https://www.mdu.se/

The electricity consumption is constantly increasing in Sweden combined with an increase in intermittent power production sources. In order to keep high energy security, energy sustainability, and energy affordability, it is important to provide accurate power estimations on photovoltaic systems. SMHI has developed a mesoscale model for surface solar irradiance that covers the Nordic countries named STRÅNG. This study aims to improve the solar radiation estimation accuracy of STRÅNG by using several machine learning techniques. Linear regressors, ensemble techniques, neural network, and deep neural networks were analysed to find the optimal machine learning technique for estimating photovoltaic power outputs in Sweden. The machine learning models were trained with time, solar altitude, solar azimuth, and horizontal solar radiation from STRÅNG as input and horizontal solar radiation from weather stations as output. The machine learning models are developed by using MATLAB and Python. A single trained model was analysed using three weather station’s locations in Sweden with differences in weather conditions for training between 2008-2014 and tested against six locations in Sweden during 2015. Three regional trained models were developed to further improve the machine learning models to decrease the variation in the training data. The machine learning techniques improved the estimation accuracy of the STRÅNG data significantly all over Sweden. The XGboost, Catboost, and LSTM improved the RMSE with ~3, MAE with ~2, and R2 with ~2%. These machine learning algorithms had the best performance overall. However, the regional trained models did not perform better than the single trained model. The regional models still had a distance between the weather stations that were far away from each other because of the limited weather stations that do solar radiation measurements. The machine learning techniques also require high-quality data; Otherwise, there will be an increased chance of poor performance. The STRÅNG global horizontal radiation data had a high variation in accuracy depending on the location and time. It is important to analyse the training and validation data taken from STRÅNG to improve the accuracy of the machine learning techniques. This could be achieved by developing seasonal trained models, filtering out data where solar radiation is zero, and comparing low and high accuracy data from STRÅNG.