Production forecasting and curtailment attribution are frequently discussed in the same breath — both involve analyzing how much a renewable asset generates, and both use similar inputs (irradiance data, SCADA actuals, weather history). But they are structurally different technical problems, with different data requirements, different model architectures, and different failure modes. Treating them as variations of the same problem causes both to be done poorly.
The distinction matters operationally. An asset manager trying to diagnose an underperformance event in a PPA contract dispute needs curtailment attribution, not a production forecast. A grid operator committing reserves for tomorrow's afternoon peak needs a production forecast, not a post-hoc attribution. Using the wrong tool for either task produces analysis that is technically coherent but operationally useless.
What production forecasting is actually doing
A production forecast answers a prospective question: given the expected atmospheric conditions over the next 15 minutes to 72 hours, what will this asset produce at each interval? The inputs are NWP model output (irradiance, wind speed, temperature), the asset's physical characteristics (capacity, orientation, power curve), and a learned correction layer trained on historical residuals between prior NWP inputs and observed SCADA output.
The goal of a production forecast is to accurately characterize the distribution of possible future production — ideally with calibrated confidence bands — so that the decision-maker (dispatcher, asset manager, trading desk) can commit resources with appropriate confidence. Forecast accuracy is measured as the deviation between the predicted distribution and the actual production outcome, evaluated over many intervals to characterize the forecast's statistical properties.
Critically, a production forecast does not distinguish between weather-driven shortfalls, mechanical availability issues, and operator-directed curtailment. It simply predicts what the asset will produce based on the expected weather and the historical pattern of production in similar conditions. If the asset was curtailed in training data, and the curtailment was correlated with high irradiance and high grid penetration conditions (which it often is), the model may partially learn the curtailment pattern as a production feature — systematically under-forecasting production in exactly the conditions when curtailment was historically imposed.
What curtailment attribution is actually doing
Curtailment attribution answers a retrospective diagnostic question: of the production shortfall observed in this period, how much was due to weather (unavoidable resource variation), how much was due to equipment availability events (forced outages, planned maintenance), and how much was due to curtailment (operator-directed output reduction, whether for grid reliability, transmission congestion, or market economics)?
The fundamental method is counterfactual reconstruction: estimate what the asset would have produced under the actual meteorological conditions if it had been fully available and operating normally. The difference between the counterfactual production estimate and the actual SCADA production — after accounting for scheduled maintenance windows — is attributable to curtailment or forced outage events.
Getting the counterfactual right requires a production model trained on periods when the asset was genuinely operating normally and was not curtailed. This is the data requirements difference: a production forecasting system is trained on all historical actuals, potentially including curtailed periods. A curtailment attribution system must identify and exclude curtailed training periods from the counterfactual model, because training on curtailed actuals teaches the model to reproduce the curtailed production rather than the uncurtailed potential.
The curtailment signal in SCADA data
Identifying curtailed intervals in SCADA data is not always straightforward. Explicit curtailment set-points are available when the curtailment is operator-directed with a documented dispatch signal — if the EMS sent a 50% output limit command to the inverters, the historian may record that command alongside the production reading. In that case, the curtailment window is clearly flagged.
But curtailment can also appear in SCADA data without an explicit set-point record. Transmission congestion curtailment may be implemented as a merchant curtailment decision by the asset manager without a formal grid operator dispatch signal. Voltage-related reactive power management can reduce real power output at inverters without a curtailment flag in the SCADA historian. Wind turbines in curtailment-avoidance noise reduction mode (common near populated areas during nighttime hours) show production below power curve expectations without a distinguishing flag.
These implicit curtailment signals require pattern-based detection rather than relying on explicit event records. The detection logic compares SCADA production against a modeled available production estimate for the same interval: if SCADA production is more than X% below the modeled available production (typically 10–15% as a threshold, conditioned on weather conditions) and no equipment availability event is recorded, the interval is flagged as a potential curtailment candidate for manual review.
Where the two problems intersect: training data contamination
The practical intersection between production forecasting and curtailment attribution is training data quality for the ML correction model. A solar asset that experiences 12% curtailment during peak irradiance periods — which is the pattern for many grid-congested solar assets during summer afternoons — will have 12% of its peak production intervals in the training set reflecting curtailed output rather than meteorologically limited output.
A correction model trained naively on these actuals learns that high irradiance does not always produce commensurately high output — because it doesn't, on the days when the grid operator curtailed the asset. This introduces a systematic downward bias in the model's production estimates for high-irradiance intervals. The bias is specifically in the regime that matters most for peak demand forecasting: the model underestimates production potential at the exact times when the grid needs the most accurate forecast.
The solution is to identify curtailed intervals before building the correction model's training set and exclude them. If curtailed intervals are concentrated in specific time windows or weather conditions, the effect on model accuracy can be substantial: up to 3–5% reduction in peak-hour forecast accuracy at heavily curtailed sites where curtailment periods were naively included in training.
We're not saying that a production forecasting system can or should substitute for dedicated curtailment attribution analysis. They require different inputs, different model training methodologies, and different validation approaches. What we're saying is that the two problems are tightly coupled at the data level, and that doing production forecasting correctly requires understanding which intervals of historical actuals represent genuine meteorological outcomes versus curtailed or unavailable operation.
When curtailment attribution feeds back into grid operations
Curtailment attribution analysis has direct operational value beyond contract dispute resolution. When a grid operator or balancing authority conducts a curtailment attribution study across their renewable portfolio, the output reveals which assets are being curtailed most frequently, under which grid conditions, and what the counterfactual clean energy generation would have been. That analysis is the diagnostic input for identifying transmission bottlenecks, revising interconnection agreements, or redesigning reserve commitment protocols to reduce unnecessary curtailment.
The connection back to production forecasting closes the loop: if the attribution analysis reveals that a specific transmission path is congested and curtailment-triggering under specific load and generation combinations, that pattern can be incorporated into the forecast system's operational context layer — informing the dispatch operator not just what the asset can produce meteorologically, but what portion of that production is likely to be physically deliverable given the expected grid conditions.
That integrated view — meteorological production potential, transmission deliverability, and reserve commitment requirement — is what the most sophisticated grid operations teams are working toward. Production forecasting and curtailment attribution are two components of that view, and conflating them makes it harder to build either component correctly.