Analysis · Finance
Why a Solar Forecast Can Miss the Value of the Clock
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
The hour of the day had almost zero linear correlation with solar power in one Danish dataset. Yet it appeared in useful forecasting models. That apparent contradiction exposes a trap: a variable can help predict a daily cycle without following a simple straight-line relationship with output.
In their 2022 study, Daniel Vázquez Pombo and colleagues developed and evaluated a physics-informed forecasting method using 15 months of recordings from a solar installation in Roskilde, Denmark. They compared five machine-learning approaches for hourly power predictions extending 24, 48 and 72 hours ahead, using historical observations rather than numerical weather forecasts.
The method combined weather measurements and panel specifications to calculate quantities such as estimated cell temperature and power. These became candidate inputs alongside measured output, humidity and clock time. The researchers varied input combinations, the amount of past data supplied and model settings. Physical reasoning helped retain potentially useful inputs that a linear-correlation filter could miss.
The winning model depended on the forecast horizon. An LSTM neural network led at 24 hours, while random forest led at 48 and 72 hours. Evaluation included both a simple benchmark that carried the latest power reading forward and a statistical model incorporating the daily cycle. Night periods were excluded under stated rules to prevent easy low-output predictions from flattering accuracy.
These were forecasting comparisons at one partially shaded site, not evidence of universal model superiority or causal effects. Some power records were replaced with physically estimated values during cleaning. The authors suggested volatile weather could explain the longer-horizon rankings, but did not establish that explanation through a separate experiment.
For assessing forecast claims, the practical lesson is to inspect input selection, benchmarks, forecast horizon and scoring rules together. This study supports that evaluation discipline; it does not demonstrate better trading returns, company performance or market growth.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Increasing the Accuracy of Hourly Multi-Output Solar Power Forecast with Physics-Informed Machine Learning Daniel Vázquez Pombo, Henrik W. Bindner, Sergiu Spataru …
