Probabilistic and Deterministic Wind Speed Forecasting Using Generative Models and Machine Learning

Abstract

This thesis investigates the critical role of wind speed forecasting in advancing wind power generation, a key component in the transition toward a carbon-neutral society by 2050. Accurate wind speed forecasting is essential for optimizing renewable energy production within the global energy mix. The study is structured to assess two distinct forecasting approaches: deterministic and probabilistic, each offering unique benefits and challenges. The project is divided into two parts. Initially, three different machine learning methods are applied to establish a baseline for deterministic wind speed forecasting. This phase aims to evaluate and compare the performance of models in predicting wind speed. Subsequently, the implementationtation of state-of-the-art deep generative model, specifically a flow-based approach designed to address uncertainties in wind speed forecasting. This model is capable of generating 100 scenarios for wind speed predictions six hours ahead, offering a robust framework for handling variability. The final objective is to provide the performance of this generative model against standard numerical weather predictions, providing a analysis of their relative accuracies and forecasting capabilities. This study aims to demonstrate the benefits of deterministic and probabilistic models in wind speed forecasting, thereby enhancing decision making in renewable energy management.