defense slides - NCSU COE People
Transcription
defense slides - NCSU COE People
Measurement-Based Methods for Model Reduction, Identification, and Distributed Optimization of Power Systems Seyed Behzad Nabavi Department of Electrical and Computer Engineering North Carolina State University 24/4/2015 1/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Part I– Identification of Dynamic Reduced-Order Models of Power Systems 2/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Introduction Introduction Mathematical modeling of dynamic equivalents of large-scale electric power systems has seen some 40 years of long and rich research history. Chow and Kokotovic established the relationship between the slow coherency and weak connections using singular perturbation theory. Slow coherency arises from the slower inter-area modes. These interarea modes, Figure: R. Podemore: Coherency in if not properly damped, lead to system Power Systems separation and extensive loss of load. 3/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Introduction Model-based Dynamic Equivalencing Real-Time Monitoring Recent evidences Wide-Area of blackouts have shown the discrepancy between the offline models and the response of the system. W. Winter, K. Elkington, G. Bareux, and J. Kostevc, “Pushing the Limits: Europe's New Grid: Innovative Tools to Combat Transmission Bottlenecks and Reduced Inertia," Power and Energy Magazine, 13(1), 2015 4/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Introduction Model-based Dynamic Equivalencing Dynamic equivalencing has seen 40 years of active research: – – – – – linear modal decomposition [Undrill, 71] circuit-theoretic approaches [de Mello, 75] machine aggregation [Germond, 78] enumerative clustering algorithms [Zaborsky, 82] software programs such as DYNEQ and DYNRED [Price, 95] Model based methods: – need the exact knowledge of the entire power system model, – are computationally challenging, – are based on idealistic assumption about system structure and clustering. 5/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Introduction Measurement-based Dynamic Equivalencing PMUs provide high-resolution GPS-synchronized three-phase measurements of voltage, current, phasor, and frequency. System operators are, therefore, inclining more towards online models constructed from PMU (Phasor Measurement Unit) measurements. We next propose two algorithms to identify these dynamic equivalent models using PMU measurements: * Identification of the equivalent linear models * Identification of the equivalent nonlinear DAE models G8 G10 PMU 8 10 26 G4s 29 28 25 G1s 9 27 31 24 G9 G4s Ips4 Vps4 18 39 17 G1 32 16 6 PMU 15 14 33 1 34 20 7 30 PMU PMU 38 3 2 G2 Vps3 Ips3 G3s 23 19 13 11 V s Ips 1 p1 22 35 36 37 G1s G6 21 12 4 5 G3 G4 G7 G2s G3s G5 6/1 Vps2 I s p2 G2s Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Power System Swing Equation Nonlinear Electromechanical Model: δ̇i (t) = ωs (ωi (t) − 1), Mi ω̇i (t) = Pmi − Pei (t) − Di (ωi (t) − 1), Pe (t) + k X l∈Nk Pkl (t) − PL (t) = 0, Qe (t) + k k X l∈Nk Qkl (t) − QL (t) = 0 k Linearized Kron-Reduced Model (around (δi0 , 1)): " ∆δ̇(t) ∆ω̇(t) # " = 0n×n ωs In×n #" ∆δ(t) M −1 L −M −1 D ∆ω(t) | {z } A T ∆δ , ∆δ1 · · · ∆δn , # + Bd(t), T ∆ω , ∆ω1 · · · ∆ωn , ∆δ, ∆ω ∈ Rn M = diag(Mi ) ∈ Rn×n , D = diag(Di ) ∈ Rn×n , d(t) : unknown disturbance [L]i,j = Ei Ej (Gij cos(δi0 − δj0 ) − Bij sin(δi0 − δj0 )) i 6= j, [L]i,i = − n X [L]i,k , k =1 7/1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Linear Dynamic Equivalent Models G8 8 G10 10 PMU 26 G1s 29 28 9 G9 27 25 31 24 18 39 G1 17 32 16 6 PMU 15 14 33 1 21 22 12 34 23 19 35 36 13 11 37 20 3 G2 Area Area2 Area Area 7 30 PMU PMU 38 2 G6 G4s G7 4 5 G3 G4 G5 8/1 G2s Aggregated Transmission Network Graph G3s Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Identification of Linear Equivalent Models The reduced-order model: " ∆δ̇ s (t) ∆ω̇ s (t) # " = 0r ×r (M s )−1 Ls | ωs Ir ×r #" ∆δ s (t) # −(M s )−1 D s ∆ω s (t) {z } As + B s d(t). Assumptions: – The area partitioning for our system is known apriori. – There is at least one PMU at a generator bus in each area (S). – As and B s are a controllable pair. Objective: – Finding the equivalent linear model of a power system from yi (t) i ∈ S: yi (t) = {Ṽi (t), Ĩi,j (t)}, i ∈ S, j ∈ Ni . Proposed Identification Steps: – Extract ∆δks (t) for each area k from yi (t), i ∈ S. G1s G2s – Identification of As . 9/1 G4s Area Area2 Area Area Aggregated Transmission Network Graph G3s Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Extraction of ∆δks (t) for Each Area k Step 1: Extract ∆δi (t) from yi (t): Ei (t)∠δi (t) = jxd0 i Ii (t)∠φi (t) jxd i Ii Vi i Ei i + Vi (t)∠θi (t) ⇒ δ̂i (t) = ∠(jxd0 i Ii (t)∠φi (t) + Vi (t)∠θi (t)) ∆δ̂i (t) = δ̂i (t) − δ̂i (t0 ). Step 2: Extract ∆δ̂ks (t) from ∆δ̂i,k (t), (generator i belonging to area k ): 0 ∆δi,k (t) = ∆δi,k (t) + r −1 X ρil e (−σl +jΩl )t ∗ (−σl −jΩl )t + ρil e n−1 X ρil e (−σl +jΩl )t ∗ (−σl −jΩl )t , + ρil e l=r l=1 | + {z ∆δ s (t), inter-area or slow modes i,k } | {z } ∆δ f (t), intra-area or fast modes i,k Use a modal decomposition technique such as Prony to decompose ∆δi,k (t) s Form ∆δi,k (t) by retaining only the modes in [0.1,1] Hz. 10 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Extraction of ∆δks (t) for Each Area k G8 8 G10 10 PMU 26 G1s 29 28 9 G9 27 25 31 24 18 39 G1 17 32 16 6 PMU 15 14 33 1 21 22 12 34 23 19 35 36 13 11 37 20 PMU G2 G7 4 5 3 Area Area2 Area Area 7 30 PMU 38 2 G6 G4s G3 G4 G5 G2s Aggregated Transmission Network Graph G3s s We truncate ∆δi,k (t) to extract ∆δi,k (t). From the coherency assumption s s s ∆δ1,k (t) ≈ ∆δ2,k (t) ≈ · · · ≈ ∆δm (t) k ,k s We set ∆δks (t) = ∆δi,k (t). 11 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Identification of As The reduced-order model: " ∆δ̇ s (t) # ∆ω̇ s (t) " = ωs Ir ×r 0r ×r (M s )−1 Ls | #" ∆δ s (t) # −(M s )−1 D s ∆ω s (t) {z } As + B s d(t). Solve the following NLS problem (assuming d(t) is a momentary perturbation at t = t0 ): Z min s A tm k t1 s ∆δ s (t, As ) ∆δ̂ (t) 2 − k dt ∆ω s (t, As ) ∆ω̂ s (t) 2 where, s ∆δ s (t, As ) ∆δ̂ (t1 ) s = exp(A (t − t1 )) , ∆ω s (t, As ) ∆ω̂ s (t1 ) ∆ω̂ s (t) is calculated from the numerical differentiation of ∆δ̂ s (t) normalized by ωs . 12 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Identifiability Analysis of As Lemma: [Bellman and Astrom-70] Consider the system ẋ = Ax + Bu, y = Cx If the matrix C is full column-rank and the system is controllable, then A and B can be determined uniquely from input output data. In our identification problem, we assume (As , B s ) to be a controllable pair, and C = I2n (full column-rank), thus As is identifiable. More results on identifiability analysis will be provided in Part III (joint work with Dr. P. P. Khargonekar). 13 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction A Case Study– NPCC 48 Machine Model G34 G35 97 96 10 G33 92 1 G4 8 103 G36 91 11 31 102 110 112 94 89 123 83 116 84 G38 6 105 71 90 G24 68 119 G39 120 114 G47 122 13 118 G40 127 124 21 1 29 35 6 8 9 126 125 128 132 32 2 2 14 133 G44 16 130 G43 34 G27 81 75 7 76 77 74 G28 1 5 13 3 138 4 73 135 G46 12 4 31 39 129 G48 G2 30 G13 136 14 17 15 5 37 38 28 134 27 5 0 40 51 G14 42 69 3 6 16 G45 131 139 G9 41 4 G10 66 10 18 9 44 43 24 20 19 17 G18 45 70 64 63 67 137 15 G41 121 12 55 49 2 117 G8 56 G19 48G 46 G26 72 G25 G6 8 61 G11 47 11 G7 25 14 13 15 26 62 58 53 3 G23 106 113 88 85 87 G5 23 11 59 52 G16 107 111 G15 65 12 115 10 22 7 12 G22 86 G42 G37 G3 99 100 57 54 104 109 93 9 G20 60 G21 G17 95 108 G32 140 5 101 G 98 7 33 36 78 79 80 82 G29 G30 16 G1 17 14 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction A Case Study- NPCC 48 Machine Model 3 ∆δ1COI ∆δ1s model-based ∆δ8COI 10 6 ∆δ8s model-based 4 0 −1 (deg) 5 1 (deg) (deg) 2 0 COI ∆δ17 s model-based ∆δ17 2 0 −2 −5 −2 −4 −10 −3 0 5 10 15 0 5 Time (sec) 3 ∆δ1COI 10 −6 0 15 ∆δ1s Case 2 ∆δ8COI 10 10 15 Time (sec) 6 COI ∆δ17 ∆δ8s Case 2 2 s ∆δ17 Case 2 4 0 −1 (deg) 5 1 (deg) (deg) 5 Time (sec) 0 2 0 −2 −5 −2 −4 −10 −3 0 5 10 15 0 5 Time (sec) 10 Time (sec) 15 −6 0 5 10 15 Time (sec) Defining the error: Ja (k ) = 1 tm − t1 Z tm |∆δks ,reduced (t) − ∆δks ,actual (t)|dt. t1 P Pk Ja (k ) = 10.2232(deg) for the model-based method, and k Ja (k ) = 4.6017(deg) for our measurement-based method. 15 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Identification of the Equivalent DAE Models The linear equivalent models are in the Kron’s form. This model is not a very suitable choice for: 1 Identification of the individual equivalent parameters such as inertia (Mi ) 2 Shunt controller design purposes 3 Describing the system behavior for large disturbances (transient stability) Area 1 Area 4 G4s ` Ips4 Vps4 G1s Ips 1 Vps1 Vps3 Ips3 G3s Vps2 I s p2 Area 2 G2s Area 3 δ̇is (t) s s Mi ω̇i (t) δ̇i (t) = ωs (ωi (t) − 1), Mi ω̇i (t) = Pmi − Pei − Di (ωi (t) − 1), 16 / 1 = ωis (t) − ωs , s − Pesi − Dis (ωis (t) − 1), = Pm i Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Identification of the Equivalent DAE Models Assumptions: – The area partitioning for our system is known apriori. – The boundary buses of all areas are equipped with PMUs (denoted by S). Objective: – Finding the equivalent DAE model of a power system from yi (t) i ∈ S: yi (t) = {Ṽi (t), Ĩi,j (t)}, i ∈ S, j ∈ Ni . Proposed Identification Steps: G4s – Finding the equivalent pilot bus voltages and currents. – Estimating the equivalent area impedances. Ips4 Vps4 G1s V s Ips 1 p1 Vps3 Ips3 G3s – Estimating the equivalent generator parameters. – Estimating the inter-area impedances. Vps2 I s p2 G2s 17 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Equivalent Pilot Bus Voltage and Current Step 1: Use yk to calculate Ṽpk (t) and Ĩpk (t) P Ĩpk (t) , Ipk (t)∠φpk (t) = X Ĩi (t), Ṽpk (t) , Vpk (t)∠θpk (t) = i∈Bk i∈Bk Ṽi (t)Ĩi∗ (t) Ĩp∗k (t) PMU PMU Coherent Area k Step 1 Vpk Coherent Area k PMU 18 / 1 Ipk Step 2 Vpsk Coherent Area k Ipsk Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Equivalent Pilot Bus Voltage and Current Step 2: Construction of Ṽpsk (t) and Ĩpsk (t) The modal decomposition of δis (t): δis (t) ≈ 2r X ρjl eλl t + l=1 2r X 2r X ρ0jkl e(λl +λk )t ⇒ Vpsk (t) = 2r X αlk eλl t + l=1 k =1 l=1 2r X 2r X α0ijk e(λi +λj )t , i=1 j=1 Use Prony to decompose Vpk (t): Vpk (t) = N X βlk eγl t l=1 Retain only those modal components within the [0.1,1] Hz. The sum of these selected modal components are classified as Vpsk (t). Apply the same procedure to extract θpsk (t), Ipsk (t), and φspk (t) PMU PMU Coherent Area k Step 1 Vpk Coherent Area k PMU 18 / 1 Ipk Step 2 Vpsk Coherent Area k Ipsk Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Equivalent Area Impedance KVL in the equivalent circuit: Eks (t)∠δks (t) = (rks + jxd0sk )Ĩpsk (t) + Ṽpsk (t). For any time instance: Φ0 , |(rks + jxd0sk )(Îpsk (t0 )∠φ̂spk (t0 )) + V̂psk (t0 )∠θ̂psk (t0 )|, .. . Φm , |(rks + jxd0sk )(Îpsk (tm )∠φ̂spk (tm )) + V̂psk (tm )∠θ̂psk (tm )|. The estimation of rks and xd0sk can be posed as the following NLS problem: min var Φ0 , . . . , Φm , xd0s , rks k rks Eks ks Ipsk jxdsk Vpsk V psk psk 19 / 1 the kth equivalent pilot bus Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Estimating the equivalent generator parameters Solve the following NLS problem Z min s s s Mk ,Dk ,Pm k tm t0 |δ̂ks (t) − δks (t, Mks , Dks , Pms k )|2 dt, where δ̇ks (t) s s Mk ω̇k (t) = ωks (t) − ωs , s = Pm − Pesk − Dis (ωks (t) − 1), k δks (t) =δks (t0 ), ωks (t) = ωks (t0 ), Pesk (t) = Re Êks (t)∠δ̂ks (t) Ĩps∗ (t) k rks Eks ks Ipsk jxdsk Vpsk V psk psk 20 / 1 the kth equivalent pilot bus Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Estimating the inter-area impedances KCL on equivalent pilot buses: Y s Ṽ s (t0 ) | · · · | Ṽ s (tm ) = Ĩ s (t0 ) | · · · | Ĩ s (tm ) , {z } | {z } | Ṽ s Ĩ s s Estimate Y by solving: min kY s Ṽ s − Ĩ s k2F , s Y s.t. Y s = (Y s )T G4s Ips4 Vps4 G1s V s Ips 1 p1 Vps3 Ips3 G3s Vps2 I s p2 G2s 21 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction A Case Study– IEEE 39 Bus Model .0074+j.0268 9 D4s 1.6541M 4s Area 4 29 30 22 V ps2 23 20 PMU 4 2 +j D3s 0 −3 −3 x 10 H 3s 267.2335 G2s D2s 0.4417 M 2s G4 −3 1.5 H 2s 82.3485 G7 G5 G3 G2 Area 2 7 5 3 2 0. 05 01 19 35 37 38 .0100+j.0267 .0037+j.0451 13 j 0.0815 24 11 1 59 .0 j0 G1s 6 34 36 -0.0301 -j 1.1996 j0.0058 21 12 G3s V ps3 7 G6 15 14 33 G1 V ps1 16 97 17 32 .1 j0 39 1+ D1s 0.3227 M 1s 18 PMU 1 V ps4 05 Area 3 PMU 27 H1s 510.6557 j0 .0 81 7 26 .0 -0 31 x 10 1 x 10 1 0.5 (rad/sec) (rad/sec) 1 0 −0.5 ω̂2s − ω̂1s ω2s − ω1s −1 −1.5 0.5 1 1.5 Time (sec) 2 2.5 (rad/sec) Area 1 PMU 28 0. 00 76 + 25 PMU 17 8 G10 10 H 4s 106.6757 G4s G9 0. 00 G8 0 −1 −2 0.5 1.5 Time (sec) 22 / 1 2 0 −0.5 ω̂3s − ω̂1s ω3s − ω1s 1 0.5 2.5 −1 0.5 ω̂4s − ω̂1s ω4s − ω1s 1 1.5 2 2.5 Time (sec) Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identification of Dynamic Reduced-Order Models of Power Systems Power System Model Reduction Future Work Investigating the utility of the reduced order models for shunt controller design purposes (such as Static Var Compensator (SVC)). Vpsk (t ) Gks Ipsk (t ) Control Inversion Coherent Area k SVC SVC 23 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Part II– Distributed Optimization Algorithms for Wide-Area Oscillation Monitoring in Power Systems 24 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Introduction In Part I, we describe methods to identify the equivalent models from PMU. In Part II, we use PMUs to identify the (inter-area) oscillation modes from PMUs in a distributed way. Majority of modal estimation algorithms are centralized such as: Eigenvalue Realization Algorithm (ERA) [Sanchez-Gasca-99], Prony analysis [Hauer-90], Robust Least Squares [Zhuo-08], and Hilbert-Huang transform [Messina-06]. Figure: http://www.eia.gov/ As the number of PMUs scales up into the thousands, the current state-of-the art centralized architectures will no longer be sustainable. 25 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Wide-Area Oscillation Monitoring Using PMU measurements to estimate the frequency, damping factor and residue of the different electro-mechanical oscillation modes G14 PMU G1 66 40 41 48 47 42 31 38 51 49 G11 52 G16 36 64 37 44 68 43 G13 53 68 6 11 54 19 10 0 5 Time (sec) 23 59 G6 PMU G2 55 G3 -0.4 -0.5 24 22 58 12 56 G4 -0.3 0.3 0.2 14 13 5 62 G12 0.4 21 7 35 45 39 27 17 15 4 8 61 G9 26 18 9 63 33 34 50 PMU 25 28 16 PMU G10 46 2 29 G8 56 G15 60 1 30 3 32 62 67 PMU 53 0 5 Time (sec) 26 / 1 G7 20 57 G5 0.24 0.22 0.2 0.18 0.16 0 5 Time (sec) Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Wide-Area Oscillation Monitoring Wide-Area Oscillation Monitoring Using •PMU measurements to estimate the the frequency, damping Using PMU measurements to estimate frequency, damping factor and residue ofand theresidue different electro-mechanical oscillation modes of the electro-mechanical oscillation modes State-of-the-Art Monitoring Architecture The Proposed Distributed Monitoring Architecture 26 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Problem Formulation Oscillation Monitoring 0 1 1 1 2 2 1 1 2 ⋯ ⋯ 2 0 E1 1 G1 1 1 1 1 2 2 1 2 2 ⋯ ⋯ ~ I L1 V11 E2 2 G2 En n Gn Vn n V2 2 ~ I L2 ~ I L ,n 0 1 Vi i Gi Vn 1 n 1 0 1 1 1 1 2 2 2 2 ⋯ ⋯ 1 2 2 2 2 ⋯ ⋯ G n 1 Ei i 1 1 1 1 En1 n1 ~ I Li ~ I L , n 1 0 1 1 1 1 27 / 1 1 2 2 2 2 ⋯ ⋯ Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Problem Formulation Centralized Prony Method Step 1. Find a1 through a2n ∆θi (2n − 1) ··· ∆θi (2n) ∆θi (2n) ··· ∆θi (2n + 1) = . . . . . . ∆θi (2n + `) ∆θi (2n + ` − 1) · · · {z } | {z | ci Hi ∆θi (0) ∆θi (1) . . . ∆θi (`) } −a1 −a2 . . . −a2n | {z } a Finding the global a using all available measurements by solving: c1 H1 . . . = . a . . Hp cp | {z } Solve this using Batch Least Squares - Centralized Prony Method Step 2. Find the eigenvalues of A (i.e., −σi ± jΩi ) by – Finding the roots of discrete-time transfer function (z1 through z2n ) – Converting them from discrete-time to continuous-time 28 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Problem Formulation Centralized Prony Method θi → (Hi , ci ), i = 1, . . . , p H1 c1 ⇒ ... a = ... PMU G14 41 47 38 67 31 62 51 63 18 G11 35 45 50 39 36 64 37 44 68 43 G13 7 62 G12 6 12 11 10 G2 27 17 21 14 13 5 54 H1 c1 1 ⇒ a = arg min k ... a − ... k22 a 2 Hp cp G9 15 4 55 G3 19 56 G4 cp 61 26 25 9 8 33 34 52 G16 28 60 16 G10 46 49 32 29 G8 53 2 1 30 3 42 G15 Hp G1 48 66 40 24 23 22 58 59 G6 G7 20 57 G5 29 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Problem Formulation Distributing the Prony Method N Computational Areas: ∆θj,i : PMU i in area j ∆θj,i → Hj,i , cj,i T T T Ĥj , [Hj,1 Hj,2 · · · Hj,N ]T , j G14 48 66 40 41 47 31 62 38 67 51 G11 35 45 39 52 G16 18 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 Nj : is the total number of PMUs in Area j, 27 17 15 4 7 61 G9 26 25 9 8 33 34 50 68 28 60 16 63 G10 46 49 32 29 G8 53 2 1 30 3 42 G15 T T T · · · cj,N ]T cj,2 ĉj , [cj,1 j G1 55 G3 19 56 G4 24 Global Consensus Problem: 23 22 58 59 G6 minimize G7 20 PN a1 ,...,aN ,z 57 G5 1 i=1 2 kĤi ai − ĉi k22 subject to ai − z = 0, for i = 1, . . . , N Use Alternating Direction Method of Multipliers (ADMM) to solve it 30 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Problem Formulation Distributing the Prony Method Three Distributed Cyber-Physical Architectures (Using ADMM): Standard ADMM – Asynchrnous ADMM G14 G1 48 66 40 41 47 42 31 62 38 67 G15 51 63 18 9 G11 35 45 39 36 64 37 44 43 G13 7 62 G12 6 12 11 10 G2 Distributed ADMM 27 17 21 14 13 5 54 61 G9 15 4 8 33 34 50 G16 28 26 16 52 68 25 1 30 3 32 G10 46 49 60 2 Hierarchical ADMM 29 G8 53 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 30 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using Standard ADMM (S-ADMM) minimize a1 ,...,aN ,z PN 1 j=1 2 kĤj aj Area 1 − ĉj k22 Area 3 PMU subject to aj − z = 0, for j = 1, . . . , N θ11(t) PMU θ12(t) θ31(t) ak k 1 a ak Central PDC at ISO k 2 a ak θ22(t) θ21(t) θ32(t) k 3 a a4k ak θ42(t) θ41(t) PMU PMU Area 2 Area 4 Augmented Lagrangian: Lρ = N X ρ 1 ( kĤj aj − ĉj k2 + wjT (aj − z) + kaj − zk2 ), 2 2 j=1 aj , z: the primal variable wj : the dual variable ρ: penalty factor 31 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) G14 G1 48 66 40 41 47 42 31 62 38 67 G15 49 51 G11 35 45 39 52 G16 32 / 1 28 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 50 68 25 16 63 G10 46 60 2 1 30 3 32 29 G8 53 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. G14 G1 48 66 40 41 47 42 31 62 38 67 G15 51 G11 35 45 39 52 G16 32 / 1 28 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 50 68 25 16 63 G10 46 49 60 2 1 30 3 32 29 G8 53 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. The central PDC receives ajk +1 G14 from all PDCs. G1 48 66 40 41 31 62 38 67 G15 51 25 28 35 45 52 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 G11 39 G16 60 2 16 63 33 34 50 68 29 G8 53 1 30 3 32 G10 46 49 32 / 1 47 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. The central PDC receives ajk +1 G14 from all PDCs. G1 31 62 38 67 The central PDC calculates z k +1 = 1 N PN k +1 . j=1 aj 48 66 40 41 G15 49 51 52 G16 25 28 G11 35 45 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 39 68 60 2 16 63 G10 46 29 G8 53 1 30 3 32 50 32 / 1 47 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. The central PDC receives ajk +1 G14 from all PDCs. G1 31 62 38 67 The central PDC calculates z k +1 = 1 N PN k +1 . j=1 aj 48 66 40 41 G15 51 32 / 1 52 G16 25 28 G11 35 45 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 39 68 60 2 16 63 G10 46 49 29 G8 53 1 30 3 32 50 The central PDC sends z k +1 to local PDCs. 47 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. The central PDC receives ajk +1 G14 from all PDCs. G1 31 62 38 67 The central PDC calculates z k +1 = 1 N PN k +1 . j=1 aj 48 66 40 41 G15 49 51 52 G16 25 28 G11 35 45 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 39 68 60 2 16 63 G10 46 29 G8 53 1 30 3 32 50 The central PDC sends z k +1 to local PDCs. PDC j calculates wjk +1 as 47 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 wjk +1 = wjk + ρ(ajk +1 − z k +1 ) 32 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using S-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 to the central PDC. The central PDC receives ajk +1 G14 from all PDCs. G1 31 62 38 67 The central PDC calculates z k +1 = 1 N PN k +1 . j=1 aj 48 66 40 41 G15 49 51 52 G16 25 28 G11 35 45 36 64 37 44 43 G13 62 G12 21 14 13 5 6 12 11 54 10 G2 27 17 15 4 7 61 G9 26 18 9 8 33 34 39 68 60 2 16 63 G10 46 29 G8 53 1 30 3 32 50 The central PDC sends z k +1 to local PDCs. PDC j calculates wjk +1 as 47 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 wjk +1 = wjk + ρ(ajk +1 − z k +1 ) The central PDC and local PDCs find the eigenvalues −σi ± jΩi using z k . 32 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using A-ADMM Iteration k Each PDC updates aj locally ajk +1 = ((Hjk )T Hjk + ρI)−1 ((Hjk )T cjk − wjk + ρz k ) PDC j sends ajk +1 and wjk to the central PDC. ajk +1 The central PDC receives S k , a subset of PDCs. The central P PDC calculates k +1 z j∈ / = N1 S : ajk +1 and wjk from G1 48 66 40 31 62 38 67 G15 51 52 G16 60 2 25 28 9 35 45 36 64 37 44 43 G13 7 21 14 13 5 6 12 11 54 62 G12 10 G2 27 17 15 4 8 G11 61 G9 26 18 16 63 33 34 39 68 29 G8 53 1 30 3 32 50 ajk +1 + ρ1 wjk ajk , wjk = wjk −1 47 G10 46 49 N j=1 = G14 41 42 55 G3 19 56 G4 24 23 22 58 59 G6 G7 20 57 G5 The central PDC sends z k +1 to local PDCs. PDC j calculates wjk +1 as wjk +1 = wjk + ρ(ajk +1 − z k +1 ), j ∈ S k , 33 / 1 wjk +1 = wjk , j ∈ / Sk Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using H-ADMM Less communication and computation overhead for the central PDC for large number of PDCs. Area 3 Area 1 θ11a(t) The same convergence properties as the S-ADMM. θ21a(t) θ11b(t) a1k ak θ12c(t) θ22c(t) θ22a(t) θ12b(t) Area 2 34 / 1 θ13a(t) θ23a(t) θ13b(t) ak ak θ11c(t) θ21c(t) θ12a(t) θ21b(t) a2k θ22b(t) θ13c(t) θ23c(t) a3k Central PDC at ISO a4k θ14a(t) θ23b(t) ak θ14c(t) θ24c(t) θ24a(t) θ14b(t) θ24b(t) Area 4 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Distributed Prony using D-ADMM Define a communication graph G(V , E). Area 1 θ11(t) A different version of the original problem defined over G: minimize a1 ,...,aN ,z a1k a3k w13k w12k − ĉj k22 a4k a1k a3k k 14 w a4k k w34 Communication Graph G subject to aj − ak = 0, for jk ∈ E(G) θ21(t) Modified Augmented Lagrangian: L0k ρ = θ32(t) θ31(t) a2k a 1 j=1 2 kĤj aj PMU θ12(t) k 1 PN Area 3 PMU θ41(t) θ22(t) θ42(t) PMU PMU Area 2 Area 4 N X 1X 1 kĤjk aj − ĉjk k2 + ρ( kavk +1 − aj − wvjk k2 + 2 ρ j=1 X v ∈Sj v ∈Pj X 1 1 X kaj − avk − wjvk k2 ) − ( kwvjk k2 + kwjvk k2 ) , ρ ρ v ∈Pj 35 / 1 v ∈Sj Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM One of the challenges of using any distributed computational architecture is ensuring their resiliency to node attacks in the form of data manipulation. It is difficult for the ISO to detect a manipulated set of measurement broadcasting from a malicious local PDC. The D-ADMM architecture has the advantage that the primal and dual updates are done by local PDCs. Let us define the following residual errors: Ejk , kĤj ajk − ĉj k, Ejlk , kĤj alk − ĉj k, ∀l ∈ Nj let us consider G to be a cycle. 36 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 37 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 2 PDC j updates aj as ajk +1 = arg min L0ρ . aj 3 PDC j updates all wlj for l ∈ Pj : wljk +1 37 / 1 = wljk − ρ(alk +1 − ajk +1 ). Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 2 PDC j updates aj as ajk +1 = arg min L0ρ . aj wljk +1 = wljk − ρ(alk +1 − ajk +1 ). 3 PDC j updates all wlj for l ∈ Pj : 4 PDC j sends ajk +1 to all l ∈ Pj ∪ Sj , and receives alk +1 from l ∈ Sj . 37 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 2 PDC j updates aj as ajk +1 = arg min L0ρ . aj wljk +1 = wljk − ρ(alk +1 − ajk +1 ). 3 PDC j updates all wlj for l ∈ Pj : 4 PDC j sends ajk +1 to all l ∈ Pj ∪ Sj , and receives alk +1 from l ∈ Sj . 5 PDC j updates all wjl for l ∈ Sj : wjlk +1 = wjlk − ρ(ajk +1 − alk +1 ). 6 PDC j calculates Ejk and Ejlk for l ∈ Pj ∪ Sj . 37 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 2 PDC j updates aj as ajk +1 = arg min L0ρ . aj wljk +1 = wljk − ρ(alk +1 − ajk +1 ). 3 PDC j updates all wlj for l ∈ Pj : 4 PDC j sends ajk +1 to all l ∈ Pj ∪ Sj , and receives alk +1 from l ∈ Sj . 5 PDC j updates all wjl for l ∈ Sj : wjlk +1 = wjlk − ρ(ajk +1 − alk +1 ). 6 PDC j calculates Ejk and Ejlk for l ∈ Pj ∪ Sj . 7 If log(Ejlk ) − log(Ejk ) > ET for any l ∈ Pj ∪ Sj , PDC j reports an alert about node j to the ISO. 37 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Distributed Prony Methods Resilient Distributed Prony using D-ADMM 1 Each PDC j receives the update of alk +1 for all l ∈ Pj . 2 PDC j updates aj as ajk +1 = arg min L0ρ . aj wljk +1 = wljk − ρ(alk +1 − ajk +1 ). 3 PDC j updates all wlj for l ∈ Pj : 4 PDC j sends ajk +1 to all l ∈ Pj ∪ Sj , and receives alk +1 from l ∈ Sj . 5 PDC j updates all wjl for l ∈ Sj : wjlk +1 = wjlk − ρ(ajk +1 − alk +1 ). 6 PDC j calculates Ejk and Ejlk for l ∈ Pj ∪ Sj . 7 8 If log(Ejlk ) − log(Ejk ) > ET for any l ∈ Pj ∪ Sj , PDC j reports an alert about node j to the ISO. If the ISO gets an alert for PDC j from all PDCs belonging to Pj ∪ Sj for K iterations, it removes PDC j, rearranges a new communication graph G 0 with the remaining PDCs, and continues the iterations. 37 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Simulation Results A case study for the IEEE 68 bus model, Area 3 Area 2 G14 PMU 66 41 PMU 48 40 68 bus, 16 generators Area 1 PMU G8 G1 53 47 60 PMU 26 2 42 PMU 32 3 51 9 15 21 37 44 68 43 G16 5 13 23 22 36 6 PMU 54 64 65 G13 Area 4 24 14 7 35 45 PMU 39 PMU 27 17 4 G11 50 G9 16 8 33 34 52 PMU 18 63 G10 46 49 30 PMU PMU G15 1 PMU 31 62 38 67 25 5 computational areas 29 61 28 12 58 11 10 19 56 59 G6 G 7 PMU G12 PMU G2 55 G3 G4 Area 5 20 57 PMU G5 38 / 1 A three-phase fault is considered occurring at the line connecting buses 1 and 2. The fault starts at t = 0.1 sec, clears at bus 1 at t = 0.15 sec and at bus 2 at t = 0.20 sec, Ts =0.2 seconds. Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Simulation Results 0.3 0.2 0.6 Ω1 Ω2 Ω3 Ω4 6 5 10 20 30 Iteration (k) 40 2 50 0.4 0.3 0 10 20 30 Iteration (k) 40 0.2 50 0 10 20 30 Iteration (k) 0.4 0.3 40 50 2 0 10 20 30 Iteration (k) 40 50 Ω1 Ω2 Ω3 Ω4 6 5 10 0 S-ADMM A-ADMM D-ADMM (G1 ) D-ADMM (G2 ) H-ADMM 10 -6 E σ 0.5 4 Figure: H-ADMM 7 Ω (rad/sec) σ1 σ2 σ3 σ4 5 3 Figure: S-ADMM 0.6 Ω1 Ω2 Ω3 Ω4 6 4 3 0 7 σ1 σ2 σ3 σ4 0.5 Ω (rad/sec) σ 0.4 7 σ σ1 σ2 σ3 σ4 0.5 Ω (rad/sec) 0.6 4 3 0.2 0.1 0 10 20 30 Iteration (k) 40 50 2 10 -12 0 10 20 30 Iteration (k) 40 50 0 5 10 15 20 25 30 Iteration (k) 35 40 45 50 Figure: A-ADMM 39 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Simulation Results (D-ADMM) with Attack ET = 8, K = 20. 1010 100 10-10 5 10 15 20 Iteration (k) 10-20 25 -5 10 Error E12 E15 E1 10-10 30 40 Iteration (k) 50 10 15 20 Iteration (k) 10-20 25 10 E24 E21 E2 10-10 10-15 30 40 Iteration (k) 10-10 5 10-20 25 10-10 30 40 Iteration (k) 5 10 15 20 Iteration (k) 10-5 E45 E42 E4 10-15 50 10 15 20 Iteration (k) -5 50 25 E51 E54 E5 10-10 10-15 30 40 Iteration (k) 50 0.7 σ1 σ2 σ3 σ4 0.6 Ω1 Ω2 Ω3 Ω4 7 6 0.5 5 Ω 10-15 5 -5 E51 E54 E5 100 10-10 σ Error G 10 10-10 Error 10-20 after detection 1010 E45 E43 E4 100 Error 100 1010 E23 E21 E2 PDC 5 Error E12 E15 E1 Error G PDC 4 Error 1010 PDC 2 Error PDC 1 before detection 4 0.4 3 0.3 2 0.2 10 20 30 Iteration (k) 40 50 40 / 1 10 20 30 Iteration (k) 40 50 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Conclusions Development of distributed algorithms is imperative considering the increasing number of PMUs in power systems. We consider the problem of estimating the frequencies and damping factors of oscillation modes using Prony method in a distributed way. We proposed three cyber-physical architecture for implementing the distributed Prony algorithm using several versions of ADMM. The results of the case studies verify that the distributed solution for the oscillation modes converges to the centralized solution. Using a heuristic cross verification method we showed how a malicious data manipulation can be detected and isolated. 41 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Future Work Investigating the resiliency of the proposed algorithms under more complicated attack scenarios. (Joint work with Jianhua Zhang) Incorporating the asynchronous wide-area communications considering the delay traffic models in both uplink and downlink: Z t P(t) = φ(s)ds = −∞ 1 µ t −µ [erf( √ ) + erf( √ )]+ 2 2σ 2σ (p − 1) ( 1 λ2 σ2 +µλ) −λt λσ 2 + µ t − λσ 2 − µ √ e 2 e [erf( √ ) + erf( )]. 2 2σ 2σ Change the update strategy for downlink (needs convergence proof) wik = wik −1 + ρ(aik − (z k −1 + γ(z k −1 − z k −2 ))), 42 / 1 i∈ / S2k Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Part III– Graph-Theoretic Identifiability Analysis of Weighted Consensus Networks 43 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Preliminaries Consider the following single-input consensus model defined over a graph G(V , E, W ): ẋi (t) = X wij xj (t) − xi (t) + bi u(t), i = 1, . . . , n j∈Ni Defining x = x1 x2 ··· xn T ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0, x ∈ Rn , L = −L ∈ Rn×n , B ∈ Rn×1 , W = {wij , ∀ i, j} [L]i,j P −wi,j wi,k = k ∈Ni 0 44 / 1 i∼j i=j otherwise Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Distinguishability/Identifiability ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 Consider two distinct parameter sets W and W 0 (W 6= W 0 ). These two sets are called indistinguishable if the respective models cannot produce different outputs y(t) for any given input, i.e., y(t, W ) = y(t, W 0 ). 45 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Distinguishability/Identifiability ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 Consider two distinct parameter sets W and W 0 (W 6= W 0 ). These two sets are called indistinguishable if the respective models cannot produce different outputs y(t) for any given input, i.e., y(t, W ) = y(t, W 0 ). u1 (t ) 2.5 7 2 1 2.5 7 2 y1 (t ) u2 (t ) 2 7 2 2 2 7 2 y2 (t ) Y1 (s) Y2 (s) 4.5 = = 4 U1 (s) U2 (s) s + 12s3 + 33s2 + 18s 45 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Distributed Oscillation Monitoring Simulation Results Distinguishability/Identifiability ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 Consider two distinct parameter sets W and W 0 (W 6= W 0 ). These two sets are called indistinguishable if the respective models cannot produce different outputs y(t) for any given input, i.e., y(t, W ) = y(t, W 0 ). If W and W 0 are not indistinguishable, they are distinguishable. A parameter set W is said to be globally identifiable if for all W 0 6= W , W and W 0 are distinguishable. 45 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis Identifiability in terms of Markov Parameters ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 Lemma(Grewel,1976): The parameter sets W and W 0 are indistinguishable if and only if CL` (W )B = CL` (W 0 )B, ` ≥ 0. W is identifiable if and only if the mapping from W to the Markov parameters is injective (one-to-one). 46 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis Our Proposition There are analytical results for identifiability analysis of generic dynamic models (as early as 70’s and 80’s). Generally, investigating the parameter identifiability for medium and large-scale systems is a difficult and intractable task. We develop a simple sensor placement algorithm to guarantee identifiability of the edge-weights W for consensus networks defined over a class of graphs. We integrate the results from the graph theory with these classical results of identifiability. 47 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis Preliminaries Let us consider a rooted graph G with the root being the input node (node indexed by 1). Let us partition V into the following sets: Si = {v ∈ V : d(v , 1) = i}, i = 0, 1, . . . , p. S0 S1 S2 1 2 S 2v1 v S22 48 / 1 S 2v3 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Studied Class of Graphs Assumption 1: For a rooted graph G, nodes vl , vq , vs ∈ Si , vj ∈ Si+1 , ∀ i ≥ 1 satisfy the following properties: vl ∼ vj , vq ∼ vj ⇒ vl = vq , (vl ∈ Sil ) ∼ (vj ∈ Sij ) ⇒ l = j, dim({qv ∈ E(G) | q, v ∈ Sij }) ≤ 1, ∀ i, j 49 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Studied Class of Graphs Assumption 1: For a rooted graph G, nodes vl , vq , vs ∈ Si , vj ∈ Si+1 , ∀ i ≥ 1 satisfy the following properties: vl ∼ vj , vq ∼ vj ⇒ vl = vq , (vl ∈ Sil ) ∼ (vj ∈ Sij ) ⇒ l = j, dim({qv ∈ E(G) | q, v ∈ Sij }) ≤ 1, ∀ i, j Assumption 2: W is identifiable if C = In . Parameter b is not identifiable regardless of choice of C. u (t ) y1 (t ) y2 (t ) 49 / 1 y3 (t ) Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis Two Supporting Lemmas Lemma 1– The following holds for L of a G satisfying Assumption 1: 0 0 ≤ k ≤ d(v , 1) − 1 [Lk ]v ,1 = W(Pv ,1 ) k = d(v , 1) Pv ,1 is the unique path of length d(v , 1) connecting nodes v and 1. W(Pv ,1 ) is the weight of path Pv ,1 : W(P) = Y we e∈P Proof: By strong induction on k . 50 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis Two Supporting Lemmas Lemma 2– Consider a node indexed as v in G and its neighboring nodes denoted by v1 , . . . , vs . Let L = −L, where L is the weighted Laplacian matrix of G. If H denotes a subgraph of G induced by the set of all edges incident to v , and VH and WH denote the vertex set and the weights of all edges belonging to H respectively, then [Li ]vs ,1 can be uniquely computed from WH and [Li ]m,1 , (m ∈ VH \{vs }), ∀ i ≥ 1. vs is called an available node. 51 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm The Proposed Algorithm Start with S0 and place a sensor at this node. 52 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm The Proposed Algorithm Start with S0 and place a sensor at this node. for k = 1 : p for each set of siblings Skj choose any |Skj | − 1 nodes belonging to Skj and place sensors at them. 52 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm The Proposed Algorithm Start with S0 and place a sensor at this node. for k = 1 : p for each set of siblings Skj choose any |Skj | − 1 nodes belonging to Skj and place sensors at them. for each neighboring siblings q, v ∈ Skj −1 q if Sk and Skv are both non-empty q place an additional sensor in either Sk or Skv . 52 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm The Main Theorem Consider the following model with G satisfying Assumptions 1 and 2. ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 If S ⊂ V is a set of sensor nodes determined by the proposed algorithm, y(t) is the corresponding output measured by S, and H(W ) is the transfer function from u(t) to y(t), then the mapping from the W to H(W ) is one-to-one. 53 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm The Main Theorem Consider the following model with G satisfying Assumptions 1 and 2. ẋ(t) = L(W )x(t) + Bu(t), y(t) = Cx(t), x(0) = 0 If S ⊂ V is a set of sensor nodes determined by the proposed algorithm, y(t) is the corresponding output measured by S, and H(W ) is the transfer function from u(t) to y(t), then the mapping from the W to H(W ) is one-to-one. Proof: Wj−1,j , {wu,v ∈ W | u ∈ Sj−1 , v ∈ Sj }, Wj,j , {wu,v ∈ W | u, v ∈ Sj }, j = 1, . . . , p.Qj , CLj B, j = 1, 2, . . . , p. The proof follows from strong induction on j. In each step we proof the S2n−1 injective mapping of (Wj−1,j , Wj−1,j−1 ) to i=1 Qi . 53 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Identifiability Analysis The Proposed Sensor Placement Algorithm More Results Proposition 1: If the proposed algorithm is applied to a rooted-tree T , then the number of placed sensors is equal to the number of non-input leaves of T , i.e., the set of leaves that are not the input node. Proposition 2: If T is a star-graph, then the minimum number of sensors to identify W is (n − 2). 54 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Summary and Conclusions Summary and Conclusions We investigate the identifiability problem in Laplacian consensus NDS. We translate the classical results of identifiability in terms of graph properties. We propose a sensor placement algorithm for a class of graphs. We prove that, our algorithm provides a sufficient condition of identifiability of the edge weights. 55 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Publications Publications (Published/Accepted) Book Chapter: B1 S. Nabavi, J. Zhang, and A. Chakrabortty. Distributed Algorithms for Wide-Area Monitoring in Power Systems: A Cyber-Physical Perspective. Invited Chapter for CyberPhysical-Social Systems and Constructs in Electric Power Engineering, IET, 2015. Journal Articles: J3 S. Nabavi and A. Chakrabortty, “A Graph-Theoretic Condition for Global Identifiability of Weighted Consensus Networks," IEEE Transactions on Automatic Control, (conditionally accepted). J2 S. Nabavi, J. Zhang, and A. Chakrabortty, “Distributed Optimization Algorithms for Wide-Area Oscillation Monitoring in Power Systems Using Inter-Regional PMU–PDC Architectures," IEEE Transactions on Smart Grid, vol: p, no: pp, 2015. J1 T. R. Nudell, S. Nabavi, and A. Chakrabortty, “A Real-Time Attack Localization Algorithm for Large Power System Networks Using Graph-Theoretic Techniques," IEEE Transactions on Smart Grid, vol: p, no: pp, 2015. Conference Proceedings: C5 J. Zhang, S. Nabavi, A. Chakrabortty, and Y. Xin, “Convergence Analysis of ADMM-Based Power System Mode Estimation Under Asynchronous Wide-Area Communication Delays," IEEE PES General Meeting, 2015 (to be appeared). C4 S. Nabavi, A. Chakrabortty, and P. P. Khargonekar, “A Global Identifiability Condition for Consensus Networks with Tree Graphs," American Control Conference (ACC), 2015 (to be appeared). C3 S. Nabavi and A. Chakrabortty, “Distributed Estimation of Inter-area Oscillation Modes in Large Power Systems Using Alternating Direction Multiplier Method," IEEE PES General Meeting, National Harbor, MD, 2014. C2 S. Nabavi and A. Chakrabortty, “A Real-Time Distributed Prony-Based Algorithm for Modal Estimation of Power System Oscillations," American Control Conference (ACC), Portland, OR, 2014. C1 S. Nabavi and A. Chakrabortty, “Topology Identification for Dynamic Equivalent Models of Large Power System Networks," American Control Conference (ACC), Washington, DC, 2013. 56 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi Publications Publications (Under review/ To be submitted) Journal Articles: – J. Zhang, S. Nabavi, A. Chakrabortty, and Y. Xin, “ADMM Optimization Strategies for Wide-Area Oscillation Monitoring in Power Systems under Asynchronous Communication Delays" IEEE Transactions on Smart Grid - Special Issue on Theory of Complex Systems with Applications to Smart Grid Operations. – S. Nabavi and A. Chakrabortty, “Identification of Equivalent DAE Models of Power Systems using Synchrophasors". – S. Nabavi and A. Chakrabortty, “Identification of Dynamic Equivalent Models of Power Systems using Synchrophasors". Conference Proceedings (submitted): – S. Nabavi and A. Chakrabortty, “An Intrusion Resilient Distributed Optimization Algorithm for Modal Estimation in Power Systems", submitted to 2015 IEEE Conference on Control and Decision (CDC). 57 / 1 Model Reduction, Identification, and Distributed Optimization of Power Systems c 2015 by S. Nabavi