Two Wind Farms, Two Islands: Physics Informed Causal Wind Analysis in New Zealand

Gururaj H C, Vasudha Hegde

Wind power forecasting models often rely on correlation-based methods, which can misinterpret the relationship between meteorological variables and power generation. A key example is air density: while physics suggests denser air should increase available wind power, observational data can show a negative correlation because high-pressure regimes (high density) often coincide with lower wind speeds. This study applies a Double Machine Learning (DML) framework to separate these confounding effects using 20 years of hourly ERA5 reanalysis and New Zealand grid wind generation data (2005–2024). Because it is not feasible to study every wind farm, we use one representative wind farm per island—Te Apiti for the North Island and White Hill for the South Island—so the estimates should be interpreted as site-based evidence for island-specific operating regimes rather than island-wide fleet averages. Using XGBoost nuisance models with time-ordered cross-fitting, and controlling for wind dynamics and diurnal/seasonal structure via engineered covariates (including wind speed, lag/ramp terms, and harmonic seasonality features), we mitigate the spurious negative association seen in raw data. The results show a positive causal effect: a 0.1 kg/m³ increase in air density raises power output by about 17.2 MW at the Te Apiti wind farm (North Island) and 3.9 MW at the White Hill wind farm (South Island). The larger North Island effect is consistent with regional differences in operating conditions and fleet characteristics, but attribution to specific turbine control technology should be treated as a hypothesis. Subgroup estimates indicate the density effect is strongest in summer and at medium wind speeds (6–10 m/s), which is relevant for operational planning under marginal wind conditions.