AI Forecasts and Virtual Power Plants Team Up to Keep the Grid Stable

As renewable energy floods onto power grids around the world, one of the trickiest challenges is no longer simply generating enough electricity — it is making sure the grid can respond fast enough when demand surges or drops. A new study published in Energy Reports tackles this problem head-on, presenting a data-driven framework that pairs a deep learning forecasting engine with an optimization model to coordinate virtual power plants in distribution networks under real-world physical constraints. The result is a blueprint for how utilities might squeeze more flexibility out of the demand side while respecting the stubborn limits of the giant synchronous generators that still anchor the transmission system.

The research, led by Sheng Chen and colleagues including Tongdong Yan, Xiaolu Zhang, Youguo Zhao, and Quanxi Yu, addresses a subtle but consequential bottleneck: ramping. Conventional power plants cannot change their output instantaneously. Synchronous machines have inherent ramp-rate limits, meaning the amount of power they can add or shed per unit of time is physically bounded. When electricity flows from the transmission system into a distribution network through an upstream substation, those generator-level limits are effectively transferred onto the distribution-level dispatch problem. At the same time, the spatiotemporal variability of electrical loads introduces substantial uncertainty into how much ramping capability the system will actually need at any given moment.

Without accurate forecasts of load ramping boundaries and a well-designed allocation of regulation resources between the upstream substation and virtual power plants, the consequences can be serious: elevated frequency and voltage regulation burdens, increased network losses, and even degraded system reliability. The authors argue that coordinated dispatch of virtual power plants under ramping constraints and uncertainties has become both a pressing research problem and a practical necessity for modern grid operators.

Virtual power plants, or VPPs, are the study’s central actors. Rather than being a single physical generator, a VPP aggregates distributed resources — in this case, controllable building loads such as air-conditioning and lighting systems — into a single schedulable entity. Each VPP is modeled as an adjustable load whose operating point can be flexibly scheduled within prescribed bounds. The key insight of the framework is that VPP flexibility and substation flexibility are not physically identical. The ramping support from the upstream substation is ultimately delivered by synchronous generators that ramp slowly but can sustain regulation over long durations, while VPP flexibility from controllable loads can respond rapidly but is constrained by user comfort, allowable operating ranges, and limited energy-shifting duration. The study treats these two reserve products as additive resources at the dispatch time scale as a first-order approximation — reasonable for coordinating minute-level ramping adequacy, though not implying the two reserves are interchangeable in all physical or market contexts.

The forecasting backbone of the framework is a quantile regression temporal convolutional network, or QR-TCN. Instead of producing a single deterministic load forecast, the model directly estimates a set of conditional quantiles of future load, allowing prediction intervals to be formed from any lower-upper quantile pair. The choice of a temporal convolutional network over recurrent architectures like LSTM is deliberate: TCNs use causal dilated convolutions to capture both local load fluctuations and longer-term temporal dependencies without sequential recurrence, enabling more efficient parallel training and more stable gradient propagation. Stacking residual blocks with exponentially increasing dilation factors lets the receptive field grow rapidly with network depth, so the model can learn multi-scale temporal patterns — from short-term load wiggles to full daily periodic cycles — without an explosion in parameters.

Training relies on the quantile loss, also known as the pinball loss, an asymmetric objective with a clear statistical meaning: for a high quantile, underestimation is penalized more heavily, while for a low quantile, overestimation receives the larger penalty. This drives the network to learn the conditional distribution of future load rather than just its expected value. The team trained the model on 366 days of urban load data from Southwest China, sampled at 15-minute intervals, using 96-step historical windows to predict the next hour of load. The results were striking: an average prediction interval coverage ratio of 98.39 percent at the 99 percent confidence level, with a mean absolute error of 0.2576 and a root mean square error of 0.3433. The same architecture, applied to the ramping series — defined as the difference between consecutive load steps — achieved a 98.18 percent coverage ratio, accurately capturing the upper and lower boundaries of load fluctuation.

These probabilistic forecasts then feed directly into a coordinated dispatch model formulated as a second-order cone program, a convex optimization problem that can be solved efficiently by standard commercial solvers. The objective function jointly minimizes three terms: the operating cost of importing power through the upstream substation, the flexible regulation cost of the VPPs, and the network power losses computed from the power flow equations. The nonlinear power flow equations are handled through a second-order cone relaxation of the DistFlow formulation, preserving network security constraints such as voltage magnitude limits and branch current limits while keeping the problem convex and tractable.

A crucial modeling subtlety deserves attention: the substation itself is a passive transformer interconnection with no intrinsic ramping limitation. The ramping constraint imposed on its power import is a modeling aggregation reflecting the collective ramp-rate limits of the upstream transmission generators. By mapping these transmission-level constraints onto the substation’s power exchange profile, the dispatch model captures the upstream flexibility bottleneck without explicitly modeling the entire transmission network. Reserve coverage constraints then require the aggregate upward and downward reserve capacity from the substation and all VPPs to match the forecast ramping requirements, with the upper and lower bounds of the predicted confidence interval serving as the data-driven reserve requirements. Notably, the authors use equality constraints rather than relaxed inequalities, explicitly determining each party’s contribution to reserve provision.

The framework was validated on the IEEE 33-bus test system, with ten VPPs connected at various buses and a 24-hour horizon divided into 96 fifteen-minute intervals. Four representative cases reveal how the dispatch logic behaves. In the baseline case, the optimizer suppresses VPP operating points to maintain low-load conditions, reducing substation operating costs, and prioritizes VPP reserves for ramping requirements before drawing on the substation’s reserve. During a sudden load surge in the seventh hour, several VPPs are forced above their minimum values to accommodate the spike in upward ramping requirements, returning to lower levels once the requirement subsides. The baseline 24-hour objective came to $1,807.57 with a network loss ratio of 2.94 percent, solved in about 17 seconds.

The sensitivity cases carry the economic punchline. When the VPP regulation boundary was tightened from [0.7, 1.3] to [0.8, 1.2] times the preferred operating point, the total objective rose to $1,888.00 and the loss ratio climbed to 3.17 percent — demonstrating the tangible economic value of demand-side flexibility. Conversely, when the substation’s ramp-rate limits were tightened to ±0.3 per unit, the cost increase was modest ($1,813.44), because VPPs were pushed away from their economically optimal operating points only during the rare peak ramping periods when the substation’s reserve was fully exhausted. Raising the reserve coverage ratio to 1.4 as a safety margin raised costs to $1,812.02 while maintaining reliability. Three insights emerge: the dispatch model prioritizes VPP reserves before substation reserves; narrowing VPP flexibility increases both losses and substation costs; and even under stringent substation ramping limits, the framework maintains reliability by adaptively reallocating the burden to the demand side. As grids worldwide absorb ever more variable renewable generation, this marriage of probabilistic deep learning and convex optimization offers a compelling template for turning millions of air conditioners and lights into a coordinated shock absorber for the power system.

Subject of Research: Coordinated dispatch of virtual power plants in distribution systems under load ramping constraints and forecasting uncertainties

Article Title: Coordinated dispatch of virtual power plants in distribution systems under ramping constraints and uncertainties

Article References: Chen, S., Yan, T., Zhang, X., Zhao, Y., & Yu, Q. (2026). Coordinated dispatch of virtual power plants in distribution systems under ramping constraints and uncertainties. Energy Reports, 16, Article 109751. https://doi.org/10.1016/j.egyr.2026.109751

Image Credits: AI Generated

DOI: 10.1016/j.egyr.2026.109751

Keywords: virtual power plant, distribution system, ramping constraints, probabilistic load forecasting, quantile regression, temporal convolutional network, second-order cone programming, demand-side flexibility, power system dispatch, network losses, reserve allocation, renewable energy integration