The behind-the-meter data gap
In Great Britain, only a fraction of installed solar is registered with Elexon's Balancing Mechanism (BM). Most domestic and small commercial rooftop systems are embedded generation, with no real-time metered output sent to National Grid ESO. Their system contribution is estimated through demand-side models that subtract expected solar generation from metered demand.
Those demand-side models use an irradiance forecast, usually converted to generation through a fleet capacity-factor model. Forecast error therefore passes into the demand-adjusted net-load signal used for dispatch and reserve decisions. If forecast irradiance is high but cloud cover is heavier than expected, solar is overstated, demand understated, and reserve may not cover the shortfall.
This visibility gap makes irradiance accuracy a first-order balancing risk for grid operators managing the BM, not a secondary issue for embedded generation. A 5% reduction in irradiance RMSE during peak-solar midday hours can reduce unplanned reserve draws during high-output periods.
Why UK irradiance is difficult to forecast
The UK is demanding for irradiance forecasting, which helps explain differences in provider performance.
Global Horizontal Irradiance (GHI), used for flat-panel generation estimates, combines the direct solar beam, Direct Normal Irradiance (DNI), with diffuse sky radiation, Diffuse Horizontal Irradiance (DHI). Standard solar geometry predicts clear-sky GHI well. Clouds are harder: broken cover can change GHI by factors of 10 or more within minutes.
The UK's maritime climate brings variable cloud, frequent broken cumulus, and fast cloud-field evolution that challenges NWP models at any resolution. In UK summer, partial cloud can drive instantaneous GHI from 100 to 900 W/m2 over 30-minute windows. That is harder to forecast than the clear-sky or overcast split common in more continental climates.
A provider that performs well in Spain or Germany may underperform in the UK if its cloud parameterisation was tuned to less variable regimes. Different market rankings can therefore reflect real physics, not benchmark design.
How to score irradiance forecasts
For deterministic forecasts, daytime GHI RMSE and MAE are the standard starting metrics. RMSE weights large midday errors more heavily, fitting generation estimates where such errors affect reserve calculations most. MAE gives a linear view of average hourly accuracy across daylight hours.
For probabilistic irradiance products, especially those supporting intraday BM decisions, Continuous Ranked Probability Score (CRPS) is appropriate. It assesses the full distribution, including central accuracy and uncertainty calibration. If an 80% confidence interval contains actual irradiance 80% of the time, calibration is good. A narrow interval is under-dispersive and understates reserve risk.
Raw irradiance RMSE is difficult to compare across seasons and regions because clear-sky irradiance at UK latitudes changes by a factor of three or more between December and June. A skill score against a clear-sky or climatological reference compares better across periods and locations. Climingo uses this normalisation in its irradiance benchmarks.
Seasonal patterns in UK irradiance skill
UK irradiance accuracy changes markedly by season, so seasonal context is essential when reading benchmark results.
November to January brings low solar generation and limited operational impact from irradiance errors. Absolute errors are small because clear-sky irradiance is low, so winter skill scores do not reliably indicate provider quality for solar-critical decisions.
March to May and September to October are the hardest shoulder-season periods. Variable cloud, intermediate solar angles, maritime airmasses, and developing convection cause rapid irradiance changes. Provider skill separates most here: higher-resolution cloud parameterisations or satellite nowcasting outperform reliance on global NWP alone.
June to August has the highest absolute irradiance and forecast-error impact. Yet clearer, more stable airmasses can make summer easier for some providers. Thus absolute summer RMSE often exceeds spring RMSE, while clear-sky-relative skill is similar or better.
A summer-only benchmark misrepresents provider performance in the shoulder seasons, when balancing risk is highest. Run irradiance benchmarks across a full annual cycle to cover the complete difficulty range.
Benchmark findings for UK grid operators
In irradiance benchmark runs completed for early-access pilot customers in the GB grid services sector, GHI skill relative to climatology ranged from about 0.52 to 0.74 at the 6-hour horizon for a southern England zone over 12 months. That spread is material: 0.52 is a modest gain over persistence, while 0.74 captures much of the forecastable variance.
Top-scoring providers combined NWP ensemble output with geostationary satellite cloud analysis refreshed at 15 to 30-minute intervals. Lower-scoring providers mainly used single-model NWP without satellite assimilation, limiting their ability to capture rapid UK cloud evolution.
At the 24-hour horizon, the spread narrowed slightly because satellite input matters less at longer lead times, where synoptic NWP dominates. For grid procurement, provider choice matters most for intraday and same-day decisions, while day-ahead skill converges within a narrower range.
Limits of irradiance forecast accuracy
More accurate irradiance forecasts improve demand estimation and reduce reserve surprises during high-solar periods. They do not remove the visibility gap: without real-time output from behind-the-meter systems, irradiance-based generation estimates retain modelling uncertainty separate from forecast quality. Even a perfect irradiance forecast can misestimate power if the capacity-factor model is poorly calibrated to local panel technology, orientation mix, and soiling.
Closing the UK's behind-the-meter visibility gap requires better irradiance forecasts and capacity-factor models. These can be benchmarked separately, so a grid operator can assess irradiance accuracy independently of model quality. Both are needed, however, and improving only one brings diminishing returns.
Irradiance benchmarking alone will not improve grid balancing. It measures one specific part of a larger estimation problem, and one where procurement choices directly and measurably affect outcome quality.