Improved local weather forecasting using machine learning

In the context of island energy systems, the accurate forecasting of demand and supply is required for advanced energy management. Despite its primary reliance on mainland energy supplies, the German North Sea Island Borkum, integrates local renewable energy sources and storage solutions, aiming to phase out fossil fuel sources by 2030. This goal necessitates the precise modelling of energy demand and supply to optimize energy flows and ensure grid stability, especially during peak demand periods. The research focuses on improving forecasts for three key weather variables—outside air temperature, solar irradiance, and wind speed— owing to their direct impact on energy consumption and production. Statistical models were developed to improve the existing forecasts for these variables up to seven days in advance with hourly granularity. This study compiled over four years of historical data from January 2018 to June 2022, sourced from the German Meteorological Office (DWD), Météo-France, and the European Meteorological Office (ECMWF), encompassing in-situ measurements and numerical weather prediction (NWP) forecasts. The initial analysis compares external NWP forecasts with in-situ measurements, revealing that Météo-France’s AROME forecast has superior accuracy for short-term predictions. The study then proposed two methods for enhancing local weather forecasts on Borkum: machine learning models and dynamic weighting of third-party NWP forecasts. After rigorous selection and testing, Multilayer Perceptron models for air temperature and wind speed, along with a support vector machine regressor for solar irradiance, demonstrated significant improvements over NWP models in terms of mean absolute error (MAE) and mean bias error (MBE). The machine learning models notably outperformed NWP predictions, with wind speed forecasts showing a 35 to 50% improvement in MAE, air temperature forecasts achieving a 12 to 30% MAE reduction, and solar irradiance forecasts improving by 9 to 12%. In contrast, the dynamic weighting method only marginally increased short-term temperature forecast accuracy without outperforming all NWP models.