Comparative Assessment of Physical and Machine Learning Models for Wind Power Estimation: A Case Study for Hungary

Abstract

In the context of the planned mid-term development of wind power plants in Hungary, the authors evaluated the applicability of a physical-based model and several machine-learning models for wind power production estimation and wind resource availability assessments based on wind speed time series retrieved from climate reanalysis. While the physical-based model relies on a national wind power plant database and follows a bottom-up approach transforming wind speed time series into aggregate power output by using type-specific power curves, the machine learning models estimate the aggregate wind power production directly from climate data. Three types of machine learning models are trained and tested: a conventional Recurrent Neural Network (RNN) model, a Long Short- Term Memory (LSTM) model, a Support Vector Regression (SVR) model. The modelling performance is evaluated against historical aggregate wind power generation data. Machine learning models achieved similar performance metrics when compared to the physical-based model. However, different use cases can be attributed to the different types of models, considering the availability of training data sets for machine learning models. A specific use case is demonstrated for the physical-based model, where the existing set of wind turbines was extended by additional, hypothetical wind turbines. This allows for analyzing the impact of geographic distribution on expected wind resource availability for different development scenarios.