HELIOS: Grid Intelligence for a High-Renewables Power System

1The balancing problem

A power grid must match supply to demand every instant, on a network with physical limits, using generators that cannot all start on command. Renewables make this harder in three distinct ways, and HELIOS is organized around answering each of them.

First, renewable output is variable and uncertain: a cloud front or a lull in the wind changes generation on minutes-to-hours timescales, so an operator must plan against a distribution of futures, not a single schedule. Second, scheduling generation and storage under real constraints is combinatorial: the unit-commitment problem is mixed-integer and NP-hard, and it grows explosively with assets, intervals, and network limits. Third, as synchronous generators retire, system inertia falls: frequency moves faster after a disturbance, shrinking the time an operator has to respond from seconds toward fractions of a second.

These are three different computational problems, forecasting under uncertainty, combinatorial optimization, and real-time control, and treating them with one tool leaves an operator either slow or blind. HELIOS pairs each with the appropriate method and connects them into a single loop that also learns from what the grid actually did.

2System overview

HELIOS runs a closed loop every few seconds to minutes: telemetry in, a dispatch and control actions out, measured response back to retrain the forecast. The loop lives inside the operator's trust boundary.

2.1Control plane and real-time plane

HELIOS separates a control plane (model deployment, configuration, market and policy data) from a real-time plane (the loop that ingests live telemetry and issues setpoints). The real-time plane is latency-bound and degrades gracefully: if the optimizer cannot return within its window, the last feasible schedule holds and control keeps acting. Nothing on the critical path depends on a quantum backend being reachable.

3Telemetry and data plane

Every decision is only as good as the picture of the grid behind it. HELIOS fuses four streams into a single, time-aligned state of the system.

3.1Sources

• Weather — numerical weather prediction (NWP) ensembles plus satellite irradiance and local met stations, the raw material of a renewable forecast.
• SCADA / telemetry — real-time measurements of generation, flows, voltages, and breaker states from across the network.
• Market — prices, bids, and schedules that define the economic objective.
• Topology — the network model (buses, lines, limits) that constrains what is physically feasible.

3.2Latency budget

Different loops run at different cadences. Real-time stability acts in the sub-second-to-seconds range; dispatch re-optimizes over minutes; forecasts refresh over minutes-to-hours. HELIOS assigns each workload a budget and keeps slow workloads off the fast path.

4Forecast

Forecast turns weather into power and predicts demand, as a distribution. A high-renewables grid plans against the tails, so a point forecast is not enough.

4.1Weather-to-power

Learned models map NWP fields (irradiance, wind speed at hub height, temperature) to plant-level output, capturing the nonlinear, asset-specific response that physical curves miss, soiling on a solar array, wake effects across a wind farm. A graph over assets with temporal attention lets a weather feature move across sites in space and time as a front or ramp does.

4.2Probabilistic output

The models are trained to emit quantiles, not a mean, using the pinball (quantile) loss. For a target quantile τ, prediction ŷ, and outcome y:

Minimizing this over a set of quantiles yields a calibrated predictive distribution (for example P10, P50, P90), which the optimizer consumes directly to plan reserves against uncertainty rather than against a single guess.

Skill is reported as MAPE and pinball loss by horizon and is monitored for drift; a forecast that decays triggers retraining through the loop of §2.

5Optimize

Optimize decides which units and batteries run when, at least cost, subject to every physical and reserve constraint. This is the combinatorial heart of grid operation, and where quantum acceleration is aimed.

5.1Unit commitment

Over a horizon of intervals t and generators g, we minimize production plus start-up cost subject to balance, ramp, reserve, and minimum up/down constraints. In compact form:

where p_g,t is power output, u_g,t ∈ {0,1} the on/off (commitment) decision, and SU_g the start-up cost. The binary commitment variables are what make this NP-hard.

5.2Quantum mapping

The commitment decisions are encoded as a QUBO over binary variables, with constraints entering as quadratic penalties, and mapped onto an Ising Hamiltonian whose ground state is the least-cost feasible commitment:

HELIOS solves it with QAOA: a depth-p circuit alternates a cost unitary with a mixing unitary, and a classical optimizer tunes the angles to minimize the measured cost.

A classical relaxation provides a warm start, and the objective minimizes CVaR of the measured energy distribution to concentrate on the best samples. Continuous dispatch p_g,t and storage are then set by a classical economic-dispatch and DC-OPF pass respecting network limits (§6).

5.3Storage co-optimization

Batteries are scheduled jointly with generation, not bolted on afterward. Their state of charge is carried as a constraint across intervals, so the optimizer charges on cheap, surplus renewable energy and discharges into the evening peak, turning a forecast surplus into avoided curtailment and avoided peaking cost.

6Stabilize

A schedule is a plan; the grid is physics. Stabilize enforces the physics in real time, keeping frequency, voltage, and line flows inside their limits as conditions move.

6.1Frequency and inertia

After a sudden power imbalance ΔP, system frequency responds according to the swing dynamics, governed by system inertia H. The initial rate of change of frequency (RoCoF) is

As synchronous machines retire, H falls and RoCoF rises, the frequency moves faster for the same disturbance. HELIOS estimates inertia in real time, schedules fast frequency response (from batteries and grid-forming inverters), and verifies that the response arrests the frequency within limits before the nadir violates protection thresholds.

6.2Power flow and contingencies

For fast screening, HELIOS uses the linearized DC power-flow model relating bus power injections P to voltage angles θ through the network susceptance matrix B, with line flows following from the angle differences:

N-1 security requires that for every credible single outage c, no remaining line exceeds its thermal limit:

HELIOS screens the full contingency list continuously and flags any outage that would drive a line into overload, so operators (or automatic schemes) can act pre-emptively. AC refinement is applied to the cases the fast DC screen marks as binding.

7Materials

A renewable grid needs cheaper, denser, longer-lived storage than exists today, and that is a chemistry problem. Materials screens candidate storage chemistries with quantum computation.

Redox potentials and formation energies depend on electron correlation, where classical density-functional theory is least reliable. Materials isolates the chemically active region of a candidate electrode or electrolyte and computes its ground-state energy with a variational quantum eigensolver, using the variational principle that the energy expectation is an upper bound on the true ground state:

From these energies, cell voltage and stability are estimated and used to rank candidates across chemistries, lithium, sodium-ion, solid-state, and flow, on energy density, cycle life, and cost. The grid does not need a battery that fits in a phone, which widens the search to chemistries mobile applications ignore.

8Hybrid orchestration

The "quantum only where it pays" principle is enforced in code by a router that chooses, per problem instance, between classical and quantum solvers and always guarantees a usable result within the operating window.

8.1The router

For each optimization job, the router considers problem size and structure, backend availability and queue depth, and the time budget from Table 1. Small commitment problems go to a tuned classical MILP solver; large ones go to the quantum/hybrid solver; when a deadline looms and no quantum time is available, the router returns the best classical result. The router is allowed to choose classical, and frequently does.

Optimize router (simplified)def route(job):
    if job.kind == "stability":            # sub-second, deterministic
        return Classical(model="dc_powerflow")  # never on the QPU path
    if job.kind == "commitment":
        n = job.num_units * job.intervals
        if n <= CLASSICAL_LIMIT:
            return Classical(solver="milp")
        if backend.available() and job.deadline > QPU_MIN:
            return Hybrid(qaoa_depth=6, warm_start=True)
        return Classical(solver="milp_heuristic")   # graceful fallback

8.2Backends and error mitigation

The same circuit description runs across simulators and gate-model QPUs through one interface, so an operator is never locked to a vendor and benefits automatically as hardware improves. On real hardware, HELIOS applies readout-error calibration and zero-noise extrapolation, and records which backend and mitigations produced each schedule in the audit trail.

9Deployment and integration

HELIOS deploys where the grid is run: inside the control room, behind the operator's security perimeter, speaking the protocols the existing systems already speak.

9.1Topology

The engine runs on-premises or in the operator's private cloud. Live telemetry and control never leave the trust boundary; only abstracted optimization instances (a cost matrix, a constraint set) are sent to a quantum backend, over an encrypted channel, and never raw grid data.

9.2Digital-twin sandbox

Before any action touches the live grid, HELIOS runs against a digital twin of the network, replaying historical events and stress scenarios so operators can validate behavior. New models are shadow-deployed (advising without acting) until they earn closed-loop authority.

9.3Security and audit

Access is role-based; every schedule, control action, and the compute path that produced it (classical or quantum, which backend) is written to an append-only audit log. Operation is auditable end to end, which regulators and system operators require.

10Validation and benchmarks

We benchmark against strong classical baselines, tuned MILP solvers and state-of-the-art forecasters, on public and operator data, and report where quantum does not yet win.

10.1Methodology

• Forecasts are scored by MAPE and pinball loss on temporally held-out data, never on random splits that leak the future.
• Commitment quality is measured by objective value at a fixed wall-clock budget and by feasibility against all constraints.
• Stability is validated on the digital twin against recorded disturbances and N-1 lists.

The verdict column is the point: quantum optimization earns its place as the commitment problem grows; on small cases and on the real-time path, classical methods win and HELIOS routes to them.

11Limitations and current maturity

This section is deliberately blunt. Grid operation is safety-critical and conservative for good reason; credibility depends on candor about what is proven and what is not.

• NISQ reality. Current quantum hardware is noisy and limited in scale. HELIOS's quantum commitment advantage is demonstrated in simulation and on small instances; classical MILP carries production commitment today, and the router falls back to it automatically.
• Crossover is workload-dependent. The size at which QAOA beats a tuned MILP solver on real hardware is still ahead for many networks. HELIOS is architected to capture that crossover without a rewrite when it arrives.
• Forecasts have irreducible uncertainty. Weather is chaotic; the value of the forecast is a calibrated distribution, not a promise. Reserves are sized against the distribution precisely because the mean will be wrong.
• Distribution loading and QPU access. Materials' quantum-chemistry advantage depends on efficient state preparation and improving hardware, and runs on simulators and pilots today.
• Simulated figures. Every number here is illustrative of the target. Production performance depends on the network, data quality, and backend, and is established per deployment on the operator's own system.

12Roadmap

—References and further reading

1. Kundur, P. Power System Stability and Control. McGraw-Hill, 1994.
2. Wood, Wollenberg, Sheblé. Power Generation, Operation, and Control. 3rd ed., Wiley, 2013.
3. Carrion & Arroyo. A computationally efficient MILP formulation for the thermal unit commitment problem. IEEE Trans. Power Systems 21 (2006).
4. Farhi, Goldstone, Gutmann. A Quantum Approximate Optimization Algorithm. arXiv:1411.4028.
5. Ajagekar & You. Quantum computing for energy systems optimization. Energy 179 (2019).
6. Barkoutsos et al. Improving Variational Quantum Optimization using CVaR. Quantum 4, 256 (2020).
7. Ulbig, Borsche, Andersson. Impact of Low Rotational Inertia on Power System Stability and Operation. IFAC 2014.
8. Koenker & Bassett. Regression Quantiles. Econometrica 46 (1978).
9. Hong et al. Probabilistic energy forecasting: Global Energy Forecasting Competition 2014 and beyond. Int. J. Forecasting 32 (2016).
10. Peruzzo et al. A variational eigenvalue solver on a photonic quantum processor. Nature Communications 5 (2014).

This brief describes a system and design intent. All performance figures are illustrative and simulated for demonstration; production characteristics are established per network and validated on the operator's own system. © Ace Hacker Research & Development Lab.