Support Vector Regression for Solar Power Prediction

Abstract In recent years, renewable energies have been covering an increasing part of the worldwide electrical power demand. The additional volatility introduced to power grids by weather dependent renewable energy sources, i.e., wind and solar, makes it necessary to improve the accuracy of energy forecasts, so that the underlying electrical grid can be operated in a cost efficient way. Governmental funding of renewable energy plants led to a vast increase of photovoltaic (PV) power plants, that now hold the second highest share of installed power capacity in Germany after onshore wind farms. The focus of this thesis lies on improving forecast accuracy current prediction models for short-term PV power predictions with horizons ranging from 15 minutes to five hours. For this, recent power measurements of PV systems throughout Germany are combined with data from parametric weather forecast approaches. Apart from commonly used numerical weather prediction (NWP) model output, satellite-based cloud motion vector predictions are applied. The data sets of these sources are preprocessed and combined with machine learning techniques. In particular, support vector regression (SVR) is optimized to work as the main regression model. SVR is an extension of the support vector machine classification method and is capable of learning regression functions in continuous space by identifying structures in the mapping of input to output data. One advantage over classical statistical models is its ability to learn non-linear dependencies between the applied data sets. In an extensive comparison, the SVR-based predictions are able to challenge a physical PV power forecasting model that is used for operational services. Instead of forecasting the PV systems output through physical irradiance to power modeling, the SVR approach does not require detailed information about a PV system. Predictions of the same quality can be achieved by SVR with access to PV measurements and weather forecasts only. To research further possible improvements in prediction quality, a study on NWP weather parameters, that are not yet included in the physical modeling system, is performed to assess their suitability as input features for PV power forecasting. The results confirm the importance of high quality irradiance forecasts. In a subsequent single PV system optimization, PV measurements and irradiance forecasts of neighboring systems are added to the input of the SVR model to make use of spatio-temporal correlations between PV systems. Furthermore, different parameter tuning techniques are evaluated with the goal to improve forecasts on a single system level, and thus develop an approach that dynamically optimizes the forecasts of all single PV systems.In recent years, renewable energies have been covering an increasing part of the worldwide electrical power demand. The additional volatility introduced to power grids by weather dependent renewable energy sources, i.e., wind and solar, makes it necessary to improve the accuracy of energy forecasts, so that the underlying electrical grid can be operated in a cost efficient way. Governmental funding of renewable energy plants led to a vast increase of photovoltaic (PV) power plants, that now hold the second highest share of installed power capacity in Germany after onshore wind farms. The focus of this thesis lies on improving forecast accuracy current prediction models for short-term PV power predictions with horizons ranging from 15 minutes to five hours. For this, recent power measurements of PV systems throughout Germany are combined with data from parametric weather forecast approaches. Apart from commonly used numerical weather prediction (NWP) model output, satellite-based cloud motion vector predictions are applied. The data sets of these sources are preprocessed and combined with machine learning techniques. In particular, support vector regression (SVR) is optimized to work as the main regression model. SVR is an extension of the support vector machine classification method and is capable of learning regression functions in continuous space by identifying structures in the mapping of input to output data. One advantage over classical statistical models is its ability to learn non-linear dependencies between the applied data sets. In an extensive comparison, the SVR-based predictions are able to challenge a physical PV power forecasting model that is used for operational services. Instead of forecasting the PV systems output through physical irradiance to power modeling, the SVR approach does not require detailed information about a PV system. Predictions of the same quality can be achieved by SVR with access to PV measurements and weather forecasts only. To research further possible improvements in prediction quality, a study on NWP weather parameters, that are not yet included in the physical modeling system, is performed to assess their suitability as input features for PV power forecasting. The results confirm the importance of high quality irradiance forecasts. In a subsequent single PV system optimization, PV measurements and irradiance forecasts of neighboring systems are added to the input of the SVR model to make use of spatio-temporal correlations between PV systems. Furthermore, different parameter tuning techniques are evaluated with the goal to improve forecasts on a single system level, and thus develop an approach that dynamically optimizes the forecasts of all single PV systems.