LAUSR

Abstract Accurate and reliable modeling of photovoltaic (PV) modules is significant for optimal design, operation and evaluation of PV systems. PV models can be classified into equivalent circuit-based white box models and data-driven black box models. Due to the difficulty to obtain the ground true model parameters and the limitation posed by the predetermined model structure, white-box modeling methods generally suffer relatively low accuracy and generalization performance for arbitrary operating conditions. In addition, reported black-box models are based on the conventional artificial neural networks (ANN) that are efficient but have limited performance. In this study, motivated by the high performance of fast developing deep learning techniques, we propose a novel black-box modeling method for the PV modules using a new modified one-dimensional deep residual network (1-D ResNet) and measured I-V characteristic curves, which can predict a whole I-V curve at a time for arbitrary operating conditions. To alleviate the overfitting issue caused by imbalanced data, original I-V curve datasets with highly non-uniform operating conditions are resampled by a grid sampling approach to obtain the datasets with relatively uniform conditions for the subsequent modeling. The proposed 1-D ResNet based model is comprehensively verified and compared with a proposed single-diode based white-box model and three other conventional ANN based black-box models, using large datasets of measured I-V characteristic curves from the National Renewable Energy Laboratory (NREL). Experimental results indicate that black-box models are generally better than the white-box model. Especially, the proposed 1-D ResNet based PV model is obviously superior to other three conventional ANN based black-box models, in terms of accuracy, generalization performance and reliability.