Power generator
Geospatial Wind-Siting ML System
Forward Deployed Engineer — Palantir
Site screening cut from ~1 month to ~1 hour
- PySpark
- Random Forest
- Spatial cross-validation
- NREL & meteorological data
- LLM rationale generation
Problem
Screening candidate sites for wind energy projects meant pulling together meteorological, grid, and terrain data by hand and running it through manual analysis — roughly a month of analyst time per round of site selection, with the reasoning behind each recommendation living in someone's head or a slide deck rather than the system itself.
Approach
Built PySpark pipelines over NREL, meteorological, and grid data to screen and rank candidate sites automatically, validated with spatial cross-validation to avoid the geographic leakage that makes standard cross-validation misleading for siting problems. Random Forest was chosen deliberately over a deep learning approach — the client needed to be able to interrogate why a site scored the way it did, not just trust a number, and a Random Forest's feature importances support that in a way a black-box model doesn't. An LLM-generated rationale layer then turns the model's output into a readable explanation for each recommendation. The approach is patented (US 20240394296).
Outcome
Site screening dropped from roughly a month to about an hour, and the system has been used across 5-10 site selections since — with an interpretable model the client's engineers can actually question and trust.