Imagine you are the load forecaster at a mid-size investor-owned utility in the mid-Atlantic. It is a Tuesday night in February. You come in Wednesday morning, pull your ARIMA model's day-ahead forecast, and notice it looks wrong. Not wrong like a bad weather day looks wrong, but wrong in a way you have never seen before: the model is showing a 400 MW gap between its prediction and the actual load that materialized overnight. You dig into the telemetry. A new hyperscale data center, which had been under construction for 18 months, flipped its primary computing systems online at 11 PM on Tuesday. Your ARIMA model, which has been running reliably for a decade with 3 to 4 percent MAPE, had never seen a load step like that in its training history. It had no pattern to match. The load curve broke. This is the most important capability story in the AI for Energy program, and it is playing out at utilities across North America right now.

The New Load Growth Reality: 166 GW in Five Years

For most of the past two decades, utility load forecasters worked in a relatively predictable environment. Total electricity demand in the United States grew slowly, often under 1 percent per year, held down by appliance efficiency gains and industrial offshoring. A forecaster could reasonably model the next five years as a modest, weather-normalized continuation of recent history. ARIMA models, regression models, and experienced human judgment worked together acceptably. MAPE in the 3 to 5 percent range was considered competent professional practice.

That environment is gone. Grid Strategies' 2025 utility-reported forecast projects approximately 166 GW of peak demand growth over a five-year horizon for North American utilities (a figure that independent analysts caution may be overstated by up to roughly 40 percent due to cross-jurisdiction double-counting of large-load requests, so treat it as a high-case planning scenario rather than a settled fact). To put that order of magnitude in perspective: 166 GW would be roughly equivalent to adding the entire installed generation capacity of Germany to the demand side of the North American grid in five years. It represents a load growth rate unlike anything utility planners have had to model in a generation.

Approximately 90 GW of that projected growth, more than half, is attributed to data centers and large compute facilities. The balance comes from manufacturing reshoring (driven by semiconductor, battery, and defense industrial policy), widespread electrification of heating and transportation, and crypto and AI compute loads. These are not slow, predictable growth stories. They arrive in discrete, large steps.

Data center electricity consumption is projected to climb from approximately 176 terawatt-hours (TWh) in 2023 to somewhere between 325 and 580 TWh by 2028. That range itself is a planning problem: the uncertainty band on data-center load growth is wider than the entire electricity consumption of many mid-size states. A planner who builds infrastructure for the low end and gets the high end has under-built by the equivalent of adding New England's electricity demand to the system in a handful of years.

The Step-Load Problem: Why ARIMA Has Never Seen This Before

Traditional forecasting models, primarily ARIMA and linear regression with temperature and calendar variables, are built on a fundamental assumption: the future load profile will be a statistically similar extension of the historical load profile. This is a reasonable assumption when load grows slowly and smoothly, dominated by gradual population change and appliance efficiency trends. It becomes a catastrophically bad assumption when load arrives in discrete, large, overnight steps.

A hyperscale data center is not a housing development. A housing development adds, say, 50 MW to a distribution feeder over three years as homes are built and occupied incrementally. ARIMA can track that growth, because at each point in time, the current load is reasonably predictable from the recent history. A data center connects to the transmission system as a single facility. When its operations team flips the switch on a new computing cluster, the connected load can jump by hundreds of megawatts within hours, at any time of day or night, regardless of weather, regardless of season.

The ARIMA model's training data has no step-loads of that magnitude in its history. It has never seen a 400 MW overnight interconnection. Its pattern-matching engine, which is looking for combinations of temperature, day-of-week, and seasonal signals, does not have a category for "new hyperscale facility came online." The model will forecast based on the patterns it knows. The actual load will be dramatically different from those patterns. The result is a forecast error that has nothing to do with the model's temperature response or seasonal logic; it is a regime-change error, a fundamental mismatch between the world the model was trained on and the world it is now forecasting.

Your ARIMA model has never seen a 400 MW overnight step load. The question for 2026 is not whether this will happen to your utility; it is whether your forecasting system is ready when it does.

The scale of the step-load problem is not hypothetical. PJM, the largest RTO in North America, has documented multiple instances of hundred-plus-megawatt demand changes tied to large-load interconnections. Several ISOs have begun requiring large-load customers to provide operational schedules and ramp-rate information as part of their interconnection agreements, precisely because the load curve behavior of these facilities is categorically different from residential and commercial load.

Why AI Handles Step Loads Better Than Statistical Models

The AI advantage in handling structural load change is not magic; it is architecture. Machine learning models, particularly ensemble methods and neural networks used in load forecasting, can be designed to incorporate fundamentally different categories of input variables than ARIMA can handle efficiently.

The most important input variable for a data-center step-load problem is information about the facility itself: Is it connected? Is it in a ramp phase or at steady state? Does it have a predictable diurnal operational pattern (cooling loads fluctuate with compute intensity, which fluctuates with user traffic and batch processing schedules)? These are not weather variables or calendar variables. They are facility-specific operational variables that the AI model can learn from if you give it the data.

An AI load forecasting system that receives a structured operational feed from a new hyperscale data center, even just daily expected load profiles by hour, can incorporate that information directly into its forecast. The model does not need to "figure out" that there is a new 400 MW step load; it is told about the facility and its expected operating profile as an explicit input. The forecaster's job shifts from hoping the model infers the step load from residual errors to actively managing the data pipeline that feeds facility-level information into the model.

Beyond facility-specific feeds, AI models can also detect and adapt to regime changes in ways that ARIMA cannot. When a large new load pattern begins appearing consistently in the training data, gradient-boosted tree models and neural networks can update their learned relationships through periodic retraining to incorporate the new baseline. ARIMA's structural assumptions (stationarity, autoregression on a single time series) make this kind of structural adaptation much harder without manual model rebuilding.

The accuracy improvement is quantified in industry benchmarks: AI-based load forecasting achieves approximately 1 to 2 percent MAPE in day-ahead conditions compared to 3 to 5 percent for ARIMA. On a stable load base without step loads, that gap is primarily driven by the AI's ability to handle weather nonlinearity and multi-variable interactions. On a load base undergoing data-center-driven structural change, the gap becomes much larger, because ARIMA's accuracy can degrade dramatically on step-load days while a properly configured AI model with facility feeds maintains its performance.

The Net Load Complication

The step-load problem is compounded by a simultaneous development on the supply side of the net-load equation: the rapid growth of behind-the-meter solar and battery storage. Net load is gross customer demand minus behind-the-meter generation. As solar panels proliferate on commercial and industrial rooftops, the utility's metered load understates gross consumption by an amount that varies by time of day, season, cloud cover, and the geographic distribution of installed panels.

A data center that adds 400 MW of gross load might add only 300 MW of net load on a sunny afternoon if it has a large rooftop solar installation. Or it might add 450 MW of net load if it runs its battery-backed UPS systems (uninterruptible power supplies) during an on-site maintenance window that coincides with overnight battery charging. The AI model that handles net load best is one that separately models the behind-the-meter solar offset and the battery dispatch behavior, using irradiance data and facility operating schedules, and then combines them into a net-load forecast. That multi-layer approach is exactly what AI architectures can support and what ARIMA-based single-equation models cannot easily accommodate.

The IRP and Rate-Case Implications

The load curve breaking is not just a forecasting problem. It is an IRP (Integrated Resource Plan) problem and a rate-case problem. The IRP is the document in which a utility commits to a multi-year capital investment program to serve projected load: generation, transmission, storage, and demand-response resources. If the load forecast driving the IRP is wrong by 10 or 20 percent because statistical models are systematically under- or over-representing data-center growth, the IRP commits resources to the wrong scale.

Under-building because the load model missed the data-center surge results in reliability violations: not enough generation capacity to serve load on peak days, transmission constraints that cannot be resolved through normal operating procedures, and emergency import dependence. NERC's reliability assessments flag this scenario explicitly as a near-term concern for several regions including PJM, MISO, and ERCOT.

Over-building because the load model over-represented projected data-center load (perhaps a large campus that was announced but delayed or cancelled) results in stranded investment: infrastructure built and rate-based to customers for load that never materialized. Commissions will scrutinize the forecast methodology in a rate case if the overbuild becomes visible. "We used the best available forecasting tool" is not a sufficient defense if the best available tool is a legacy ARIMA model when AI-based alternatives with better step-load performance are commercially available.

This is the professional standard that the 2026 environment demands: a utility forecaster who uses a traditional statistical model without evaluating whether it adequately handles step-load regimes, and who does not document that evaluation in the IRP support record, is exposed to regulatory and professional scrutiny when forecast errors materialize. The AI forecasting case is not just an accuracy argument; it is an IRP governance argument.

What Good Step-Load Forecasting Looks Like in Practice

A load forecasting team that has adapted to the step-load environment does several things differently from a team running a legacy ARIMA model.

First, they maintain a large-load customer registry as a live, actively managed data asset. When a new commercial or industrial customer begins the interconnection application process, they enter the forecasting system's large-load tracking file. The registry includes expected connection date, expected load profile by hour (provided by the customer as part of the interconnection application), and any known operational schedule constraints. This registry is updated as the project progresses through the queue and as the customer's operational plans become clearer.

Second, they run AI and statistical models in parallel, particularly during periods when large-load interconnections are pending. The parallel run serves two purposes: it provides a real-time comparison of which model is performing better on the current load profile, and it provides a fallback if the AI model's performance degrades because its training data has not yet been updated to include a recent step load.

Third, they have defined criteria for when a manual adjustment or override is applied to the AI forecast. If the large-load registry indicates that a new 300 MW facility is expected to come online this week but has not yet done so, the forecaster may apply a probability-weighted adjustment to the AI forecast rather than trusting the model to handle an event that is not yet in its training history.

Fourth, they have a documented model-update cadence: at minimum quarterly retraining that incorporates the most recent load data, with an out-of-cycle retraining trigger if a step-load event of a defined magnitude occurs. The trigger magnitude is set low enough to catch meaningful structural changes but not so low that trivial load fluctuations generate unnecessary retraining cycles.

The FERC and NERC Regulatory Context

The step-load problem is not invisible to regulators. FERC acted in 2026 on a rulemaking specific to loads over 20 MW connecting to the transmission grid, the first national reset of interconnection policy in a generation. The rulemaking addresses how large loads are studied, what operational data they must provide to the connecting utility, and how their load variability is reflected in reliability assessments. This is directly responsive to the forecasting accuracy problem: if data centers are required to provide operational load profiles as part of their interconnection agreements, utilities gain the facility-specific data feeds that AI forecasting systems can use.

NERC's Computational Load Entity (CLE) initiative, committed in a March 2026 FERC filing for delivery by December 31, 2026, goes further. The CLE framework creates a registry category that makes large compute loads registered grid actors with explicit reliability obligations. This means compute facilities above a defined threshold will have the same data-reporting and operational transparency obligations as generators. For a load forecaster, that means the entity that is causing the step-load problem will, under the new framework, be required to provide the data that helps forecast it.

NERC issued a Level 3 Alert in May 2026 specifically addressing data-center load growth and its reliability implications. A Level 3 Alert is NERC's most serious public advisory category, indicating that NERC has identified a systemic reliability risk that requires action across the industry. For a forecasting professional, the Level 3 Alert is a direct signal that the status quo forecasting methodology is not adequate for the current environment.

Key Takeaways

• Utility-reported forecasts (Grid Strategies 2025) project approximately 166 GW of peak demand growth over five years, with roughly 90 GW attributable to data centers. Independent analysts caution this may be overstated by up to roughly 40 percent due to cross-jurisdiction double-counting. It remains a high-case planning scenario that arrives in discrete step-loads rather than smooth curves.
• ARIMA and traditional statistical forecasting models are built on stationarity assumptions that break when load arrives in large, overnight steps from new hyperscale facilities. A 400 MW overnight step load sits entirely outside the training distribution of a model calibrated on pre-2022 load history.
• AI load forecasting achieves approximately 1 to 2 percent MAPE versus 3 to 5 percent for statistical models under stable conditions. The accuracy gap grows substantially larger on step-load days, where statistical model errors can be multiples of the AI model's errors.
• The AI advantage is architectural: ML models can incorporate facility-specific operational data feeds, handle nonlinear relationships, and adapt through periodic retraining to structural regime changes that statistical models cannot accommodate without manual rebuilding.
• The IRP and rate-case implications are real: a utility that uses a statistical model without evaluating its step-load performance and documenting that evaluation in IRP support records is exposed to regulatory scrutiny when forecast errors materialize.
• FERC's 2026 large-load rulemaking and NERC's Computational Load Entity framework directly address the data availability problem by requiring large compute facilities to provide operational load profiles as registered grid actors.
• Good step-load forecasting requires a live large-load customer registry, parallel model runs during pending interconnection periods, defined manual adjustment criteria, and a documented retraining cadence with step-load event triggers.
• Data center electricity consumption is projected to nearly double or triple by 2028 (176 TWh in 2023 to 325-580 TWh), and the uncertainty range itself is a planning problem. Treat any single data-center load growth projection as a scenario input, not a point estimate.