Metered solar farms: research plan

A lower number in the "Priority" column runs sooner.

Priority

Id	Question	Priority	Waits for	State
300	How good is the assembled best-guess day-ahead forecast for the six metered farms, and what does a new farm with no history get?	1	Nothing	Planned

Planned experiments

Experiment	Hypothesis	Relevance

Proposals in detail

Better irradiance for the metered farms

Experiment 218 found that the plant's shape is not what limits the fit: giving each farm its own plant parameters raised the fraction of output explained by only 0.003 over one shape shared by all farms. The next question is whether a solar-irradiance product other than the Copernicus Atmosphere Monitoring Service (CAMS), or a blend of products, raises the fraction of the six farms' output explained by the plant model. The experiment fails if the best blend raises that fraction by less than 0.02 over CAMS alone.

XGBoost with the plant's mean and the members' spread

Experiment 224 found that an XGBoost model scores lower pinball loss at the 0.01 to 0.95 quantiles and the plant at the 0.99 quantile. The proposal is to give an XGBoost model two extra input columns, the plant's mean response and the standard deviation of the 51 plant responses, and to refit it weekly. The experiment fails if the quantile score improves by less than 2% over plain XGBoost, or if the 0.99 pinball loss stays above the plant's.