Numerical weather prediction (NWP) has matured considerably over the past decade. Global models like ECMWF's IFS and Japan Meteorological Agency's GSM now produce 10-day forecasts at horizontal resolutions below 10 km with skill scores that would have been considered exceptional fifteen years ago. For weather-dependent decisions in shipping, agriculture, or road construction, these models are genuinely useful.
For battery dispatch optimization tied to solar generation, they are not enough. And the gap is not about overall skill. It is about what these models optimize for versus what dispatch logic actually needs.
What Dispatch Needs That Global NWP Does Not Provide
A grid-scale solar forecast for dispatch purposes needs to deliver accurate irradiance estimates at 30-minute intervals, site-specific, over a 24 to 48 hour horizon. The practical requirement is not just accuracy on average: it is low error during the specific windows when the dispatch optimizer is making consequential decisions. Those windows are typically the morning ramp-up (critical for JEPX day-ahead bidding), the midday plateau (when curtailment risk is highest in surplus-prone areas), and the evening ramp-down (which determines the overnight SOC target).
Global NWP models have a few structural limitations in this context:
- Grid resolution: Even at 10 km horizontal resolution, a model cell covers an area that may include multiple elevation bands, coastal vs. inland terrain, and different cloud formation dynamics. A solar farm at the edge of a ridge line and one in a valley five kilometers away can see very different irradiance. The model cell averages across that heterogeneity.
- Cloud parameterization: Cumulus convection and thin cirrus cloud formation are parameterized rather than resolved in global models. This produces systematic irradiance errors during partly cloudy periods, which are precisely the conditions when forecast uncertainty is highest for dispatch purposes.
- Update latency: Major global models update every six to twelve hours. For intraday dispatch decisions being made at 09:00 for the 10:00 to 16:00 window, a forecast initialized at 00:00 UTC may already be six to nine hours stale before it enters the dispatch system.
The Local Correction Layer
The standard approach to bridging this gap is bias correction and statistical downscaling on top of a global NWP backbone. A local correction model is trained on historical pairs of NWP output and observed irradiance (or proxied via actual generation data) at or near the target site. The correction adjusts for systematic biases in the NWP that are specific to the site's topography and climate regime.
Japan's topography makes this non-trivial. The Japanese archipelago has steep terrain gradients close to the coast, frequent orographic cloud formation, and sharp seasonal transitions that affect the correction model's performance differently across the year. A bias correction model trained on summer data does not generalize cleanly to winter, and vice versa. Seasonal recalibration of the local correction layer is not optional if you want consistent forecast skill across the full operating year.
For sites in Kyushu and Shikoku, where wet-season cloud regimes in June and July produce sustained overcast periods, the NWP cloud representation errors tend to be largest. A site near the Ariake Sea or along Shikoku's Pacific coast can experience mesoscale marine cloud incursion patterns that global models systematically underestimate, leading to irradiance over-forecasts during these events.
Nowcasting Fills the Short-Range Gap
For the zero to four hour window, satellite-based nowcasting outperforms NWP consistently. Japan Meteorological Agency's Himawari satellite provides visible and infrared imagery at 10-minute intervals and 500m resolution over Japan. Cloud motion vector extrapolation from Himawari data gives a short-range irradiance nowcast that is often more accurate than a six-hourly NWP update for the next two to three hours.
The practical implication: an intraday dispatch system that relies only on NWP for its short-range solar forecast is leaving a more accurate information source unused. The nowcast layer does not replace NWP for the four to 48 hour range, where NWP dynamics-based prediction outperforms cloud motion extrapolation. Both are needed, blended based on the forecast horizon.
We handle this as a time-horizon-weighted ensemble in our forecasting stack: nowcasting dominates out to roughly three hours, NWP with local correction dominates from four hours onward, and the blend region uses confidence weighting based on recent nowcast and NWP skill at the specific site. This is not exotic; it is the approach used by serious solar forecasting operators globally. What varies is how well the local correction models are calibrated for Japan-specific cloud regimes.
What Poor Forecast Quality Costs in Dispatch
Forecast error translates directly to dispatch error through a chain that is easy to trace. A 15% irradiance over-forecast for the 13:00 to 15:00 window means the dispatch optimizer expects 15% more solar generation than will actually arrive. If the battery was pre-discharged in anticipation of that generation filling the JEPX position, and the solar generation does not materialize, the site runs short. The shortfall must be covered by purchasing in the JEPX intraday market or by incurring an imbalance penalty under the area TSO's settlement rules.
Imbalance penalties in Japan's market are not trivial. The area imbalance settlement price can be significantly higher than the prevailing JEPX spot price during constrained periods. A small forecast error compounded over multiple such events in a month adds up to a measurable P&L impact on the storage asset.
We are not saying that getting the forecast right eliminates all imbalance risk. Inherent stochasticity in weather means some forecast error is unavoidable. The question is whether you are operating with the best available forecast or with a generic NWP feed that was never calibrated for your site's specific irradiance environment. The difference in Mean Absolute Error between a generic NWP and a locally corrected ensemble is typically in the range of 20 to 40% reduction for partly-cloudy conditions in the Japanese context, which is the condition class where the error matters most for dispatch.
Matching Forecast Architecture to Dispatch Decisions
Each dispatch decision has a different time-horizon dependency, and the forecast architecture should be matched to it. Day-ahead JEPX bidding needs a 24-hour forecast with good skill on daily totals and peak generation windows. Intraday BESS dispatch needs a rolling 30-minute to four-hour nowcast. Curtailment pre-positioning needs a forecast that is specifically tuned for midday plateau accuracy and the onset/clearance timing of cloudy periods.
A single global NWP feed does none of these jobs optimally. That is not a criticism of NWP: it was not designed for this use case. The forecast quality gap is a product mismatch, not a model failure. Closing it requires treating solar irradiance forecasting as a purpose-built component in the dispatch stack, not as a generic weather data subscription.
