Linear Grid Models Lose Accuracy Under Wide Operating Conditions

Linear power-flow models become less accurate when power systems operate far from the state around which they were calibrated, according to a modeling study that examined optimal power flow, unit commitment, stressed power-flow calculations and chance-constrained optimal power flow. Simulations on the IEEE 30-bus test system found that enlarging the range of load and generation conditions broadened the distributions of errors in active and reactive power flows, bus voltages and bus angles. The authors say the results identify two linked limitations: a system state can move beyond the neighborhood where a first-order approximation remains reliable, and one variable transformation may not represent nonlinear behavior adequately throughout a broad operating range.

The work focuses on a general physics-based linear model derived from a first-order Taylor expansion of the nonlinear alternating-current power-flow equations. Rather than relying only on a local approximation, the model transforms state variables before linearization and can include a compensation term for network losses. A voltage-magnitude exponent for each bus is fitted by minimizing differences between linear and AC power-flow calculations in historical samples. This exponent is an approximation parameter, not a direct measure of a bus’s physical type, and its fitted value can reflect voltage profiles, network parameters, topology and power-injection patterns together.

Two mechanisms govern the approximation error. The first is distance from the expansion point: neglected higher-order terms become more consequential as voltages, angle differences and injections move away from the initial state. The second is the ability of the chosen independent-variable function to represent the curvature of the AC equations. These effects are coupled rather than cleanly separable, because a transformation changes both the coordinates used in the approximation and the curvature seen by the linear model. For that reason, the study evaluates their combined effect through physical-unit errors and empirical averages across generated operating cases, instead of claiming a universal error bound.

The numerical tests used a unified linear power-flow model on the IEEE 30-bus system. “Conventional” cases allowed bus-load fluctuations within plus or minus 10% of the baseline, while “wide” cases extended the range to plus or minus 50%. The researchers generated 1,000 samples for each category and solved the optimization problems with CPLEX 12.9.0 in MATLAB on a computer equipped with an AMD Ryzen 5 3500X processor and 16 GB of memory. These fluctuation limits were test settings for the selected system, not universal thresholds separating normal and wide operation.

In optimal power flow, which is characterized by bus loads and network topology, the mean branch active-power error increased from 0.246 MW under conventional conditions to 0.311 MW under wide conditions, while the maximum rose from 0.533 MW to 1.017 MW. Mean reactive-power error increased from 1.521 to 1.951 MVar, and the maximum from 4.614 to 7.642 MVar. The authors note that an approximate solution can incur different operating costs or fail AC feasibility checks because the linear result may contain voltage or branch-flow violations.

Unit commitment showed a similar pattern because both bus loads and generator outputs varied. Under conventional conditions, the standard deviation of branch active-power error was 0.047 MW, compared with 0.075 MW under wide conditions. Reactive-power-error standard deviation rose from 0.065 to 0.11 MVar, voltage-magnitude-error standard deviation from 7.37×10⁻⁴ to 1.36×10⁻³ per unit, and angle-error standard deviation from 0.021 to 0.062 degrees. The mean errors across the tested state variables were approximately 50% lower in the conventional category, indicating that the fixed local approximation was better suited to the smaller neighborhood.

Errors also increased in power-flow cases selected near the numerically identified non-convergence boundary of the IEEE 30-bus system. These “extreme” cases were system-specific stress samples rather than a general load-percentage category. Mean active- and reactive-power errors rose from 0.024 MW and 0.031 MVar under conventional conditions to 0.038 MW and 0.047 MVar in the extreme set. Mean voltage-magnitude errors increased from 4.51×10⁻³ to 6.28×10⁻³ per unit, while angle errors rose from 0.17 to 0.26 degrees. The paper also warns that a constant, full-rank linear Jacobian can mask the ill-conditioning that appears in the AC Jacobian near a convergence boundary.

Chance-constrained optimal power flow tests examined three illustrative renewable-energy penetration levels: 10%, 30% and 50%. Each scenario used a predefined 5.0% constraint-violation threshold and 10,000 out-of-sample Monte Carlo scenarios for evaluation. Operating-cost errors rose from 0.257% at 10% penetration to 0.293% at 30% and 0.342% at 50%. Measured constraint-violation probabilities increased from 5.02% to 5.17% and 5.26%, respectively. At 50% penetration, mean active-power-flow error was 2.51 times its value in the 10% case, while the corresponding reactive-power error was 1.57 times as large. The authors interpret these probabilities as empirical outcomes combining operating-point uncertainty and model mismatch, not analytical security bounds.

Based on the error patterns, the study proposes a framework for adaptively selecting segmented linear power-flow models. The wide operating domain would first be divided into sub-regions with similar operating features and error behavior. For each sub-region, the authors’ existing error-minimization procedure would fit a separate set of model parameters, and an offline-trained classifier would later identify the relevant sub-region from application-specific features. During dispatch, the selected coefficient set would replace the single unified model. The framework remains general: the distance measure, number of sub-regions, classifier and operating-condition features depend on the application, and the paper does not report a quantitatively validated implementation or comparisons with other linearization approaches on a common dataset.

Subject of Research: Accuracy limits and adaptive improvement of linear power-flow models under wide power-system operating conditions

Article Title: Error analysis and improvement framework for linear power flow models under wide operating conditions

Article References: Wu, Y., Long, J., Guo, Z., Lin, W., & Yang, Z. (2026). Error analysis and improvement framework for linear power flow models under wide operating conditions. Frontiers in Energy Research, 14, Article 1954278. https://doi.org/10.3389/fenrg.2026.1954278

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1954278

Keywords: linear power flow, power-system dispatch, Taylor approximation, optimal power flow, unit commitment, chance-constrained optimization, renewable variability, error analysis, adaptive modeling, IEEE 30-bus system, Error, analysis