Dalarna University's logo and link to the university's website

Photovoltaic (PV) systems become an important technology for generating electric power, but their output is not as stable and consistent as energy output from traditional fossil fuels. To help grow and integrate these new technologies into larger power systems and power grids, methodologies need to be improved to accurately forecast PV output.

One particularly important type of forecast is One-Day-Ahead forecasting, and it is the focus of this paper. Factors such as solar irradiation, air temperature, and relative humidity can influence the output of PV systems, and the accuracy of the forecasting relies heavily on the chosen input variables. The research associated with this paper tests the hypothesis that, given the availability of weather forecast variables, these variables can be used instead of instant meteorological variables measured onsite (local variables) in predicting One-Day-Ahead (from sunrise to sunset) solar energy.

This research was conducted to forecast the power production of an 8.4 kW solar site located at Västerås. Research and findings are reviewed, focused on two major techniques – linear, using Multiple Linear Regression (MLR), and non-linear, using Support Vector Regression (SVM). Both approaches are explored relative to developing the best One-Day-Ahead (sunrise to sunset) solar energy output forecast.

The research utilized solar energy output data obtained from Västerås solar site as well as metrological variables from Copernicus Atmosphere Monitoring Service (GMES) for the period 01/05/2014 – 10/12/2015. Stepwise techniques were used to identify the most significant input variables as well as the application of a" sliding windows" technique for dataset construction (with varying historical time parameters).

The results indicate a 30-day historical horizon provides the most accurate forecasts. Even though the research also showed that SVM achieved 16,4% more accurate predictions than MLR, Mean Absolute Percentage Error (MAPE) analysis indicates that SVM can only adequately predict (with MAPE below 30%) just over 50% of the time (188 days) over the course of a year. Thus, there are opportunities for future expansion and improvement to the research and methodologies described in this paper.