| dc.description.abstract | The increasing demand for PV (photovoltaic) modellers brings forward the need for a clear and comprehensive assessment of the most applied PV prediction models. Such an overview should consist of both computer simulation models and machine learning models, as the latter has expanded to the field of PV assessment. Comparative studies have so far been incomplete, limited in accuracy assessment and have never compared simulation modelling with machine learning techniques.
This comparative study determined the modelling accuracy for simulation models PVLib, PVSyst, SAM, PVWatts and Helioscope and for eight different machine learning models. The accuracy is determined by comparing the modelled with the measured DC power output of a commercial PV module, for which the meteorological and performance data is obtained from the Utrecht Photovoltaic Outdoor Test facility. The accuracy is evaluated on a macro- and microlevel, which differentiates between the error of annual electricity yields and the aggregated errors for each data point. The differentiation between the macro- and micro- accuracy provides further insights in a model’s optimal application. In addition, the influence of the source of meteorological data, type of solar input irradiance and the resolution of input data on a model’s accuracy is determined as well. The sensitivity of the machine learning accuracies to the amount of training data is also determined. Every modelling step is elaborately described to ensure absolute transparency and examination of the model configurations.
It is concluded that four different model combinations under PVLib are the most accurate on both the macro-and microlevel. SAPM is found to most accurately model from global panel-of-array irradiance and the combination of Physical and FSSC the most accurate using global horizontal irradiance. PVSyst and SAM also obtained decent micro-accuracies, but they generally underestimated the electricity yields. The machine learning models proved to accurately predict electricity yields, but generalisation and wider application require more research.
The influence of the source of meteorological data and the type of solar input irradiance influenced a model’s micro-accuracy but were not found to consistently influence the macro-accuracy. The time resolution of meteorological data is found to slightly influence both macro- and micro-accuracy. The minimum amount of training data needed for all used data sets that guaranteed decent machine learning accuracies was found to be 8 months.
The results found in this comparative study facilitate in selecting the most suitable PV prediction model for each objective, incorporating both simulation and machine learning options. | |
