A research team led by the University of Sharjah in the United Arab Emirates has developed a novel machine learning approach for fault detection in bifacial PV systems.

The method combines a one-dimensional convolutional neural network (1D-CNN) with the random forest (RF) ensemble learning method to identify specific faults such as dust, shading, aging, and cracks. The latter is a machine learning approach that utilizes multiple decision trees to improve predictive accuracy for both classification and regression tasks, while the former is a feedforward neural network that learns features via filter optimization.

“This study pioneers a multi-fault classification framework for bifacial PV systems, integrating mathematical modeling of dual-sided fault impacts on current–voltage (I-V) curves and a 180-day synthetic dataset with randomized severities,” the group said. “Combining I-V curve and maximum power point-based power profile analyses, it demonstrates bifacial resilience, achieving up to 100% power advantage under severe shading and 11.7–30% superior outputs across faults.”

The academics started by developing a mathematical framework for modeling the I-V characteristics of monofacial and bifacial PV modules. The simulated monofacial module has a short circuit current of 9.30 A, an open circuit voltage of 46.86 V, a maximum power point current (Impp) of 8.72 A, a maximum power point voltage (Vmpp) of 38.26 V, and maximum power point power (Pmpp) of 333.8 W. The front side of the bifacial module had values of 9.09 A, 46.89 V, 8.57 A, 38.32 V, and 328.4 W, respectively. The bifaciality factor was 0.8.

For both module types, the baseline performance is established by simulating regular operation. Further formulas were established to simulate various fault scenarios, namely shading effect, dust accumulation, degradation effect, aging effect, and cracks. All of these profiles were simulated for 12 hours, from 6:00 to 18:00, at 15-minute intervals, reflecting a typical solar day with realistic irradiance and temperature dynamics.

“The Stage 1 dataset comprises Baseline, which consists of 90 days, 4,410 samples, allocating 50.1% of the dataset, Obstruction consists of 60 days, 2,940 samples, allocating 33.3% of the dataset, and Degradation, which consists of 30 days, 1,470 samples, allocating 16.6% of the dataset,” the academics explained. “For Stage 2, the 2,940 Obstruction samples are distributed as Shading (592 samples, 20.1%), Dust (588 samples, 20.0 %), Cracks (585 samples, 19.9%), None (585 samples, 19.9%), and Aging (594 samples, 20.2%). This distribution reflects realistic fault prevalence, with Degradation and Cracks as minority classes.”

For each block, 80% of the samples are randomly assigned to train the CNN-RF model, and 20% are allocated to testing. In the combination, the 1-D CNN is responsible for extracting features such as patterns in current, voltage, power, irradiance, and others. At the same time, the RF receives these extracted features as input and performs the final classification. They both operate at both stages – the first, a general operating state, and the second, a specific fault type.

“In Stage 1, it achieves 100% accuracy in classifying baseline, degradation, and obstruction states,” the results showed. “In Stage 2, it identifies specific faults such as none, dust, shading, aging, and cracks with 97.6% accuracy, an area under the curve of 0.999, and a false positive rate of 0.006, surpassing the standalone CNN with 89.7% accuracy, RBF-SVM with 96.1%, and GBRT with 91.8%.”

The results have appeared in “Enhancing fault detection in bifacial photovoltaic systems: a two-stage CNN-RF approach with I-V curve analysis,” published in Energy Conversion and Management: X. Scientists from the United Arab Emirates’ University of Sharjah and the United Arab Emirates University have participated in the study.