Four-Second AGC Control Keeps Renewable Export Sections Secure in Simulations

As wind and solar farms take on a growing share of electricity supply, the fluctuations of their output can push multi-level transmission sections past their security boundaries within seconds. A study published in Frontiers in Energy Research on 27 October 2026 proposes a seconds-scale automatic generation control (AGC) security-control method that redistributes the commands of AGC-capable renewable stations every four seconds, coordinating transmission-section security, 15-minute dispatch tracking, and renewable energy accommodation in a single rolling optimization.

The work, led by Guofang Zhang of State Grid Sichuan Electric Power Company and Yi Zhou of the College of Electrical Engineering at Sichuan University, addresses a specific gap in existing practice. According to the authors, current section security control relies mainly on two measures: security-constrained economic dispatch based on day-ahead plans or ultra-short-term forecasts, and the designation of specific plants for section control in the real-time layer. Both approaches work reasonably well in traditional grids, but they face evident limitations in high-penetration renewable energy export scenarios, where rapid changes in wind and photovoltaic output or load can invalidate a power-flow verification formed tens of minutes in advance.

The problem is compounded by the structure of renewable export systems. Wind and photovoltaic stations are typically connected to local collection buses through local outgoing lines and then transmit power to the main grid through regional aggregation corridors, creating nested layers of export constraints: a local renewable export section, a regional collection section, and a total export section. Traditional AGC section control, which predefines control resources and a single target section, is difficult to extend to such multi-level nested sections, the paper argues.

The proposed method rests on three integrated components. The first is a deep neural network (DNN) for fast state estimation, trained with a composite loss that combines state supervision against AC power flow labels with a measurement-consistency loss that forces quantities recalculated from the estimated states to match actual SCADA and PMU measurements. Bus voltage magnitudes and phase angles serve as the state variables, from which line power flows and, by direction-weighted summation, the power flows of each monitored transmission section are computed. The second component is a sensitivity matrix built by aggregating direction-weighted line power transfer distribution factors (PTDFs), which characterizes how each AGC-capable renewable station affects each nested section. The third is a quadratic-programming model that allocates AGC commands under section security, station output, and ramping constraints while weighing dispatch tracking against renewable accommodation in its objective function.

Because some renewable stations lack AGC closed-loop regulation capability, their output fluctuations cannot be directly commanded. The method treats non-AGC renewable output fluctuations and load disturbances as external disturbances whose influence on section flows is captured through real-time measurements and the fast state estimator. The PTDF matrix, determined by network topology, branch reactances, and the balancing-bus configuration, is calculated once before the rolling optimization and remains fixed during the simulated 15-minute interval, since the topology does not change; it would need recalculation after line switching, network reconfiguration, or a change in the balancing bus.

A key design choice is the preventive control boundary. The real-time control boundary of each section is set at 0.95 of its transmission capacity limit, leaving a margin for residual state-estimation errors, PTDF linearization errors, renewable power fluctuations, and communication and control delays. The authors justify this coefficient with noise-robustness tests: on the IEEE 118-bus system, the overall active- and reactive-power relative estimation errors remained below 3 percent even when the measurement-noise level was raised to three times its baseline value. For systems with larger uncertainties or longer control delays, they note, a smaller security coefficient could be adopted.

The estimation tests themselves are detailed. A mixed PMU and SCADA measurement dataset of 3,000 samples was constructed on the IEEE 118-bus system, with 80 percent used for training and 20 percent for testing. Compared with traditional weighted least squares estimation, the DNN achieved a mean voltage deviation of 0.00023 per unit versus 0.00035 for WLS, and a mean phase-angle deviation of 0.0068 radians versus 0.0031. For a key test branch, the active- and reactive-power relative errors under baseline noise were 2.163 percent and 0.985 percent, rising to 2.949 percent and 2.089 percent at triple noise. Computing time also scales favorably: the DNN required 0.0029 seconds on the 118-bus system and 0.0054 seconds on a 2,383-bus system, against 0.0438 and 2.8465 seconds for WLS, though the authors caution that these figures compare core computational costs rather than complete end-to-end execution times.

The control performance was evaluated on an equivalent high-penetration renewable energy export system derived from a practical grid in Southwest China, containing eight wind power stations, five of which participate in AGC. Three strategies were compared over one 900-second dispatch interval: 15-minute plan execution, 5-minute real-time control, and the proposed 4-second rolling control. Under 15-minute plan execution, all three monitored sections exceeded the 0.95 preventive boundary, with a maximum loading rate of 0.9833, a control-boundary violation duration of 448 seconds, a maximum boundary exceedance of 52.754 megawatts, and a renewable accommodation rate of 92.690 percent. The 5-minute strategy reduced the violation duration to 300 seconds but could not eliminate violations between update instants, and its accommodation rate fell to 91.456 percent. The 4-second strategy held all sections within the preventive boundary, cutting violation duration and maximum exceedance to zero while maintaining an accommodation rate of 92.596 percent. The average optimization solution time was 0.772 milliseconds, with a maximum of 3.17 milliseconds, well within the 4-second cycle.

To test robustness, the authors constructed ten 900-second simulation scenarios varying renewable fluctuation intensity, load disturbance level, the number of AGC-participating stations, and initial section margins, plus a combined severe scenario with high fluctuations, only three AGC stations, and tight initial margins. The proposed strategy eliminated physical overload in all ten scenarios and reduced control-boundary violations to zero in eight of them. By contrast, the 15-minute and 5-minute strategies exceeded the physical section limit in the high-renewable-fluctuation, three-AGC-station, tight-margin, and combined severe scenarios, with physical-overload durations of 248, 76, 140, and 404 seconds for the 15-minute strategy and 128, 84, 84, and 280 seconds for the 5-minute strategy. Even when three-station AGC participation left residual boundary violations of 72 and 268 seconds in two scenarios, the reserved margin prevented them from developing into physical overloads; under the combined severe condition the maximum loading was limited to 0.98732 per unit.

Renewable energy utilization remained high across the scenario range, with the system-wide accommodation rate of the proposed strategy spanning 90.447 percent to 98.388 percent. The highest rates, 98.388, 95.396, and 97.188 percent, occurred under low renewable fluctuations, high AGC participation, and wide initial margins respectively, while the combined severe condition required additional curtailment to maintain section security, lowering the rate to 90.447 percent. The authors emphasize that the weight settings in the objective function, 1.0 for dispatch tracking and 1.2 for accommodation, do not relax the section security requirements, which are imposed as hard constraints, and only affect how commands are allocated within the feasible region. They present the results as evidence that the method can coordinate multi-level transmission-section security, dispatch-plan tracking, and renewable energy accommodation under stochastic disturbances, while noting that future work will address topology-adaptive PTDF updating, end-to-end hardware-in-the-loop validation, and coordinated regulation involving renewable energy, energy storage, and conventional generation.

Subject of Research: Seconds-scale AGC security control of multi-level transmission sections in high-renewable energy export power systems

Article Title: Seconds-scale renewable AGC security control considering multi-level transmission section constraints

Article References: Zhang, G., Xi, L., Zhao, J., Deng, Z., Liu, S., & Zhou, Y. (2026). Seconds-scale renewable AGC security control considering multi-level transmission section constraints. Frontiers in Energy Research, 14, Article 1905686. https://doi.org/10.3389/fenrg.2026.1905686

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1905686

Keywords: automatic generation control, transmission section security, renewable energy accommodation, deep neural network state estimation, power transfer distribution factors, quadratic programming, IEEE 118-bus system, wind power, grid congestion, dispatch tracking, smart grids, Seconds-scale