Performance Analysis of Mechanistic and Machine Learning Models for Photovoltaic Energy Yield Prediction
Date
2019Author
![ORCID logo](https://orcid.org/sites/default/files/images/orcid_16x16.png)
![ORCID logo](https://orcid.org/sites/default/files/images/orcid_16x16.png)
![ORCID logo](https://orcid.org/sites/default/files/images/orcid_16x16.png)
Sutterlueti, Juergen
Ransome, Steve
![ORCID logo](https://orcid.org/sites/default/files/images/orcid_16x16.png)
Place of publication
Marseille, FranceSource
36th European PV Solar Energy Conference, EUPVSEC 2019, 9-13 SeptemberGoogle Scholar check
Metadata
Show full item recordAbstract
In this work, the prediction performance of the mechanistic performance model (MPM) and a machine learning Feed Forward Neural Network (FFNN), was evaluated using yearly datasets, containing sixty-minute average and instantaneous field measurements obtained from the outdoor test sites in Nicosia, Cyprus and in Arizona, USA, respectively. The model exhibiting the lowest prediction error was derived based on different model training conditions. The obtained results demonstrated that both models provide good predictive quality using both instantaneous and average measurements. The performance of the models was strongly dependent on the duration of the train set, since for a random 70:30 % train and test set split using the yearly dataset from the GI OTF, a mean absolute percentage error (MAPE) of 1.95 % and 1.55 % was obtained for the MPM and FFNN, respectively. Alternatively, for a random 30:30 % train and test set data partition, the MPM and the FFNN achieved a MAPE of 2.03 % and 1.67 %, respectively. Moreover, by applying a medium irradiance condition filter (GI>0.3 kW/m 2) to the dataset during a random 70:30 % train and test set approach, the MPM and the FFNN achieved a MAPE of 1.77 % and 1.37 %, respectively, demonstrating that the predictive accuracy of the models was enhanced by data filtering. Additionally, the MPM and the FFNN achieved the lowest MAPE of 2.05 % and 2.11 % when the train set (10 % of the entire dataset) contained 80 % of clear and 20 % of variable and diffuse measurements (same amount of weather type measurements as the amount of the irradiance profile classes of the location). Finally, for accurate predictions a random train and test approach should be utilized, the training process should be performed using the greatest possible amount of train data samples and by simultaneously applying an irradiance condition filter.