Colloquia

Summary

The stochastic nature of wind imposes one of the greatest challenges for the competitiveness of wind systems in energy markets: predicting wind power output from wind farms ahead of time. As inherent biases in meteorological models reduce their reliability in accurate simulating essential variables for electricity modeling, enhancing our understanding on meteorological factors influencing renewable energy, e.g. wind speed, is crucial in the energy transition. In this thesis, long-term wind speed forecasting for long-term wind power estimation was performed based on the numerical meteorological ERA5 model’s wind speed data and experimental data from multiple anemometric towers using five machine learning methods. Decision Tree Regressor is the method that showed the best performance by accurately modeling wind power production, preserving bimodality and giving the smallest difference in respect to observational cumulative distribution functions. Best predictions ocurred on sites with highly contrasting spatial features. Areas influenced by large-scale effects, such as the interaction between large-scale atmospheric circulation and orography were better represented by all methods. Demonstrating ML potential to reproduce local wind conditions through global reanalysis. Training ML methods on bias corrected data can be either beneficial or redundant. While it may improve capacity factor estimates by supporting the integration of wake models, DTR method can still yield reliable, albeit slightly lower, estimates when trained solely on observational data.