Estimation of solar irradiance at the North Sea using machine learning

Abstract

Further integration of renewable energy will result in higher pressure on land availability especially in densely populated areas. Therefore, several companies are exploring the possibility of offshore PV systems. To estimate the potential of offshore PV in the North Sea, this research provides an offshore irradiance estimation. This estimation is conducted comparing the Machine Learning Methods (MLMs), Random forest (RF), Extreme Gradient Boosting (XGB) and the Artificial Neural Network (ANN). The estimation is performed using onshore training data provided by the Royal Dutch Meteorological Institute (KNMI) containing 15 different climate variables as input for the MLMs. This research presents a new approach in the solar resource estimation field by evaluating the role of external factors that influence the performance of the selected MLMs. An onshore case study is conducted to identify how the distance, cardinal direction and temporal differences between training and validation data affect the performance of the selected MLMs. This onshore case study is performed using data of 15 selected stations in the Netherlands. The ANN produced the overall best performance with an average MAE of 21.5 J/cm^2 and a relative error of 1.10. XGB (22.63 J/cm^2, 1.092) and RF (22.67, 1.149) produced slightly higher errors. Based on the onshore case study it is concluded that the external factors distance and cardinal direction strongly affect the performance of MLMs. Larger distances between training and validation stations resulted in considerable higher relative errors and models validated west of the training station showed above average relative errors for all MLMs. Temporal differences between training and validation data moderately affected the performance of the MLMs. The results of the offshore estimation produced average irradiance levels of 73.1 J/cm^2 which is slightly lower compared to onshore levels of 75.6 J/cm^2. These results are conflicting as a satellite based study by the KNMI concluded offshore irradiance levels are 4-8% higher compared to onshore levels. Since most offshore stations are located north-west of onshore training stations and distances are relatively large, the result of this estimation should be interpreted with caution. Considerable improvements on the offshore estimation can possibly be made by combining ground measurements with satellite data. Nevertheless, this research provides valuable conclusions on the performance of MLMs to estimate offshore irradiance.