The problem the researchers tackled is one that plagues every real-world photovoltaic installation: missing data. Photovoltaic power plants continuously monitor voltage, current, power, temperature, and irradiance, feeding this stream into fault-detection algorithms. But communication blockages, sensor malfunctions, asynchronous triggering between devices, unstable sunlight, temperature drift, hidden component defects, and simple wiring errors can all interrupt the flow. When data gaps appear, most existing fault diagnosis methods simply fail to account for them — a blind spot that can turn a healthy array into a mystery box.

Di Liu, Xiaojuan Zhu, and Changyu Du of Beijing Fibrlink Communications built their solution as a pipeline of four cooperating technologies: Tucker decomposition, the sparrow search algorithm, the Informer network, and a novel multi-scale convolutional model they call MSCNet. The first job belongs to Tucker decomposition, a tensor-based mathematical technique that treats the high-dimensional photovoltaic data — power, voltage, current, irradiance, and timestamps — as a structured spatial object rather than a flat table. By transforming the data into a Hankel tensor through a standard delay transformation known as Hankification, the method exploits the inherent temporal structure of the measurements to reconstruct the missing values.

The tensor completion model is solved using alternating least squares, a procedure that iteratively refines the core tensor and its factor matrices until the reconstruction converges. Once the completed Hankel tensor is obtained, the data pass through a delayed inverse transformation using the Moore–Penrose pseudo-inverse, delivering a filled-in dataset that preserves the original temporal patterns. In randomized missing-data experiments, where the team removed between 5 and 50 percent of the current measurements in 5 percent steps, their method achieved a relative squared error of just 3.51 percent at the brutal 50 percent missing rate — outperforming cubic spline interpolation, K-nearest neighbors, support vector machines, and low-rank matrix filling by margins of 1.28 to 3.74 percentage points.

Continuous data loss, the kind that occurs when remote devices lose communication or sensors fail outright, proved even more challenging. At a 30 percent continuous missing ratio, the Tucker-based method delivered a relative squared error of 5.27 percent, while KNN, low-rank matrix, and SVM-based alternatives hovered between 7.98 and 10.28 percent. The team’s algorithm beat the comparison methods at every missing percentage tested, confirming that the tensor approach is not a one-trick trick but a robust foundation for the diagnostic pipeline that follows.

With the data restored, the Informer network takes over. Informer is a Transformer variant purpose-built for long time-series feature extraction, and it outperforms the standard Transformer architecture in this domain thanks to a probabilistic sparse self-attention mechanism. Rather than forcing every query to attend to every key in the sequence, Informer lets each focus only on the first n queries with the highest relevance, dramatically reducing computational cost while capturing the global patterns that matter. The encoder converts the input into dense vectors of fixed dimensions and extracts features in order, generating a mapping that preserves the long-range dependencies in photovoltaic operation data.

But global patterns alone are not enough. Fault signals in photovoltaic arrays are often subtle, localized, and multi-scale — a short circuit in one string, a shadow across a few modules, an aging cell drifting out of spec. To capture these granular features, the team embedded MSCNet into the Informer encoder. MSCNet operates through two parallel branches: one performs global max pooling and global average pooling, then passes the results through a fully connected layer and a sigmoid gating mechanism to recalibrate the weights of each feature channel. The other branch expands the number of channels and solves the gradient vanishing problem through convolution, batch normalization, and ReLU activation, then extracts multi-scale features using depthwise separable convolutions. Because depthwise convolutions ignore relationships between channels, a channel shuffling operation recombines the outputs, and the two branches merge through a fully connected layer to produce the final enhanced representation.

The third component, the sparrow search algorithm, optimizes the global parameters of the combined Informer-MSCNet model. Inspired by the foraging and anti-predatory behavior of actual sparrows, the algorithm divides its population into discoverers, joiners, and vigilantes. Discoverers find food and lead the group, joiners follow them, and vigilantes monitor the surroundings for danger. When a warning value exceeds a safety threshold, the sparrows move to a safe area, and the algorithm iteratively updates positions until the optimal configuration emerges. This swarm intelligence approach accelerates convergence of the deep learning model, which would otherwise require extensive manual tuning of its many parameters.

To test the system, the researchers built a real photovoltaic simulation platform in a city in northern China, comprising eight PV modules connected in a four-series-by-two-parallel configuration, along with resistor boxes, electricity meters, temperature sensors, and an irradiance meter. Data were collected at 15-minute intervals, producing a dataset of 4,320 by 5 sampling points with ambient temperatures between 25.47 and 34.69 degrees Celsius and irradiance between 527.25 and 765.14 watts per square meter. They simulated four single fault types — open circuit, line short circuit, abnormal aging, and partial shadow — and four compound fault types combining two simultaneous failures. After filling the missing data with their Tucker-based method and diagnosing with the full pipeline, the model achieved 98.73 percent accuracy on single faults, beating Informer, Transformer, TCN, SVM, and CNN by 3.05 to 15.07 percentage points.

Under noise, the advantages became even clearer. When 30 dB of noise was added to the compound fault scenario, the proposed model maintained 96.12 percent accuracy, improving on Informer, Transformer, TCN, SVM, and CNN by 4.61, 5.9, 11.51, 13.96, and 17.88 percentage points respectively. The recall rate of 96.76 percent similarly outpaced all comparison algorithms. Ablation experiments confirmed that MSCNet is the critical ingredient: when the team removed it and ran a TCN-Informer-SSA variant, accuracy on open-circuit faults dropped 16 percentage points, and the variant fell below 90 percent on several fault types that the full model handles with ease. The researchers conclude that their Tucker decomposition-SSA-Informer-MSCNet pipeline offers a practical, noise-robust foundation for smart grid construction — a system that can keep watch over the world’s solar farms even when the sensors go dark.

Subject of Research: A high-precision fault diagnosis method for photovoltaic arrays considering the effect of missing data

Article Title: A high-precision fault diagnosis method for photovoltaic arrays considering the effect of missing data

Article References: Liu, D., Zhu, X., & Du, C. (2025). A high-precision fault diagnosis method for photovoltaic arrays considering the effect of missing data. Frontiers in Electronics, 6, Article 1656864. https://doi.org/10.3389/felec.2025.1656864

Image Credits: AI Generated

DOI: 10.3389/felec.2025.1656864

Keywords: high-precision, fault, diagnosis, method, photovoltaic, arrays, considering, effect, missing, data, scientific research