AI Load Forecasting for Utilities: Planning for Data Center-Driven Demand

AI load forecasting replaces the trend-plus-weather regression that utilities have used for decades with machine learning models that can handle what regression cannot: discrete gigawatt-scale data center additions of uncertain probability, electrification reshaping load curves, and behind-the-meter solar turning net load into a weather derivative — and the utilities getting it right pair short-term ML forecasts for market operations with probabilistic, scenario-based long-term forecasts for resource planning.

Resource planners are searching for better forecasting tools because the old ones are visibly breaking. NERC's latest Long-Term Reliability Assessment forecasts summer peak demand across the bulk power system growing by roughly 224 GW over ten years — a 24 percent increase over 2025 peak and a dramatic jump versus the prior year's assessment — driven largely by new data center load. EPRI projects data centers could consume 9 to 17 percent of U.S. electricity by 2030. In ERCOT, the large-load interconnection queue reached roughly 239 GW by early 2026, over three-quarters of it data centers, while observed large-load consumption was under 4 GW. Somewhere between those numbers is the load a utility must actually plan for, and finding it is now a machine learning problem as much as an engineering one.

Gain America works both sides of this equation — advising on AI data center development and power procurement for compute operators, and staffing the data science teams inside utilities that must forecast that same load. This article covers what modern AI load forecasting looks like on both horizons, how to model speculative hyperscale load honestly, and how to build the team.

Why legacy regression load forecasts break under data center growth

For fifty years, utility load forecasting rested on a stable premise: load grows slowly and predictably with population and economic activity, modulated by weather. Regression against degree days, calendar effects, and customer-count trends was accurate to within a couple of percent — good enough for rate cases, IRPs, and day-ahead scheduling. Three structural changes have invalidated that premise simultaneously:

• Lumpy, binary load. A single hyperscale campus can add 300 MW to 1 GW-plus — years of "normal" organic growth arriving as one interconnection agreement. The load either shows up or it doesn't; there is no trend line to extrapolate. Regression models built on continuous historical growth have no mechanism for representing a discrete event with a 30 percent probability of materializing in 2029.
• Electrification changing load shape, not just level. EV charging, heat pump adoption, and industrial electrification shift when energy is consumed. Winter peaks are re-emerging in summer-peaking territories; evening residential charging is steepening ramps. A model that forecasts peak magnitude but misses peak timing produces the wrong resource plan.
• DER-driven volatility. Behind-the-meter solar means the utility no longer observes true consumption — it observes net load, which now swings with cloud cover. On high-solar feeders, a passing weather front moves apparent demand faster than any historical regression coefficient anticipated. Metered load history itself becomes a corrupted training signal unless BTM generation is reconstructed and separated out.

Each of these alone strains classical methods. Together they explain why forecast misses are now measured in whole power plants, and why forecasting modernization appears in IRPs and rate filings as an explicit investment line — a shift we cover in our guide to AI consulting for energy and utilities.

Machine learning for short-term load forecasting: day-ahead and intraday

Short-term load forecasting (STLF) — next-hour through day-ahead — is where ML has the longest production track record and the most direct financial payoff, because every MWh of forecast error settles at market prices.

The production-proven stack looks like this:

• Gradient-boosted trees (LightGBM, XGBoost) as the workhorse. Trained on lagged load, multiple weather-forecast vendors' outputs, calendar and holiday features, and BTM solar estimates, boosted trees routinely cut day-ahead MAPE versus tuned regression baselines, retrain nightly on commodity hardware, and offer feature attributions that operators and regulators can interrogate.
• Sequence models (LSTM variants, temporal transformers) where granularity is high — feeder-level, nodal, or 5-minute intraday — and where probabilistic output matters. Quantile forecasts, not point forecasts, are what a scheduling desk actually needs: procurement and reserve decisions should be made against the P90, not the median.
• A statistical baseline that never leaves the ensemble. Similar-day methods and well-tuned exponential smoothing remain competitive on ordinary weather days and cost nothing. Champion-challenger backtesting per zone and horizon — not vendor claims — decides what runs.

Intraday, the highest-value ML application is now net-load forecasting: jointly forecasting gross demand and BTM solar/storage behavior. Utilities with meaningful rooftop solar penetration have learned that improving the solar irradiance forecast improves the "load" forecast more than any change to the demand model itself.

The unglamorous prerequisite is the data pipeline: AMI interval data, SCADA, weather feeds, and DER registrations unified with quality controls, plus retraining automation and drift monitoring. This is standard MLOps discipline, and it is where most utility forecasting pilots stall — the model works in a notebook and never becomes an operational system.

Long-term forecasting for IRP: scenarios, not single trajectories

Long-term forecasting for integrated resource planning is a different problem, and applying STLF techniques to it is a category error. Twenty-year horizons are dominated not by weather but by structural uncertainty: which large loads materialize, how fast electrification proceeds, what DER adoption does to net peak.

The modern approach replaces the single "expected" trajectory with probabilistic scenario ensembles:

• Bottom-up econometric and end-use modeling for organic load — still valid, now augmented with ML-driven adoption models for EVs, heat pumps, and rooftop solar that learn from granular adoption data rather than assumed S-curves.
• Explicit large-load layers, modeled separately (next section) rather than smeared into the trend.
• Monte Carlo combination across scenarios, producing a distribution of load trajectories. The resource plan is then stress-tested against the P10 and P90 paths, not optimized against a single line.

Regulators are adapting too: commissions increasingly ask utilities to justify large-load assumptions in rate cases, because the cost of overbuilding for phantom demand lands on ratepayers. A forecast that can show its probability logic survives that scrutiny; a point forecast cannot.

Modeling large-load interconnection requests: phantom load, ramp schedules, curtailability

This is the hardest and newest sub-problem, and the one utility planners ask about most. The interconnection queue is no longer a reliable demand signal. Industry analyses consistently estimate five to ten times more data center load requests than projects that will actually be built, as developers shop identical projects to multiple utilities to compare speed and price — the "phantom load" problem. PJM's 2026 long-term forecast explicitly trimmed near-term growth after applying stricter vetting to large-load adjustment requests; ERCOT's queue-versus-energized gap makes the same point at larger scale.

A defensible large-load forecasting methodology has four components:

1. Materialization scoring. Score each request on observable commitment signals: site control, permits, capital deployed, long-lead equipment orders, developer track record, and whether the same project appears in other utilities' queues. Convert scores to probabilities, and update them as milestones are hit or missed. This is a supervised learning problem once a utility has even a few years of request-outcome history.
2. Ramp modeling. A 500 MW campus does not arrive as a 500 MW step. Phased construction, staged server deployment, and utilization ramps mean load builds over 2-5 years — and load shape differs by workload: training clusters run near-flat at high utilization, while inference load tracks user demand with diurnal shape. Forecasting nameplate on the in-service date overstates near-term peak contribution badly.
3. Curtailability and flexibility. Contract structures increasingly matter more than nameplate. A load with a curtailment obligation or economic demand-response participation contributes differently to peak than firm load. Take-or-pay minimums, collateral requirements, and curtailability commitments should feed directly into both the materialization probability and the coincident-peak model. NERC's 2026 alert on large computational loads unexpectedly disconnecting during voltage disturbances is a reminder that these loads also behave differently in real time — a topic that connects forecasting to grid operations and NERC CIP-compliant AI.
4. Locational realism. Large-load siting follows power availability, fiber, water, and tax structure — the same calculus we describe from the developer's side in AI data center site selection. Utilities that understand what makes their territory genuinely attractive can separate serious requests from queue tourism far faster.

The forecasting question is no longer "how much will load grow?" It is "which of these forty requests are real, when does each one actually ramp, and how firm is it at system peak?" That is a probabilistic modeling problem, and regression has no answer to it.

Forecast inputs that matter: weather, behind-the-meter solar, EV charging, economic indicators

Model architecture gets the attention; input engineering delivers most of the accuracy. The inputs that separate modern forecasts from legacy ones:

• Ensemble weather, not single-vendor point forecasts. Ingesting multiple numerical weather prediction models and learning per-zone bias corrections typically beats any single feed. Extreme-event tails matter most — forecast error during heat domes and winter storms is where the money is.
• Reconstructed behind-the-meter solar. Combining DER interconnection registries, irradiance data, and disaggregation models to estimate BTM generation, so the model trains on true consumption rather than net load contaminated by solar variability.
• EV charging telemetry and adoption data. Registration data, managed-charging program telemetry, and public-charger utilization give both the shape (evening residential ramps, fleet depot schedules) and the trajectory of transport electrification.
• Economic and development indicators. Building permits, industrial announcements, and — increasingly — the utility's own large-load pipeline milestones, treated as leading indicators with learned lags.

ISO/RTO settlement exposure: what forecast error actually costs

Forecasting accuracy is one of the few analytics investments a utility can tie directly to market P&L. Load-serving entities buy day-ahead against their forecast and settle deviations at real-time prices. Under-forecast into a scarcity event and the balance is bought at prices that can reach administrative caps in the thousands of dollars per MWh; systematically over-forecast and you overpay day-ahead and over-procure capacity year after year. Peak forecasts also drive capacity obligations and transmission cost allocations, where errors persist for years.

A sustained one-percent MAPE improvement on a multi-GW book, settled every interval of every day, is worth millions annually — before counting the avoided capital of resource plans built on inflated peaks.

This is also the argument for utility-specific rather than generic forecasting talent: the objective function is not statistical accuracy in the abstract, it is asymmetric financial exposure across day-ahead, real-time, and capacity markets. A data scientist who has never read a settlement statement will optimize the wrong metric.

Building the utility forecasting team: power systems context via staff augmentation

The binding constraint on forecasting modernization is rarely software — it is people who combine ML engineering with power systems and market context. That profile is scarce, and utilities compete for it against tech-sector compensation. The staffing pattern that works, across the utilities we support:

• A small embedded pod to build the foundation — typically a data engineer for the AMI/SCADA/weather pipeline, an ML engineer for models and backtesting, and an MLOps engineer for retraining, drift monitoring, and serving. Contract or forward-deployed engineers with prior energy-sector work compress the first year dramatically because they arrive knowing what a settlement interval and an interconnection queue are.
• Internal ownership from day one. The utility's own forecasting analysts and resource planners work inside the pod, because the forecast must survive rate-case scrutiny and operator trust long after contractors roll off. The goal is capability transfer, not a black box.
• Convert or scale deliberately. Once the forecast is trusted in production, some utilities convert contractors to staff; others keep a smaller augmented bench for model expansion (nodal forecasts, DER programs, large-load analytics). The staff augmentation versus direct hiring trade-off favors augmentation early, when speed and scarce skills dominate, and shifts toward internal hires as the system stabilizes.

Because Gain America staffs engineers on both sides of the meter — data center development teams sizing hyperscale power requests, and utility teams forecasting them — the consultants we deploy have seen how large-load requests are actually assembled, which is precisely the context a materialization model needs. That dual vantage point is rare, and it is the difference between a forecasting team that treats the queue as data and one that understands it as negotiation.

The utilities that navigate this decade's load growth will not be the ones with the largest queues or the boldest forecasts — they will be the ones whose forecasts are probabilistic, auditable, and run by teams that understand both the models and the markets they settle in.