Chapter VI: Failure Mitigation: Adaptive Network Response
6.1 The Bridge-block Decomposition and the Unified Controller
The bridge-block decomposition and the unified controller have recently emerged as two important tools for grid reliability [3, 102, 103]. The two concepts operate on different timescales to improve the power system robustness: the bridge-block decomposition aims to localize the failure propagation, while the unified controller aims to stabilize a disturbed system. In this section, we review these concepts and elaborate on how they can be integrated as a novel control framework for failure localization and mitigation.
Bridge-block Decomposition
Given a power networkG =(N,E), apartitionofGis defined as a finite collection P = {N1,N2,· · ·,Nk} of nonempty and disjoint subsets ofN such thatÐk
i=1Ni = N. For a partition P, each edge can be classified as either a tie-line if the two endpoints belong to different subsets ofN or aninternal lineotherwise.
We define an equivalence relation on N such that two nodes are in the same equivalence class if and only if there are two edge-disjoint paths connecting them.
For this specific partition, the tie-lines connecting different components are exactly thebridges(cut-edges) of the graph. We thus refer to this partition asbridge-block
Figure 6.1: Bridge-block decomposition of a graph.
decompositionPBBof the power network (see Fig. 6.1 for an example).
We show in Chapter 3 that each graph has a unique bridge-block decomposition, which can be found in linear time. In particular, the bridge-block decomposition encodes rich information on failure propagation.
Unified Controller (UC)
Before we introduce the unified controller, we first summarize the integrated failure model proposed in Chapter 2. We describe the cascading failure process by the set of outaged linesB(n) ⊂ E over stagesn ∈ {1,2, . . . ,N}. For each stagen, the system evolves according to the fast-timescale dynamics on the topologyG(n):= (N,E(n)) whereE(n):=E \ B(n)and converges to an equilibrium point. When it eventually converges to an equilibrium, we compare the branch flow with the steady-state thermal capacityπefor the surviving transmission linee. Overloaded lines are then tripped and the cycle repeats, i.e.,B(n+1)= B(n) ∪ {e: |fe(n)| > πe,e∈ E(n)}. We adopt the linearized frequency dynamics as the fast-timescale dyanmics after line failures:
θÛj =ωj, j ∈ N (6.1a)
MjωÛj =rj +dj−Djωj−Õ
e∈E
Cjefe, j ∈ N (6.1b) fi j = Bi j(θi−θj), (i, j) ∈ E. (6.1c) UC is a control approach recently proposed in the frequency regulation literature [3,
41–43, 81]. Compared to classical droop control or AGC [82], UC simultaneously integrates primary control, secondary control, and congestion management on a fast timescale. The key feature of UC is that theclosed-loopequilibrium of (6.1) under UC solves the following optimization problem on thepost-contingency network:
θ,ω,d,minf
Í
j∈Ncj(dj) (6.2a)
s.t. ω= 0, (6.2b)
r +d−C f =0, (6.2c)
f = BCTθ, (6.2d)
EC f = 0, (6.2e)
fe ≤ fe ≤ fe, e∈ E, (6.2f)
dj ≤ dj ≤ dj, j ∈ N, (6.2g)
where cj(·)’s are associated cost functions that penalize deviations from the last optimal dispatch (and hence attain minimum at dj = 0), (6.2b) ensures secondary frequency regulation is achieved, (6.2c) guarantees power balance at each bus, (6.2d) is the DC power flow equation, (6.2e) enforces zero area control error [82], (6.2f) and (6.2g) are the flow and control limits. The matrix E encodes balancing area information as follows. Given a partition PUC = {N1,N2,· · · ,Nk} ofG that specifies the balancing areas for secondary frequency control,E ∈ {0,1}|PUC|×n is defined by El j = 1 if bus j is in balancing area Nl and El j = 0 otherwise. As a result, the l-th row of EC f = 0ensures that the branch flow deviations on the tie-lines connected to balancing areaNl sum to zero.
UC is designed so that its controller dynamics, combined with the system dynamics (6.1), form a variant of projected primal-dual algorithms to solve (6.2). It is shown in [3, 41–43, 81] that when the optimization problem (6.2) is feasible, under mild assumptions, the closed-loop equilibrium under UC is globally asymptotically stable and it is an optimal point of (6.2). Such an optimal point is unique (up to a constant shift ofθ) if the cost functionscj(·)are strictly convex. This means that, after a (cut or non-cut) failure, the post-contingency system is driven by UC to an optimal solution of (6.2) (under appropriate assumptions). We refer the readers to [3, 41–43, 81] for specific controller designs and their analysis.
Connecting UC and the Bridge-Block Decomposition
We have introduced two partitions of a power network: the bridge-block decompo- sitionPBBand the balancing area partitionPUC, which in general are different from
each other. However, when they do coincide, the underlying power grid inherits analytical properties from both bridge-block decomposition and UC, making the system particularly robust against failures. Our proposed control strategy leverages precisely this feature, as we present in Section 6.2.
In practice, the balancing areas over which UC operates are usually connected by multiple tie-lines in a mesh structure. However, in order to align with the bridge- block decomposition, we may have to switch off a few tie-lines of PUC. The selection of these tie-lines can be systematically optimized, e.g., to minimize line congestion or inter-area flows on the resulting network; see Section 6.2 for more details. We henceforth assume thatPBB= PUC. We refer to such a network as the tree-partitionednetwork since the balancing areas are connected in a tree structure prescribed by its bridge-block decomposition.
Definition 6.1. Given a cascading failure process described by B(n), with n ∈ {1,2, . . . ,N}, the set B(1) is said to be its initial failure. An initial failure B(1) is said to be criticalif the UC optimization (6.2) is infeasible over G(1) := (N,E\B(1)), ornon-criticalotherwise.
To formally state our localization result, we define the following concept to clarify the precise meaning of an area being “local” with respect to an initial failure.
Definition 6.2. Given an initial failureB(1), we say that a tree-partitioned balancing area Nl isassociatedwith B(1) if there exists an edgee = (i, j) ∈ B(1) such that eitheri ∈ Nlor j ∈ Nl.
As we discuss below, our control strategy possesses a strong localization property for both non-critical and critical failures in the sense that only the operation of the associated areas are adjusted whenever possible.