Solar photovoltaics (PV) are a promising candidate for utility-scale use due to the low cost per mWh of silicon-PV and large available solar resource in the continental United States. At the utility scale, the last few years have seen a massive increase in deployment of Single Axis Trackers (SATs), which rotate a PV module east-west so that it always faces towards the Sun. However, the effect of SAT failure on the costs of utility scale PV is not well studied, mainly because data on tracker failure in field is either nonexistent or proprietary. We simulate tracker failure in a capacity expansion model to quantify the costs of theoretical failure modes, including tracker stall, improper backtracking, and other reasonable failure conditions. This work presents a capacity expansion model that quantifies the losses incurred by these failures, in terms of costs of required additional generation to meet future demand. We rank failures from most costly to least, and provide operational recommendations to operators to mitigate potential failure.

In recent years, as the total photovoltaic (PV) capacity in the power grid continues to increase, PV installations have become more diverse than the simpler fixed-tilt, utility-scale systems of the past. Systems are installed on rooftops, trackers, deployed with batteries, and in many other configurations. Virtual models of a system, called “digital twins,” are often constructed to predict future performance, quantify losses, schedule O&M, etc. However, due to increasing system diversity and complexity, constructing these digital twins from traditional physical models (such as PVSyst or the Sandia Array Performance Model) can be quite time consuming, prone to modeler error, and tend to require detailed, expensive site-specific metadata. Therefore, models that can learn a representation of the system from data (called “data-driven digital twins”) can be used alongside or in place of traditional, end-to-end physics-based models. In this work, we present a comparison of a empirical digital twin (using pvlib-python physics-based models) and a data-driven digital twin (using spatiotemporal graph neural networks) and quantify error metrics for the models. It is found that data-driven models best represent the average case of a PV system, whereas the empirical model better corresponds to oscillation due to weather variation. Data used in this work is open access on OSF, and reference implementations in Python are provided on GitHub.

In this work, we introduce and compare the results of several methods for determining the horizon profile at a PV site, and compare their use cases and limitations. The methods in this paper include horizon detection from time-series irradiance or performance data, modeling from GIS topology data, manual theodolite measurements, and camera-based horizon detection. We compare various combinations of these methods using data from 4 Regional Test Center sites in the US, and 3 World Bank sites in Nepal. The results show many differences between these methods, and we recommend the most practical solutions for various use-cases.

Bulk AlN single crystal has a great demand as a substrate material for AlGaN-based optical and electronic devices such as deep ultraviolet light-emitting diodes and high-power transistors. We have previously proposed a novel solution-growth method using Ni-Al solution with an in-situ observation system for solution growth of AlN crystal using electromagnetic levitation. In this paper, to investigate AlN formation behavior on 40 mol%Al-Ni droplets, two precisely synchronized high-speed cameras from the horizontal and vertical directions were installed. AlN formation behavior was evaluated quantitatively using computer vision image processing tech-niques. Based on the results, we demonstrated the growth of thick AlN film at 2030 K for 1 h. A 3.8-mu m-thick c -axis oriented AlN film successfully formed on the droplet, and the film was also oriented in-plane. The + c -di-rection AlN film grew towards the droplet center from the surface by reacting dissolved nitrogen and Al atoms.

In this work, we present two different algorithms to aid in real-time weather predictions. This information can be used to inform the movement of a tracker or short-term power predictions. Since cloud cover significantly affects the resulting insolation on a PV module, identifying and tracking cloud motion is useful to this end. This work presents a convolutional autoencoder (CAE) to identify clouds and a particle tracker to predict cloud movement. The CAE model integrates information from multiple approaches to cloud segmentation. Particle tracking is useful in areas such as Albuquerque, NM where clouds move in smaller fragments due to rapid variance in wind direction caused by nearby mountains. By combining neural networks and more classical technologies, the system becomes more robust and explainable then either image processing or pure neural network technologies, respectively.

This article presents a notable advance toward the development of a new method of increasing the single-axis tracking photovoltaic (PV) system power output by improving the determination and near-term prediction of the optimum module tilt angle. The tilt angle of the plane receiving the greatest total irradiance changes with Sun position and atmospheric conditions including cloud formation and movement, aerosols, and particulate loading, as well as varying albedo within a module's field of view. In this article, we present a multi-input convolutional neural network that can create a profile of plane-of-array irradiance versus surface tilt angle over a full 180 $^{\circ }$ arc from horizon to horizon. As input, the neural network uses the calculated solar position and clear-sky irradiance values, along with sky images. The target irradiance values are provided by the multiplanar irradiance sensor (MPIS). In order to account for varying irradiance conditions, the MPIS signal is normalized by the theoretical clear-sky global horizontal irradiance. Using this information, the neural network outputs an N -dimensional vector, where N is the number of points to approximate the MPIS curve via Fourier resampling. The output vector of the model is smoothed with a Gaussian kernel to account for error in the downsamping and subsequent upsampling steps, as well as to smooth the unconstrained output of the model. These profiles may be used to perform near-term prediction of angular irradiance, which can then inform the movement of a PV tracker.

Cloud cover significantly affects the solar irradiance incident on a photovoltaic (PV) module, so identifying and predicting cloud motion is useful for PV applications, such as informing tracker movements and predicting short-term power. This work presents two algorithms to aid in real-time weather predictions, i.e., a convolutional autoencoder (CAE) to identify clouds and a particle tracker to predict cloud movement. The CAE model integrates information from multiple cloud segmentation approaches, and then utilizes transfer learning on these unreliable, automatically generated masks to bootstrap model performance. The presented model improves upon the state-of-the art metrics with a resultant pixelwise accuracy greater than 90% while remaining lightweight in number of samples used. For tracking and prediction of cloud movements, particle tracking is useful in areas where cloud coverage is transient and clouds move in smaller fragments. By combining neural networks and more classical image processing techniques, the system becomes more robust and explainable than image processing or pure neural network technologies alone, while also demonstrating the power of transfer learning techniques in application.

Cell cracking in PV modules can lead to a variety of changes in module operation, with vastly different performance degradation based on the type and severity of the cracks. In this work, we demonstrate automated measurement of cell crack properties from electroluminescence images, and correlate these properties with current-voltage curve features on 35 four-cell Al-BSF and PERC mini-modules showing a range of crack types and severity. Power loss in PERC modules was associated with more total crack length, resulting in electrical isolation of cell areas and mild shunting and recombination. Many of the Al-BSF modules suffered catastrophic power loss due to crack-related shunts. Mild power loss in Al-BSF modules was not as strongly correlated with total crack length; instead crack angles and branching were better indicators of module performance for this cell type.

Cell cracking in PV modules can lead to a variety of changes in the modules operation, with vastly different performance degradation based on the type and severity of crack. In this work, we correlate cell crack metrics in images with current-voltage (I-V) curve features on a sample set of 38 four-cell Al-BSF and PERC mini-modules showing a range of fracture and electrical properties. Impacts of cracking on electrical performance demonstrated in this work include cell shunting and electrical isolation of cell regions, shown with electroluminescence (EL) images and I-V tracing. Cracks and other EL image properties are quantified with algorithmic computer vision techniques, and correlated with I-V properties at the cell and module levels.

Electroluminescence (EL) imaging of photovoltiac (PV) modules offers high-speed, high-resolution information about device performance, affording opportunities for greater insight and efficiency in module characterization across manufacturing, research and development, and power plant operations and management. Predicting module electrical properties from EL image features is a critical step toward these applications. In this article, we demonstrate quantification of both generalized and performance mechanism-specific EL image features, using pixel intensity-based and machine learning classification algorithms. From EL image features, we build predictive models for PV module power and series resistance, using time-series current-voltage (I-V) and EL data obtained stepwise on five brands of modules spanning three Si cell types through two accelerated exposures: damp heat (DH) (85 degrees C/85% RH) and thermal cycling (TC) (IEC 61215). In total, 195 pairs of EL images and I-V characteristics were analyzed, yielding 11 700 individual PV cell images. A convolutional neural network was built to classify cells by the severity of busbar corrosion with high accuracy (95%). Generalized power predictive models estimated the maximum power of PV modules from EL images with high confidence and an adjusted-R-2 of 0.88, across all module brands and cell types in extended DH and TC exposures. Mechanistic degradation prediction was demonstrated by quantification of busbar corrosion in EL images of three module brands in DH, and subsequent modeling of series resistance using these mechanism-specific EL image features. For modules exhibiting busbar corrosion, we demonstrated series resistance predictive models with adjusted-R-2 of up to 0.73.

In order to characterize the degradation of solar modules, it is necessary to systematically recognize common signatures, such as cracks and corrosion. This information can be useful in a multitude of ways, such as comparing failure models of different brands or types of cells. The process of electroluminescence imaging can illuminate these signatures such that automated image processing techniques can identify them. Both supervised (via C.N.N.) and unsupervised models can be applied; the focus of this work is on improvement of the unsupervised approach. Feature extraction, a pivotal step to computer vision, is used to produce localized descriptions of relevant regions in an image. In order to identify features, numerous feature extraction algorithms, such as ORB [12] and FAST [11] have been applied, which yield feature vectors that are invariant to location, scale, and orientation, as opposed to Haralick features [4]. The resulting feature vectors can be clustered via hierarchical clustering and a bag of visual words can be constructed that can identify similar features across a sample set. This method allows clustering around a large number of unidentified features both apparent and non-obvious to human inspection, including cracking, busbar corrosion, and other common forms of degradation. The resultant model and code implementation can be used as a form of exploratory data analysis before labeling in preparation for supervised machine learning, or as a classifier on its own.