Spatially transferable machine learning wind power prediction models: v−logit random forests

• Published in: Renewable energy 2024-03, Vol.223, p.120066, Article 120066
• Main Authors: Arrieta-Prieto, Mario, Schell, Kristen R.
• Format: Article
• Language: English
• Subjects: Hierarchical clustering; Random forest; Ripley’s K function; v-logit; Wind direction prediction; Wind power prediction
• ISSN: 0960-1481; 1879-0682
• DOI: 10.1016/j.renene.2024.120066

Wind power prediction models provide essential information to wind farm developers and power system operators on the power available at an undeveloped location. Traditionally, statistical models require recalibration of the model’s parameters in order to fit the model to a specific location’s dynamics. To mitigate this computational expense, this research develops a data-driven wind power prediction model that is spatially transferable, without the need for recalibration of the model’s parameters at a new location. This study also develops a wind direction prediction and interpolation model, as well as a transferability metric that demonstrates the model’s predictive accuracy at a new location. The transferability metric is evaluated in an exhaustive experimental setting. Results show that a v-logit random forest with minimum information requirements is the most transferable random forest variant for wind power prediction, with an error rate as low as 9% and never more than 20% when a transferred model is applied to a new location. The case study illustrates that the transferred model outperforms a site-specific, in-situ model for several existing wind farms. Clustering based on similarity, validated by cophenetic correlation and mapping, reveal that both the latitude of the farms, and the similarity of their turbine layouts, increase model transferability.