Loss-of-load probability is one of the foundational reliability metrics in power system planning. Traditional LOLP calculations rely on historical reliability data with long averaging windows, generating a number that is statistically defensible but slow to reflect current conditions. Machine learning forecasting is now being applied to give LOLP calculations a shorter feedback loop and more granular inputs at the distribution level.
How traditional LOLP is calculated
LOLP at the transmission and bulk power level is well-established methodology: you model the probability distribution of generator forced outages, the probability distribution of load over a planning horizon, and compute the frequency with which load exceeds available generation capacity. The inputs are well-defined, the math is tractable, and the resulting number is accepted by regulators as a planning standard.
At the distribution level, LOLP calculations are less standardized. Distribution LOLP typically looks at the probability that peak feeder load will exceed feeder capacity ratings, accounting for N-1 contingency requirements. The load input is a peak load forecast with an uncertainty band. The wider the uncertainty band, the higher the implied LOLP, and the more reserve capacity you need to maintain to keep LOLP below a target threshold.
Where ML forecasting changes the calculation
The key input to distribution LOLP that ML forecasting improves is the load probability distribution. Instead of a single-point peak forecast with an assumed uncertainty band based on historical forecast errors, an ML ensemble approach can provide a probabilistic forecast: here is the 10th, 50th, and 90th percentile prediction for peak feeder load on this day, given current weather forecast and historical load patterns.
A tighter probability distribution at the 90th percentile means your reserve requirement to achieve a given LOLP target is smaller. A model with 6 percent MAPE has a tighter 90th percentile range than a model with 14 percent MAPE. Connecting that improvement to the LOLP calculation converts forecast accuracy improvement into a quantifiable reliability margin.
The shorter feedback loop argument
Traditional LOLP inputs use long historical averaging windows, often 10 to 20 years of reliability data, which is appropriate for generation planning but problematic for distribution planning in markets where load shapes are changing rapidly. A distribution utility in a high-EV-penetration corridor has different feeder peak probability distributions today than it did five years ago. A planning tool using 10-year historical averages will systematically underestimate current risk on affected feeders.
ML forecasting models that retrain quarterly on recent interval data reflect the current load shape, including EV charging patterns, recent solar additions, and recent large customer changes. Using the current model's probabilistic outputs as LOLP inputs produces a distribution LOLP estimate that is more responsive to structural load change than one based on historical averages.
Practical implementation
Implementing ML-based probabilistic forecasts as distribution LOLP inputs is not a single-step process. It requires a forecasting pipeline that generates ensemble or quantile outputs rather than single-point estimates, a method for translating feeder-level probabilistic forecasts into system-level LOLP calculations, and a regulatory framework that accepts the methodology. The last item is often the longest-lead constraint. Discussions with state commissions about alternative LOLP input methodologies take time.
Quantile forecasting as a LOLP input
To use ML forecasting as a direct input to distribution LOLP calculations, the forecasting pipeline needs to produce probabilistic outputs rather than single-point estimates. The standard approach is quantile regression, where the model generates separate estimates for the 10th, 50th, and 90th percentile of the load distribution at each forecast horizon. These quantiles can then be used as inputs to a LOLP calculation in place of a single-point peak forecast with an assumed uncertainty band.
The operational advantage is that the quantile estimates are derived from the model's actual experience with the specific feeder's load distribution, rather than from a generic uncertainty assumption applied to all feeders equally. A feeder with weather-stable load shows tight quantile intervals. A feeder with high EV penetration shows wider quantile intervals during peak charging hours. Using feeder-specific quantile estimates produces more accurate LOLP calculations than applying a uniform uncertainty assumption across all feeders.
Distribution LOLP in rate cases and resource planning
Distribution LOLP metrics are increasingly appearing in state commission proceedings, particularly in markets where renewable integration and EV adoption are creating new distribution-level reliability questions. Utilities that can demonstrate rigorous distribution LOLP tracking are better positioned to support capital investment justifications for feeder upgrades and new substation capacity.
The argument is that better forecasting inputs produce more defensible LOLP estimates, which in turn support more credible capacity planning justifications. A utility that can show its distribution LOLP methodology uses current-period probabilistic feeder forecasts rather than multi-year historical averages is making a stronger evidentiary case for its capital plans than one relying on traditional inputs. This matters particularly for distribution capital projects that are evaluated against a commission-defined reliability standard.
Near-term versus planning-horizon applications
There is an important distinction between LOLP as a near-term operational tool and LOLP as a planning-horizon metric. The near-term application is using short-horizon probabilistic forecasts (4 to 24 hours) to calculate real-time LOLP for specific feeders, flagging feeders where the current probabilistic forecast produces a LOLP above threshold for the next peak period. This is an operational dispatch support application.
The planning-horizon application is using seasonal probabilistic forecasts (monthly to annual) to estimate peak period LOLP for resource and capital planning purposes. The two applications use different forecast horizons and different accuracy requirements. Near-term LOLP uses 4-hour ahead quantile forecasts where accuracy is high. Planning-horizon LOLP uses seasonal probabilistic load forecasts where uncertainty is wider and the methodology for estimating uncertainty must account for structural load growth, not just weather variation.
Ampgrove's current product focuses on the near-term operational application. We generate 4-hour ahead quantile forecasts that support real-time distribution LOLP calculations for operational dispatch support. The planning-horizon application is a longer-term development area. Utilities that are primarily interested in improving their capital planning LOLP methodology will benefit more from the operational forecasting accuracy improvements in the near term, and from the expanding model dataset as a foundation for longer-horizon probabilistic forecasting once the production deployment is established.