Skip to content

Qian Zhang

Ph.D. Candidate, Electrical EngineeringHarvard John A. Paulson School of Engineering and Applied Sciences

Research

Live demosDifferentiable Reliability live demoDeliverable Flexibility live demo

Building the decision infrastructure for the AI-era grid

How can we build a unified decision infrastructure that extracts decision-relevant information from stochastic simulations and connects decisions across timescales and engineering tools?

From two grid challenges and one deployment bottleneck to three research thrusts AI data centers bring a structurally new class of electricity demand and renewables a variable supply; together they create two coupled challenges. First, planning uncertainty: uncertainty in the location, timing and size of load growth, on a scale of years to decades. Second, multi-scale variation: variation spanning sub-second dynamics to minute-to-hour ramps. Both run into a shared deployment bottleneck. Even where better methods exist, adoption in the power sector can remain slow, because decisions are distributed across legacy simulators, heterogeneous data formats, manual interfaces and institutional procedures; this blocks better methods at every timescale. Three research thrusts answer these, in the order the sections below take them. Thrust I, Differentiable Reliability: efficient accreditation and planning via gradient estimation, answering planning uncertainty over years to decades. Thrust II, Deliverable Flexibility: feasible operation via receding horizon screening, answering multi-scale variation, working over minutes to hours. Thrust III, Deployable Agentic Intelligence: plan studies, orchestrate tools and verify results, recursive self-improving under human oversight, answering the deployment bottleneck of legacy workflows and tools. 1AI data centers + renewables New load at scale, on variable supply. 2aPlanning uncertainty In the location, timing, and size of load growth. Years to decades 2bMulti-scale variation From sub-second dynamics to minute-to-hour ramps. Sub-second to hours Both challenges · every timescale 3Deployment bottleneck Even when better methods exist, adoption in the power sector can remain slow: decisions are distributed across legacy simulators, heterogeneous data formats, manual interfaces, and institutional procedures. Legacy workflows & tools From two grid challenges and one deployment bottleneck to three research thrusts AI data centers bring a structurally new class of electricity demand and renewables a variable supply; together they create two coupled challenges. First, planning uncertainty: uncertainty in the location, timing and size of load growth, on a scale of years to decades. Second, multi-scale variation: variation spanning sub-second dynamics to minute-to-hour ramps. Both run into a shared deployment bottleneck. Even where better methods exist, adoption in the power sector can remain slow, because decisions are distributed across legacy simulators, heterogeneous data formats, manual interfaces and institutional procedures; this blocks better methods at every timescale. Three research thrusts answer these, in the order the sections below take them. Thrust I, Differentiable Reliability: efficient accreditation and planning via gradient estimation, answering planning uncertainty over years to decades. Thrust II, Deliverable Flexibility: feasible operation via receding horizon screening, answering multi-scale variation, working over minutes to hours. Thrust III, Deployable Agentic Intelligence: plan studies, orchestrate tools and verify results, recursive self-improving under human oversight, answering the deployment bottleneck of legacy workflows and tools. 1AI data centers + renewables New load at scale, on variable supply. 2aPlanning uncertainty In the location, timing, and size of load growth. Years to decades 2bMulti-scale variation From sub-second dynamics to minute-to-hour ramps. Sub-second to hours 3Deployment bottleneck Even when better methods exist, adoption in the power sector can remain slow: decisions are distributed across legacy simulators, heterogeneous data formats, manual interfaces, and institutional procedures. Legacy workflows & tools Both challenges · every timescale

Three thrusts

11 papers

Thrust one. Differentiable Reliability for Long-Term Planning

Stochastic Gradient Estimation, Differentiable Optimization, Integrated Planning

Accreditation is a derivative, not a re-run.

Capacity accreditation sets how much each resource counts toward reliability, and so what gets procured and built. As the mix shifts toward renewables, storage and flexible demand, that value turns on weather, location, portfolio and dispatch rather than nameplate. U.S. markets are moving to probabilistic, marginal accreditation, but the methods are computationally heavy.

Same simulator · same random numbers Conceptual
IPA (our method) infinitesimal perturbation analysis
1 run

adequacy simulator pathwise aggregation

1 run, every gradient
Perturbation-based finite difference, per resource
2 runs × N resources

baseline run +perturbed run finite difference

the pair shares common random numbers

2N runs, N gradients

One IPA run returns every non-storage resource’s gradient; a perturbation-based method needs a baseline and a perturbed run for each resource.

Live demo. Race IPA against perturbation. One simulation against ten, run in your browser

Challenge

How to make the capacity accreditation process computationally efficient? And how to build a scalable resource planning framework ensuring the system reliability and investment incentive?

Solution

  • Accreditation as the directional derivative of a reliability metric: under expected unserved energy, ELCC and MRI reduce to the same local object.
  • An infinitesimal perturbation analysis (IPA) estimator recovers every non-storage gradient from a single Monte Carlo simulation, over 1000× faster than perturbation-based methods.
  • Optimization-based accreditation for intertemporal resources, read from the duals of a storage dispatch problem with ISO New England, extends to multi-zone production-cost models in one or two solves.

Impact

Begun during a research internship at ISO New England. Invited to an EPRI ISO/RTO working group, reaching engineers across all U.S. regions, and carried toward industry standards with consulting firms and software vendors. A related industry report has shaped the wider accreditation-reform discussion.

Papers in thrust I 2
Thrust I — Differentiable Reliability: papers
No.TitleAuthorsStatus
01 A Gradient-Based Capacity Accreditation Framework in Resource Adequacy: Formulation, Computation, and Practical Implications Q. Zhang, F. Zhao, G. Stephen, et al. To appear
02 Reformulating Energy Storage Capacity Accreditation Problem with Marginal Reliability Impact Q. Zhang, F. Zhao, T. Zheng, L. Xie Early access

Thrust two. Deliverable Flexibility for Short-Term Operation

Multi-interval Optimization, Duration-aware Algorithm, Incentive Compatibility

Flexibility is a duration, not a quantity.

Differentiable reliability settles what can credibly count toward adequacy. Deliverable flexibility asks whether that capability actually arrives when the system needs it. Solar-driven net-load ramps and coordinated data-center workloads push variation from seconds to hours, and what a fleet can deliver depends on ramp rate, headroom and location.

Same generators · same total power Illustrative example
Cheapest-first economic dispatch
G1 1 min
G2 9 min
+30MW below the requirement
Duration-preserving duration-aware dispatch
G1 3 min
G2 7 min
+40MW requirement met

Same fleet, same output: where the last dispatch left each unit decides whether the 35 MW two-minute need is met.

Live demo. Dispatch a fleet through a ramp. Ten generators, four methods, one four-hour 600 MW ramp

Challenge

When does a resource’s ramp capability become truly deliverable flexibility? How should flexibility shortages be detected and priced?

Solution

  • With MISO: an enhanced ramp product reproduces look-ahead dispatch at a single decision point, but the equivalence breaks under rolling operation.
  • A duration-aware ramp adequacy framework flags insufficient-duration bands before they bind. Its dispatch policy matches a perfect-foresight benchmark’s minimum security loss, absent transmission constraints.
  • Conventional settlement leaves the generators that absorb ramp scarcity uncompensated for their opportunity costs, so securing flexibility and pricing it have to be solved together.

Impact

Grew out of collaboration with MISO’s research team. Applied with the Brattle Group to MISO’s realized and forecast operating data, the framework identified potential ramp-adequacy shortfalls under projected 2030 conditions and now supports evaluation of mitigations — alternative product durations, procurement portfolios and dispatch policies.

Papers in thrust II 4
Thrust II — Deliverable Flexibility: papers
No.TitleAuthorsStatus
03 Comparative Assessment of Look-Ahead Economic Dispatch and Ramp Products for Grid Flexibility Q. Zhang, L. Xie, L. Zhao, C. Wang In press
04 Duration-Aware Ramp Adequacy Screening Q. Zhang, A. Looney, C. Tian, et al. Preprint
05 Lost Opportunity Costs Under Ramp Stress: A Comparison of Ramp-Product Dispatch and Look-Ahead Economic Dispatch A. Looney, Q. Zhang, L. Xie To appear
06 Locational Marginal Pricing for Adaptive Robust Look-Ahead Dispatch with Causal Affine Recourse A. Looney, Q. Zhang, L. Xie Preprint

Thrust three. Deployable Agentic Intelligence for Grid Decision-Making

Open-source Community, Evaluation Benchmark, AI for Engineering Discovery

Agents as an orchestration layer — grid-native, and physically verifiable.

Better reliability and flexibility methods still reach the grid slowly, because decisions are spread across legacy simulators, heterogeneous data formats, manual interfaces and institutional procedure. LLM-based agents can connect validated models, data and tools without replacing the physical model or human authority.

Schematic

fixes this case — fail, mitigate, simulate again improves across cases — only what passed becomes experience

Example Application in Power Systems: The agent proposes, the grid-code test checks, and only passed cases feed back.

Challenge

How can AI agents become grid-native engineering decision makers that reason with power-system structure, learn from verified experience, and remain physically verifiable?

Solution

  • Founded the PowerAgent open-source community (foundation models, model context protocol, benchmarks), and made it grid-native with grid topology, knowledge graphs and unified data schemas.
  • Physics-grounded benchmarks — PowerAgentBench-SS, PowerAgentBench-Dyn and RestoreBench — test contingency analysis, dynamic studies and power-flow restoration through independent physical validation.
  • With ISO New England, self-improving agents for transmission planning surface risks and propose mitigations, learning only from verified experience.

Impact

PowerAgent now counts more than 800 engagements from engineers and researchers worldwide and has passed 6,000 downloads; the paper was selected among the most popular in IEEE Power & Energy Magazine. The platform is being built with ISO New England’s advanced technology team and tested by transmission-planning engineers on their daily workflow. Related work has been presented to ERCOT, Eversource, CenterPoint Energy, Appian Way Energy Partners and several national laboratories.

Papers in thrust III 5
Thrust III — Deployable Agentic Intelligence: papers
No.TitleAuthorsStatus
07 PowerAgent: A Road Map Toward Agentic Intelligence in Power Systems: Foundation Model, Model Context Protocol, and Workflow Q. Zhang, L. Xie Published
08 PowerAgentBench-SS: A Benchmark for Agentic AI in Power System Steady-State Studies C. Mylonas, M. Foti, A. Pomarico, M. Duarte, Q. Zhang, et al. To appear
09 PowerAgentBench-Dyn: A Benchmark for Agentic AI in Power System Dynamic Studies Q. Zhang, A. Pomarico, C. Mylonas, et al. To appear
10 RestoreBench: Can AI Agents Restore Power Flow Convergence? R. Mansutti, A. Pomarico, R. Jakob, Q. Zhang, et al. Preprint
11 From Crypto Mining to Artificial Intelligence Training: Economic Drivers and Grid Impacts of the Data Center Transition Q. Zhang, S. Chung, L. Xie Published