<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
Due to the inherent uncertainty in photovoltaic (PV) energy generation, an accurate power forecasting is essential to ensure a reliable operation of PV systems and a safe electric grid. Machine learning (ML) techniques have gained popularity on the development of this task due to its increased accuracy. Most literature, however, focuses only on less than 5 PV systems during training process, which does not ensure generalization to unseen systems. When in presence of a large feet, regional forecasts are the norm. Nevertheless, none of these approaches are usable when it comes to monitoring residential PV systems. In this work, we propose a single ML model that is able to predict the individual power of a large fleet of 1102 PV systems. XGBoost algorithm was selected as the most suitable algorithm for the task of PV yield nowcasting due to its performance and ease of use. This algorithm obtains Mean Absolute Error (MAE) of 0.877 kWh (considering an average system size of 4.44 kWp) and Mean Absolute Percentage Error (MAPE) of 23% for hourly data aggregated to daily values. XGBoost predictions for individual PV systems are on average two times better than currently used commercial software. We discuss the lack of a suitable loss function that can combine absolute and relative errors for residential PV yield forecasting. We also point out the lack of an adequate metric to compute the error made on the predictions and provide hints on developing a suitable one.
...
Due to the inherent uncertainty in photovoltaic (PV) energy generation, an accurate power forecasting is essential to ensure a reliable operation of PV systems and a safe electric grid. Machine learning (ML) techniques have gained popularity on the development of this task due to its increased accuracy. Most literature, however, focuses only on less than 5 PV systems during training process, which does not ensure generalization to unseen systems. When in presence of a large feet, regional forecasts are the norm. Nevertheless, none of these approaches are usable when it comes to monitoring residential PV systems. In this work, we propose a single ML model that is able to predict the individual power of a large fleet of 1102 PV systems. XGBoost algorithm was selected as the most suitable algorithm for the task of PV yield nowcasting due to its performance and ease of use. This algorithm obtains Mean Absolute Error (MAE) of 0.877 kWh (considering an average system size of 4.44 kWp) and Mean Absolute Percentage Error (MAPE) of 23% for hourly data aggregated to daily values. XGBoost predictions for individual PV systems are on average two times better than currently used commercial software. We discuss the lack of a suitable loss function that can combine absolute and relative errors for residential PV yield forecasting. We also point out the lack of an adequate metric to compute the error made on the predictions and provide hints on developing a suitable one.
A quick-scan yield prediction method has been developed to assess rooftop photovoltaic (PV) potential. The method has three main parts. For each roof, first (i) virtual 3D roof segments were reconstructed using aerial imagery, then, (ii) PV modules were automatically fitted onto roof segments using a fitting algorithm and finally, (iii) expected annual yield was calculated. For each roof, the annual yield was calculated by three different quick yield calculation approaches. Two approaches are commercial software packages of Solar Monkey (SM) and Photovoltaic Geographical Information System (PVGIS) whereas the other one is the simplified skyline-based approach developed in photovoltaic material and devices (PVMD) group of Delft University of Technology. To validate the quick-scan method, a set of 145 roofs and 215 roof segments were chosen in urban areas in the Netherlands. For the chosen roofs, the number of fitted modules and calculated yield were compared with the actual modular layout and the measured yield of existing PV systems. Results showed a satisfactory agreement between the quick-scan yield prediction and measured annual yield per roof, with relative standard deviations of 7.2%, 9.1%, and 7.5% respectively for SM, PVGIS, and PVMD approaches. It was concluded that the obstacle-including approaches (e.g. SM and PVMD) outperformed the approaches which neglect the shading by surrounding obstacles (e.g. PVGIS). Results also showed that 3D roof segments had added value as input for the quick-scan PV yield prediction methods since the precision of yield prediction was significantly lower using only 2D land register data of buildings.
...
A quick-scan yield prediction method has been developed to assess rooftop photovoltaic (PV) potential. The method has three main parts. For each roof, first (i) virtual 3D roof segments were reconstructed using aerial imagery, then, (ii) PV modules were automatically fitted onto roof segments using a fitting algorithm and finally, (iii) expected annual yield was calculated. For each roof, the annual yield was calculated by three different quick yield calculation approaches. Two approaches are commercial software packages of Solar Monkey (SM) and Photovoltaic Geographical Information System (PVGIS) whereas the other one is the simplified skyline-based approach developed in photovoltaic material and devices (PVMD) group of Delft University of Technology. To validate the quick-scan method, a set of 145 roofs and 215 roof segments were chosen in urban areas in the Netherlands. For the chosen roofs, the number of fitted modules and calculated yield were compared with the actual modular layout and the measured yield of existing PV systems. Results showed a satisfactory agreement between the quick-scan yield prediction and measured annual yield per roof, with relative standard deviations of 7.2%, 9.1%, and 7.5% respectively for SM, PVGIS, and PVMD approaches. It was concluded that the obstacle-including approaches (e.g. SM and PVMD) outperformed the approaches which neglect the shading by surrounding obstacles (e.g. PVGIS). Results also showed that 3D roof segments had added value as input for the quick-scan PV yield prediction methods since the precision of yield prediction was significantly lower using only 2D land register data of buildings.
