Presentation Details
| From Climate to Cashflow: Machine Learning Forecasts of Solar Yield for Sub-Seasonal Energy Planning Marc Perez1, Richard Perez1, 2, Jing Huang1. 1Clean Power Research, Bellevue, WA, USA.2SUNY Albany, Albany, NY, USA |
Abstract
Sub-seasonal (2–8 week) forecasting of solar photovoltaic (PV) yield is critical for energy finance yet remains challenging. We present a machine-learning framework that forecasts month-ahead anomalies in the clearness index (KT*), integrating lagged El Niño–Southern Oscillation (ENSO) signals, meteorological anomalies, and geospatial covariates. Random Forest and gradient-boosted tree (XGBoost) models, trained on 45 years of ERA5 reanalysis data [1],[2] and the NOAA Multivariate ENSO Index (MEI) [3] over California, significantly outperform a smart persistence baseline. We translate KT* anomaly forecasts into PV yield predictions and demonstrate their financial value using a case study of an 850-MW utility-scale plant. Climate-informed forecasts reduce month-ahead irradiance error by nearly an order of magnitude, lower mean effective hedging costs by ~6%, and reduce delivered-price volatility by more than 90%. These results highlight the value of explicitly incorporating large-scale climate teleconnections into scalable ML-based solar forecasting workflows.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.