- Calibsun | Expertise
Probabilistic Forecast Managing solar asset uncertainty
Delivering forecast scenarios of solar power production
Probabilistic solar power forecasting estimates the range of possible solar energy production outcomes rather than a single expected value. A deterministic forecasting model delivers one number per timestep (Q50). A probabilistic forecast provides a full distribution typically expressed through quantiles from Q05 to Q95. It quantifies the forecast uncertainty attached to each prediction of photovoltaic power output. This distinction is becoming structural for IPPs, energy traders, aggregators and grid operators, who must arbitrate between solar power generation scenarios, manage imbalance penalties on the spot and intraday electricity markets, and optimise dispatch on solar assets whose power output remains inherently variable under changing weather conditions.
The level of uncertainty depends on the forecast horizon, on weather variability, particularly cloud cover dynamics, and on input data quality, from on-site measurements of solar irradiance (GHI) to satellite imagery and numerical weather prediction (NWP). CalibSun combines these sources within an analog ensemble framework, continuously recalibrated through on-site measurements, to natively deliver probabilistic solar power forecasts that reflect the true distribution of outcomes rather than relying on deterministic averages that hide the variability structure.
Key takeaways
- CalibSun is the only forecaster to natively deliver both probabilistic and deterministic forecasts, built on the analog ensemble method and combining on-site.
- Probabilistic solar power forecasting of CalibSun expresses PV power production as a distribution of outcomes from quantiles ranging from Q5 to Q95, range of Q5.
- The methodology is backed by a long scientific track record and collaboration with Mines Paris – PSL and grounded in peer-reviewed publications (IEEE, EUPVSEC…).
- Operational benefits concentrate on energy trading on the spot market, imbalance management, and storage management for PV plants.
- Probabilistic Forecast
Deterministic vs probabilistic forecast: two approaches, two operational uses
CalibSun’s probabilistic forecasting method relies on the analog ensemble, a non-parametric statistical approach.
Mechanisms and key concepts
A deterministic forecast produces a single expected value of solar power generation per timestep. A probabilistic forecast describes the full conditional distribution of possible outcomes, usually summarised through quantiles. Q10 represents the value below which production has a 10% probability of falling; Q50 is the median; Q90 the high-production threshold. The Q10–Q90 envelope is the 80% prediction interval that materialises the forecast uncertainty.
Input Data
Building this distribution requires three families of input data:
- On-site measurements provide ground truth on solar irradiance (GHI, GTI) and local weather conditions (temperature, humidity…), capturing short-term cloud cover transitions invisible to satellite.
- Satellite imagery extends spatial awareness to the surrounding cloud field.
- NWP supplies the synoptic weather variables driving day-ahead production. The relative weight of each source evolves with the forecast horizon: on-site dominates short-term, NWP dominates day-ahead, satellite bridges intraday.
Research for past similar situations
The model searches historical data for past situations meteorologically analogous to the current forecast conditions, then reconstructs the distribution of observed outcomes from these analogues. For each analogue retrieved, the actually observed photovoltaic power output is recorded. The ensemble of these observed values forms the empirical distribution of the probabilistic forecast.
This non-parametric approach has two structural advantages:
- it makes no assumption on the shape of the distribution (Gaussian, log-normal),
- it naturally captures non-linear and asymmetric behaviours of solar power generation — for instance, the truncation of production at clear-sky ceiling.
Generation of quantiles Q5 to Q95 and calibration
From the empirical distribution produced by the analog ensemble, quantiles Q5, Q10, …, Q95 are extracted. Calibration ensures statistical reliability: a Q10 forecast must, over a long observation period, be exceeded by realised production approximately 90% of the time. CalibSun validates calibration through reliability diagrams and continuous post-processing on historical data, with corrections applied per site and per forecast horizon.
What are the limits of each approach?
Deterministic
A deterministic forecast model hides the underlying variability structure of the weather conditions. Two predictions pointing to the same expected solar energy production may carry radically different risk profiles in terms of sources of uncertainty. Under stable weather conditions, the range of possible outcomes (expressed through prediction intervals from Q5 to Q95) remains narrow. While under high variability cloud cover, this ensemble spread widens substantially. A single deterministic value cannot convey this difference and communicate the full probability estimates needed by solar players.
Probabilistic
Probabilistic forecasting methods relying only on satellite imagery lack the local short-term resolution required to capture site-specific cloud dynamics. Approaches based only on NWP carry coarse spatial resolution and limited responsiveness on intraday horizons. Conversely, approaches limited to on-site measurements lose accuracy as the forecast horizon extends. Robust probabilistic solar power forecasts require the combination of all three sources, weighted according to the forecast horizon.
Comparative table: deterministic vs probabilistic forecast
Deterministic forecast
Probabilistic forecast
Output
Single expected value
Full distribution Q5–Q95
Uncertainty quantification
None
Explicit, per timestep
Trading on the spot market
Limited arbitrage
Risk-adjusted bidding (Q30, Q70, adaptive methods)
Imbalance management
Reactive
Anticipated, quantified
Storage / hybridisation control
Not optimal
Risk-based optimisation
Data sources
Often satellite-only
On-site data + satellite + NWP
Building the probabilistic model: CalibSun's scientific approach
Our market differentiation
CalibSun is currently the only forecaster to natively deliver both probabilistic and deterministic forecasts from a single integrated model.
- Expected value serves as the deterministic reference for standard operational monitoring;
- Q5 to Q95 provide the uncertainty envelope for risk-based decision-making.
This dual output is enabled by the analog ensemble approach, which produces the distribution directly — the deterministic value is a by-product, not a separate model.
A second differentiator is continuous recalibration through on-site measurements. Pyranometers and sky imagers installed on the PV plant feed the model in real time, adjusting the distribution to the actual local solar resource behaviour. This recalibration is decisive on multi-GW portfolios, where small biases compound rapidly into significant imbalance exposure.
Related Scientific publications
The methodology draws on peer-reviewed publications covering probabilistic solar forecasting using quantile regression, weather scenario generation, analog ensemble approaches and reliability assessment of probabilistic forecasts published in IEEE journals or presented at scientific conferences such as EUPVSEC and other technical reviews on energy forecasting and solar photovoltaic systems
Internal CalibSun research
CalibSun’s probabilistic forecasting method is grounded in long-running research on solar resource assessment lead by PhD data scientist Thomas Carrière on analog ensemble techniques applied to photovoltaic power.
The model was developed in scientific collaboration with Persee (Mines Paris – PSL), a reference laboratory in solar energy and atmospheric data processing, ensuring methodological rigour on the analog selection function, the calibration framework and the validation protocols.
Backed by Research & Science
Probabilistic forecast: Operational use cases
Trading, balancing and imbalance penalties
On the spot electricity markets, probabilistic forecasts allow energy traders to position bids on the quantile minimising expected imbalance cost. When penalties are asymmetric, heavier for over-production than under-production, or vice versa, the optimal bid is rarely Q50 median: it is the quantile aligning with the penalty structure (Q30, Q40, Q70 depending on the market conditions and conditional probability assessments). This optimization of bid strategy is impossible with a deterministic forecasting model.
Storage, hybridisation and dimensioning
For PV plants coupled with battery storage or hybridised with wind, probabilistic forecasts feed energy management systems with the input required for risk-based dispatch. Probabilistic forecasting methods are relevant at the operational stage to optimise BESS charge/discharge cycle decisions and therefore to maximise solar energy storage system lifetime.
- Questions
Frequently Asked Questions about probabilistic forecast
What are the benefits of probabilistic forecasts?
Probabilistic forecasts provide explicit quantification of uncertainty, enabling risk-adjusted decisions in trading on the spot market, energy management, imbalance reduction, and storage dispatch. They improve operational performance compared to deterministic outputs by exposing the range of plausible power output scenarios.
Does CalibSun produce both deterministic and probabilistic forecasts?
Yes. CalibSun’s forecasts services NEXT and INSTANT provide deterministic (expected value) and probabilistic distribution natively from Q5 to Q95 uncertainty envelope. This dual output supports both standard operational monitoring and advanced risk-based decision-making.
How is the reliability of probabilistic forecasts validated?
Reliability is validated through reliability diagrams and statistical analysis on historical data: a Q10 forecast should be exceeded by realised production approximately 90% of the time over a long observation period. CalibSun applies continuous recalibration per site and per forecast horizon to maintain this statistical consistency.
What are the other applications of these probability forecasts beyond trading?
Anticipating electricity generation is mandatory for optimising renewable energy management. By providing a range of power output rather than a single value, Calibsun helps make solar energy steerable and to optimize its integration within the energy grid, with or without storage.
