Solar forecasting, made simple: ideas and methods
A useful solar forecast starts with the right target, a meaningful reference and an evaluation that reflects the application. Additional model complexity must earn its place.
This guide connects my work and that of my colleagues. The linked studies provide the full author lists, evidence, limitations and code.
Start with the signal and its forecastability
Day and night form a predictable cycle; clouds are less regular. Separating these components helps ask how much information remains beyond the recurring cycle.
Our team's published contribution: the stochastic coefficient of variation, sCV, and associated forecastability indices. Their interpretation depends on their definitions and evaluation conditions.
Research story and coauthors · Published paper
Require a baseline before adding complexity
Persistence is an established idea: carry recent information forward. Define it clearly and avoid an artificially weak reference.
Our team's published contribution: BLEND uses the cyclic statistical structure of energy series to build an analytical reference forecast. This does not establish universal superiority over all models or sites.
BLEND study and authors · Methods and code · Published paper
Measure what the forecast improves
An isolated error score does not always show whether a model improves on a useful reference.
Our team's published contribution: NICEᵏ compares error components with explicit baselines. Interpret each quantity using its formula, identical observations and a common horizon. It does not measure economic value by itself.
NICE study and coauthors · Definitions and implementation
Distinguish value errors from displacement
A slightly displaced cloud structure can receive a large pointwise penalty. Whether displacement is acceptable depends on the application.
Published contribution: my forecast-validation γ-index combines spatial, temporal and intensity tolerances. It adapts tolerance-based evaluation to forecasting; it does not claim the invention of the gamma-index in medical physics.
Study and limitations · Guide and code · Published paper
Keep useful information, then size the model
Input reduction and model-capacity reduction are different levers. Count the cost of the whole pipeline, including the transformation.
Ongoing research, unsubmitted manuscripts: the dimensionality-reduction study has Hassen Bouzgou as first author; the frugal-forecasting study has Cyril Voyant as first author. Both explainers credit the complete teams. They describe methods and public software rather than presenting the manuscripts as peer-reviewed publications.
Direct PV forecasting and dimensionality reduction · Deploying frugal models
Learn through a reproducible course calculation
The UE14 course uses Ajaccio observations to explain clear sky, persistence, autoregression and ELMs. Its numbers are reproducible teaching calculations, distinct from journal results and from the fictitious examples in the explainers above.
Worked example · Course and references
A collective, verifiable approach
Persistence, linear models, ELMs and dimensionality reduction have a collective scientific history. The linked pages distinguish these foundations from each study's contributions, credit coauthors and link verified profiles.
Cyril Voyant: identity and career · All guides and software · Lire en français