所示。由于子系统的最大维数是20, 并且定理2必须满足kni-1, 所以误差协方差的上界从k=19开始得到, 而之前有限时间的结果都是有界的。误差协方差的迹及其上界的变化曲线如图 3所示。从图 23可以看出, 该算法得到的误差协方差是正定有界的, 并且严格小于给定的上界, 验证了本研究的主要结果。

图 2 误差协方差的迹 Figure 2 The trace of error covariance
图 3 误差协方差的迹和上界的比较 Figure 3 Comparison of the trace of error covariance and upper bound
5 结语

本研究讨论的动态系统适用于很多实际应用, 例如电力系统、传感器网络、多智能体系统等。主要目的是继续研究文献[19]提出的分布式状态估计算法, 通过将限制条件减弱为每个子系统可观测, 证明得到的误差协方差矩阵是正定并且有上界的, 保证了该算法在实际应用中的有效性和可行性, 可作为电力系统动态状态估计的有效方法。IEEE 39节点系统的模拟仿真验证了本研究的主要结论。下一步将研究非线性的时变系统, 给出新的分布式状态估计算法, 并且对模型参数进行辨识, 进一步优化算法, 得到误差协方差矩阵有界的结论。

参考文献
[1] ABUR A, EXPOSITO A G. Power system state estimation: theory and implementation[M]. New York, USA: Marcel Dekker Press, 2004 .
[2] WU F F, MOSLEHI K, BOSE A. Power system control centers: past, present, and future[J]. Proceedings of the IEEE, 2005, 93 (11) : 1890-1908 DOI:10.1109/JPROC.2005.857499
[3] MONTICELLI A. Electric power system state estimation[J]. Proceedings of the IEEE, 2000, 88 (2) : 262-282 DOI:10.1109/5.824004
[4] SHIH K R, HUANG S J. Application of a robust algorithm for dynamic state estimation of a power system[J]. IEEE Transactions on Power Systems, 2002, 17 (1) : 141-147 DOI:10.1109/59.982205
[5] LIN J M, HUANG S J, SHIH K R. Application of sliding surface-enhanced fuzzy control for dynamic state estimation of a power system[J]. IEEE Transactions on Power Systems, 2003, 18 (2) : 570-577 DOI:10.1109/TPWRS.2003.810894
[6] DO COUTTO FILHO M B, DE SOUZA J C S. Forecasting-aided state estimation-part I: panorama[J]. IEEE Transactions on Power Systems, 2009, 24 (4) : 1667-1677 DOI:10.1109/TPWRS.2009.2030295
[7] LEITE D A SILVA A M, DO COUTTO FILHO M B, DE QUEIROZ J F. State forecasting in electric power systems[J]. IEE Proceedings-Generation, Transmission and Distribution, 1983, 130 (5) : 237-244 DOI:10.1049/ip-c.1983.0046
[8] HUANG S J, LIN J M. Enhancement of anomalous data mining in power system predicting-aided state estimation[J]. IEEE Transactions on Power Systems, 2004, 19 (1) : 610-619 DOI:10.1109/TPWRS.2003.818726
[9] WANG Shaobu, GAO Wenzhong, MELIOPOULOS A P S. An alternative method for power system dynamic state estimation based on unscented transform[J]. IEEE Transactions on Power Systems, 2012, 27 (2) : 942-950 DOI:10.1109/TPWRS.2011.2175255
[10] VALVERDE G, TERZIJA V. Unscented Kalman filter for power system dynamic state estimation[J]. IET Generation, Transmission & Distribution, 2011, 5 (1) : 29-37
[11] REGULSKI P, TERZIJA V. Estimation of frequency and fundamental power components using an unscented Kalman filter[J]. IEEE Transactions on Instrumentation and Measurement, 2012, 61 (4) : 952-962 DOI:10.1109/TIM.2011.2179342
[12] LI Yao, HE Xing, ZHANG Weidong. Applications of adaptive CKF algorithm in short-term load forecasting of smart grid[C]//Proceedings of the 33rd Chinese Control Conference (CCC). Nanjing, China:[s.n.], 2014: 8145-8149.
[13] 陈亮, 毕天姝, 李劲松, 等. 基于容积卡尔曼滤波的发电机动态状态估计[J]. 中国电机工程学报, 2014, 34 (16) : 2706-2713
CHEN Liang, BI Tianshu, LI Jinsong, et al. Dynamic state estimator for synchronous machines based on cubature Kalman filter[J]. Proceedings of the CSEE, 2014, 34 (16) : 2706-2713
[14] 李阳林, 卫志农, 万军彪. 一种新的分布式电力系统状态估计算法[J]. 继电器, 2007, 35 (20) : 13-22
LI Yanglin, WEI Zhinong, WAN Junbiao. A new algorithm for the distributed state estimation of power systems[J]. Relay, 2007, 35 (20) : 13-22
[15] SHARMA A, SRIVASTAVA S C, CHAKRABARTI S. Multi-agent-based dynamic state estimator for multi-area power system[J]. IET Generation, Transmission & Distribution, 2016, 10 (1) : 131-141
[16] FENG Jianxin, ZENG Ming. Optimal distributed Kalman filtering fusion for a linear dynamic system with cross-correlated noises[J]. International Journal of Systems Science, 2012, 43 (2) : 385-398 DOI:10.1080/00207721.2010.502601
[17] LANG Jinling, WANG Zidong, LIU Xiaohui. Distributed state estimation for discretetime sensor networks with randomly varying nonlinearities and missing measurements[J]. IEEE Transactions on Neural Networks, 2011, 22 (3) : 66-86
[18] MARELLI D, FU Minyue. Distributed weighted least-squares estimation with fast convergence for large-scale systems[J]. Automatica, 2015, 51 : 27-39 DOI:10.1016/j.automatica.2014.10.077
[19] SUN Yibing, FU Minyue, WANG Bingchang, et al. A distributed MAP approach to dynamic state estimation with applications in power networks[C]//Proceedings of the 2015 European Control Conference (ECC). Linz, Austria:[s.n.], 2015: 235-240. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=7330550
[20] MAKRIDAKIS S, WHEELWRIGHT S C. Forecasting methods and applications[M]. New York, USA: Wiley Press, 1978 .