Wind Power Prediction with Machine Learning Methods in Complex Terrain Areas

• Main Author: Sæther, Brynhild Bentsen
• Format: Master Thesis
• Language: English
• Published: UiT The Arctic University of Norway 2021
• Subjects: EOM-3901 Master’s thesis in Energy; Climate and Environment; VDP::Technology: 500::Electrotechnical disciplines: 540::Electrical power engineering: 542; VDP::Teknologi: 500::Elektrotekniske fag: 540::Elkraft: 542; VDP::Social science: 200::Urbanism and physical planning: 230; VDP::Samfunnsvitenskap: 200::Urbanisme og fysisk planlegging: 230; EOM-3901; Norway; Northern Norway
• Online Access: https://hdl.handle.net/10037/21939

The increasing amount of intermittant wind energy sources connected to the power grid present several challenges in balancing the power network. Accurate prediction of wind power production is identified as one of the most important measures for balancing the power network while maintaining a sustainable integration of wind power in the power grid. However, the volatile nature of wind makes wind power forecasting a complicated task, and it is known that the performance of already established wind power prediction models decreases for wind farms in complex terrain sites. This thesis aims to forecast the future wind power output for five different wind farms in Northern Norway using methods from statistics and machine learning. The wind farm sites are generally characterized as complex terrain areas with good wind resources. Four different prediction models are developed for short to medium-term, multi- step prediction of wind power, ranging from traditional statistical models such as the arimax process to complex machine learning models. Additionally, two of the models are implemented both using the recursive and the direct multi- step forecasting technique. For each wind farm, the models are evaluated for an entire year and utilize multivariate input data with variables from a nwp model. The results of the experiments varied greatly across all locations. It was seen that the implemented models were outperformed by the persistence model for short forecasting horizons. However, when the forecasting horizon increased, several models showed a lower error than the persistence model.