Impact of Data Quality on Real-Time Locational Marginal Price

I. INTRODUCTION

• Real-time LMP is computed ex post from SCADA-dependent data, so undetected errors or attacks that evade detection can affect prices relevant to possible demand response.The day-ahead market instead determines LMP from an OPF solution using generator offers, demand forecasts, topology, and security constraints.Source: §I · p. 1

A. Summary of Results and Organization

• The power-system state space is partitioned into polytope price regions, each associated with a unique real-time LMP vector and bounded by a set of congested lines.This geometric characterization provides a basis for analyzing how bad data affect real-time prices.Source: A. Summary of Results and Organization · p. 1
• Meter errors can shift the estimated state into another price region, whereas breaker or switch errors change the topology and price-region structure, sometimes changing prices despite little state change.The paper analyzes meter errors by seeking measurement errors that move the correct state estimate into the worst price region.Source: A. Summary of Results and Organization · pp. 1–2
• Three meter-data models span state-independent to fully adaptive corruption; the tractable worst-case search restricts congestion to prone lines and uses greedy algorithms.The adversarial framing allows the analysis to capture worst-case performance exactly in some cases and as bounds in others.Source: A. Summary of Results and Organization · pp. 1–2
• On IEEE-14 and IEEE-118 simulations, state-independent bad data had limited LMP impact, state-dependent data caused greater perturbations, and topology data increased them; meter-data impacts fell as networks grew at constant redundancy.The nonlinear state estimator may mitigate attacks designed using a DC model, while bad topology data combined with bad analog data can have greater LMP impacts.Source: A. Summary of Results and Organization · p. 2

B. Related Work

• Prior work leaves weak-attack impacts poorly understood and topology-error effects on electricity markets unreported; this study addresses worst cases, adaptive data, topology errors, and AC/nonlinear estimation.Its evaluation extends beyond a feasible price-changing case to a broader class of bad data and the AC network model with a nonlinear state estimator.Source: B. Related Work · p. 2

II. STRUCTURES OF REAL-TIME LMP

• The section models real-time pricing with a simplified ex-post incremental OPF and uses the estimated congestion pattern to connect measurements, state estimates, and prices.Once the congestion pattern is fixed, the incremental OPF no longer depends on measurement data.Source: §II · pp. 2–3
• Real-time LMP is the overall cost increase from adding one unit of load at a bus, with η and µ_k denoting the energy and line-flow dual variables.The model treats the congestion pattern as determined from state estimates.Source: §II · p. 3
• Each state-space polytope corresponds to a unique congestion pattern and real-time LMP, with boundary hyperplanes defined by individual transmission lines.This partition assumes an LMP exists for every possible congestion pattern.Source: §II · p. 3
• Meter-data errors must move the estimate into a different price region to alter real-time LMP.The analysis illustrates movement from the true region X0 to another region such as X3.Source: §II · pp. 3–4
• Topology errors can perturb the state estimate, change the price partition, and alter the prices attached to each region through a changed shift matrix.Even when the estimated state changes little, topology errors can substantially change region prices.Source: §II · p. 4

1) Meter data:

• Bad data represent deliberate or erroneous perturbations, while measurement noise represents random noise independently distributed across meters.Corruption is confined to a suspectable subset S; the feasible bad-data vectors are zero outside S.Source: 1) Meter data · p. 4
• The three meter-data models progress from state-independent corruption, to corruption adapted to the operating state, to fully adaptive corruption based on actual measurements.The fully adaptive model is described as the most powerful, but is generally unrealistic and serves as a conservative proxy for worst-case performance bounds.Source: 1) Meter data · p. 4
• The analysis includes line-congestion constraints but excludes topology-driven changes to contingency constraints to avoid complicating the data-to-price analysis.The authors note that topology changes also affect contingency analysis.Source: 1) Meter data · p. 4
• The measurement model assumes a linear function under the DC power-system model, which the authors describe as a potentially crude approximation.They use this simplifying assumption for tractability and also simulate using the actual nonlinear system model.Source: 1) Meter data · p. 4

2) Topology data:

• Topology data encode breaker states as a binary vector, so erroneous switch or breaker readings can lead to an incorrect topology estimate.The incorrect estimate is passed to subsequent operations unless the bad-data detector catches it.Source: 2) Topology data · p. 4

2) Topology data:

• Bad topology data produce a topology estimate corresponding to the corrupted breaker-state vector, which proceeds to later operations unless detected.The possible bad topology data are represented by a set B of binary vectors.Source: 2) Topology data · p. 4

B. State Estimation

• The real-time congestion pattern is inferred from estimated branch flows that meet or exceed their flow limits.The estimated congestion pattern is used in real-time LMP calculation.Source: B. State Estimation · p. 4
• Meter corruption affects the WLS state estimate linearly, but corruption confined to a k-dimensional feasible subspace limits the direction of state perturbation.The correct estimate is the estimate without bad data.Source: B. State Estimation · pp. 4–5
• Under the DC model, the control center uses weighted least squares; topology errors affect the estimate by changing the measurement matrix from H to H̄.The matrix K̄ is defined using H̄ and the measurement-noise covariance R.Source: B. State Estimation · pp. 4–5

C. Bad Data Detection

• The assumed J(x̂)-detector tests the state-estimation residue using a threshold set from a prescribed false-alarm probability.The paper focuses on cases where bad data are present but this detector fails to detect them.Source: C. Bad Data Detection · p. 5

IV. IMPACT OF BAD DATA ON LMP

• Worst-case analysis seeks the congestion pattern with the greatest LMP effect from the control center’s perspective, not the attacker’s discovery difficulty.The analysis does not assess how easily an attacker could find a locally worst pattern.Source: §IV · p. 5

A. Average Relative Price Perturbation

• ARPP averages bus-level RPP across the system to measure system-wide price perturbation.RPP is the expected percentage price perturbation caused by bad data, and varies by bus because LMP varies by location.Source: A. Average Relative Price Perturbation · p. 5
• Comparisons among ARPP and other price-impact metrics, including price-increase and price-decrease ratios, are outside the paper’s scope.The authors state that similar worst-case analyses could use those other metrics.Source: A. Average Relative Price Perturbation · p. 5

B. Worst ARPP under State Independent Bad Data Model

• The state-independent model makes bad data independent of real-time measurements and assumes a Gaussian state distribution centered on a nominal state, typically day-ahead dispatch.The covariance is denoted by Σx.Source: B. Worst ARPP under State Independent… · p. 5
• The attack moves the expected state estimate toward the center of a target price region by maximizing its shortest distance to the region boundaries.The target is the region with the largest price perturbation; the bad-data magnitude is constrained to keep the weighted squared residue at or below ε as a proxy for detection probability.Source: B. Worst ARPP under State Independent… · pp. 5–6
• Among feasible congestion patterns, the method selects the one yielding the largest ARPP, with each fixed-pattern attack formulated as a convex program.A congestion pattern is feasible when the corresponding optimization problem is feasible.Source: B. Worst ARPP under State Independent… · p. 6

C. Worst ARPP under Partially Adaptive Bad Data

• Under M2, the adversary observes only part of the measurements, estimates the state from those observations and a prior, then chooses its attack.The MMSE state estimate given the observations is the conditional mean.Source: C. Worst ARPP under Partially Adaptive… · p. 6
• For M2, the attack targets the center of a price region while controlling detection probability, and the feasible congestion pattern with the largest price perturbation is selected.The fixed-pattern problem is convex; its solution for the selected pattern is the worst bad-data vector.Source: C. Worst ARPP under Partially Adaptive… · p. 6

D. Worst ARPP under Fully Adaptive Bad Data

• Under M3, the adversary knows the full measurement set, so the attacked state and branch-flow estimates are deterministic.The worst fully adaptive bad data can be any vector satisfying the feasibility conditions for the congestion pattern with the largest price perturbation.Source: D. Worst ARPP under Fully Adaptive Bad… · p. 6

E. A Greedy Heuristic

• The greedy heuristic reduces the cost of exhaustive search by considering likely congestion patterns rather than all possible patterns.It restricts attention to lines near their flow limits and flips one candidate line at a time, stopping when no change improves the candidate.Source: E. A Greedy Heuristic · p. 6

V. BAD TOPOLOGY DATA ON LMP

• The topology-attack analysis considers an adversary changing breaker data and meter measurements to make the control center use a target topology.The target topology is formed by removing lines and is assumed to remain observable; this analysis focuses on state-preserving attacks that evade bad-data detection.Source: §V · p. 7
• Setting the meter-data attack to a = H̄x − Hx makes the received measurements consistent with the target topology, defeating any bad-data detector.The method relies on changing the breaker status to produce the target topology at the control center.Source: §V · p. 7
• In the noiseless construction, H̄x − Hx differs only in entries associated with modified lines, and the attack adjusts local flow and injection measurements around those lines.For each removed line, the procedure subtracts its directional flow measurements from the endpoint injection measurements and sets those flow measurements to zero.Source: §V · p. 7
• A topology target is considered feasible only when the adversary can perform the required breaker and meter modifications using its available attack vectors.For pricing to proceed, the system under the target topology must be observable.Source: §V · pp. 7–8

• Simulations cover IEEE-14 and IEEE-118 systems under both DC and AC state-estimation settings.Source: §VI · p. 8
• The DC estimator provides the linear setting, while the AC estimator represents the practical state estimator used in real-world power systems.Source: §VI · p. 8
• Measurements include real-power injections at every bus and real-power flows in both directions on every branch.Source: §VI · p. 8

A. Linear model with DC state estimation

• Simulations generate Gaussian system states and noisy DC measurements, then add bad data before state estimation and detection at a false-alarm probability constraint of 0.1.Source: A. Linear model with DC state estimation · p. 8
• Attack effort is measured by the number of modified meter readings; the simulations allow two lines in IEEE-14 and three in IEEE-118 to be targeted.Source: A. Linear model with DC state estimation · p. 8
• Even at low detection probabilities, average real-time price perturbations are large, especially for fully adaptive meter data and topology data.Source: A. Linear model with DC state estimation · p. 8
• Bad topology data cause much greater price perturbation than bad meter data, because they restructure price regions rather than merely move the state estimate between regions.Source: A. Linear model with DC state estimation · p. 8

B. Nonlinear model with AC state estimation

• The simulations exclude the additional effort needed to modify breaker-state data so target lines appear disconnected to the topology processor.Source: B. Nonlinear model with AC state… · p. 8
• The nonlinear AC estimator makes attack effects less significant than in the linear case, but topology data can still increase LMP impacts.Source: B. Nonlinear model with AC state… · p. 9, Figure 5 · p. 9

C. Performance of the greedy search heuristic

• Greedy search is reported to process much faster than exhaustive search without losing much accuracy on the IEEE-118 congestion-pattern search.Table I compares average search time and the percentage of runs finding the same worst congestion pattern.Source: C. Performance of the greedy search… · p. 9, Table I · p. 9

VII. CONCLUSION

• The geometric characterization of real-time LMP provides the basis for analyzing how bad data affect prices.Source: §VII · p. 9
• The adversarial framing can serve as an assurance measure when worst-case impacts are deemed acceptable, and the authors connect the analysis to smart-grid cybersecurity.Source: §VII · p. 9
• The proposed operational guidance is to flag large potential perturbations and check affected data, add protection devices, or add redundant meters.Source: §VII · p. 9
• As network size grows with measurement redundancy held constant, worst meter-data effects on LMP decrease, while topology-data effects remain nonnegligible.The authors note that the findings come from academic benchmarks involving relatively small networks.Source: §VII · p. 10