Explainable Deep Learning Localizes Photovoltaic Array Faults With 99.8 Percent Accuracy

Engineers have designed an explainable deep temporal learning network, termed ECBL-Net, capable of dynamically identifying and localizing electrical and physical faults within large-scale photovoltaic arrays with 99.84 percent classification accuracy using solely raw time-series sensor data. As photovoltaic installations expand globally, solar arrays face operational vulnerabilities from environmental stressors and electrical anomalies, including arc faults, module soiling, open circuits, and inverter failures. These malfunctions degrade power output, reduce equipment lifespan, and generate substantial maintenance overhead when left undetected. Traditional monitoring protocols frequently fail to provide the temporal granularity or interpretability needed for safety-critical solar management. To address this operational gap, the new convolutional bidirectional architecture eliminates expensive emulator hardware and artificial feature engineering. Instead, the model couples automated sequence learning with post-hoc Shapley additive explanations to deliver transparent diagnostics directly from operational electrical measurements, establishing a dependable framework for automated maintenance in renewable utility grids.

The development of ECBL-Net directly resolves persistent limitations in conventional solar diagnostic algorithms and recent machine learning methodologies. Traditional supervisory control systems rely heavily on hardcoded operational thresholds, static statistical inference, and handcrafted feature pipelines that cannot adapt to changing solar irradiance or fluctuating ambient temperatures. While standard machine learning algorithms such as support vector machines, decision trees, and multilayer perceptrons achieve acceptable baseline classification, they operate as spatially fixed classifiers that treat each incoming data sample independently, ignoring the sequential progression of developing faults. More recent deep learning approaches combining convolutional neural networks with recurrent units have demonstrated improved detection rates, yet they function predominantly as opaque black boxes without diagnostic transparency. Furthermore, many competing systems depend on complex physical simulation layers, synthetic grid emulators, continuous curve tracing, or visual scalograms derived from continuous wavelet transforms, all of which introduce computational latency that prevents real-time deployment on standard photovoltaic sensor streams.

To build a realistic and robust diagnostic system, the researchers trained and evaluated ECBL-Net using real-world operational sensor recordings retrieved from the public photovoltaic data acquisition platform PVDAQ. Specifically, the team examined operational data collected from PVDAQ Systems 2105 and 2107, hosted through the Open Energy Information repository under a public domain dedication. Unlike synthetic benchmarks generated by computer simulations, these datasets capture genuine physical phenomena from active solar installations operating under natural environmental fluctuations. The underlying sensor infrastructure continuously logs synchronized inverter-level electrical variables and meteorological parameters at a high temporal resolution of one-second intervals. Across the monitored installations, the research team isolated six core sensor dimensions: inverter direct-current voltage, direct-current input current, metered power output, solar irradiance, module surface temperature, and ambient air temperature. These channels were selected due to their acute sensitivity to electrical abnormalities, such as sudden voltage collapses during ground faults or disproportionate power drops during localized inverter disruptions.

Before feeding the sensor streams into the neural architecture, the raw observational records underwent a structured preprocessing procedure to ensure temporal alignment and numerical consistency. Sensor channels were synchronized along a unified Coordinated Universal Time index. Because communication disruptions occasionally interrupted signal logging, the researchers applied linear interpolation across feature channels, noting that missing values accounted for approximately 0.35 percent of recorded values across both photovoltaic systems. Linear interpolation was preferred over forward-filling or spline interpolation because it maintains smooth operational transitions without introducing synthetic artifacts that could mask abrupt electrical transients. To eliminate high-frequency noise generated by inverter switching jitter and measurement fluctuations, each channel was processed through a fourth-order low-pass Butterworth filter configured with an empirical cutoff frequency of 0.1 hertz. The preprocessed time series was subsequently segmented into overlapping sliding temporal windows of 120 seconds with a 30-second stride, after which each window was normalized using channel-wise min-max scaling.

The structural framework of ECBL-Net operates as an end-to-end deep temporal pipeline consisting of four interconnected processing stages designed to preserve both feature representation and interpretability. Incoming normalized windows pass first into a one-dimensional convolutional temporal encoding layer configured with 64 filters, a kernel size of five timesteps, a stride of one, and rectified linear unit activations. This unpadded convolutional block systematically extracts fine-grained localized signatures, capturing rapid voltage dips, current spikes, and irradiance changes that signal the immediate onset of an electrical disturbance. The resulting temporal feature maps are channeled into a bidirectional temporal abstraction layer composed of a two-layer bidirectional long short-term memory network featuring 64 hidden units in each direction. By simultaneously processing temporal sequences in forward and backward directions, the recurrent layer concatenates hidden representations into a 128-dimensional latent vector at each timestep, capturing both retrospective conditions leading up to a failure and subsequent recovery dynamics across the observation window.

To identify the precise moments when operational anomalies manifest, the architecture incorporates a single-head additive temporal attention mechanism directly above the bidirectional latent representations. Unlike multi-head transformer mechanisms that compute dispersed contextual scores, this temporal attention module generates a single dynamic relevance weight for every timestep using a shared context vector and a learnable projection matrix. To enhance interpretability and eliminate noisy background fluctuations, the researchers integrated a temporal attribution masking layer during backpropagation that explicitly discards timesteps where attention weights fall below a significance threshold of 0.05. Model training was conducted using the Adam optimizer with an initial learning rate of 0.001, accompanied by dropout regularization set at 0.3 and early stopping with a 10-epoch patience threshold. Crucially, the training regimen introduced an interpretability-aware gradient suppression regularizer, weighted by a factor of 0.05, which penalizes divergence between attention gradients across sample pairs belonging to the same fault class.

A core distinction of ECBL-Net is its unified explainability framework, which couples internal temporal attention scores with post-hoc feature attribution using the DeepSHAP algorithmic implementation. By computing Shapley additive explanations over attention-weighted latent vectors, the network produces a localized relevance matrix that assigns an exact marginal importance score to each of the six sensor channels at every individual second. To ensure that post-hoc explanations accurately reflect internal decision dynamics, the calculated Shapley values are synchronized with temporal attention scores, establishing that a timestep constitutes credible diagnostic evidence only when attention intensity and feature attribution coincide. Furthermore, the researchers established a quantitative clarity measure based on a domain significance threshold to track the temporal concentration of explanations within active fault intervals. This dual-verification strategy overcomes the fundamental limitations of post-hoc surrogate methods like LIME, which lack temporal resolution, and gradient saliency maps, which frequently suffer from gradient saturation when applied to noisy time-series signals.

Rigorous quantitative evaluation demonstrated that ECBL-Net delivers exceptional multi-class classification performance across five distinct operational fault categories and normal operation. The network was evaluated against baseline conditions comprising No Fault, Open Circuit, Arc Fault, Soiling, Inverter Error, and Partial Shading across five independent experimental runs with randomized seed initializations. The proposed architecture attained an overall classification accuracy of 99.84 percent, an average precision of 99.87 percent, an average recall of 99.46 percent, and an overall F1-score of 99.65 percent. In comparative benchmarks against existing solar diagnostic architectures, ECBL-Net surpassed hybrid convolutional long short-term memory networks and multi-strategy whale optimization fusion networks by more than 1.2 percent in overall classification accuracy. Receiver operating characteristic curves revealed area-under-the-curve metrics exceeding 0.97 across every modeled fault category, with near-perfect boundary separation observed for open circuits and arc faults. Even for closely related operational conditions exhibiting overlapping electrical characteristics, the model maintained high diagonal dominance across confusion matrices.

Beyond aggregate classification metrics, empirical testing confirmed the system’s high temporal localization precision by measuring the Mean Absolute Localization Error between predicted and actual fault onset moments. For sharp electrical disruptions such as Arc Faults, ECBL-Net achieved a remarkable Mean Absolute Localization Error of just 3.2 seconds, followed closely by Inverter Errors at 4.1 seconds. Conversely, gradual environmental disruptions such as Soiling and Partial Shading exhibited larger temporal error spreads, reflecting the naturally diffuse onset and ambiguous signal progression inherent to progressive dust accumulation and moving cloud cover. Analysis of attention concentration through the Attention Sharpness Index and Mean Absolute Deviation further corroborated these findings. Arc faults registered an Attention Sharpness Index of 0.86 and a low Mean Absolute Deviation of 0.078, indicating tightly clustered temporal focus around transient current spikes. In contrast, normal operating states registered diffuse attention distributions with an Attention Sharpness Index of 0.51 and a Mean Absolute Deviation of 0.146.

Computational benchmarking verified that ECBL-Net achieves practical deployment viability for continuous real-time monitoring without requiring specialized supercomputing infrastructure. With a total memory footprint of less than 10 megabytes and minimal inference latency across tested 120-second temporal windows, the trained architecture satisfies the hardware constraints of standard industrial edge controllers and embedded graphical processing units deployed at photovoltaic substations. Nevertheless, the authors highlighted several engineering boundaries that warrant subsequent investigation before full-scale commercial adoption. The current empirical validation relied exclusively on two operational photovoltaic installations from the PVDAQ database, meaning that generalization across alternative solar panel manufacturing technologies, diverse inverter topologies, and disparate geographic climate zones remains to be proven. Additionally, real-world deployment must address long-term sensor degradation, environmental noise beyond the filter passband, and unprecedented combinations of compounding faults. For ultra-low-power microcontroller platforms operating at the edge, model compression techniques such as weight quantization or knowledge distillation will be required to maintain real-time latency.

Subject of Research: Explainable deep temporal learning for dynamic fault localization in large-scale photovoltaic arrays

Article Title: An explainable learning architecture for dynamic fault localization in large scale photovoltaic arrays

Article References: Subha, K., & Sharanya, S. (2026). An explainable learning architecture for dynamic fault localization in large scale photovoltaic arrays. Discover Artificial Intelligence, 6(1), Article 1786. https://doi.org/10.1007/s44163-026-02487-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02487-3

Keywords: ECBL-Net, photovoltaic fault diagnosis, deep temporal modeling, time-series sensor data, PVDAQ dataset, dynamic fault localization, explainable artificial intelligence, temporal attention mechanism, DeepSHAP, bidirectional LSTM, one-dimensional convolution, arc fault detection, solar inverter monitoring, Butterworth filter, gradient suppression regularization, attention sharpness index, mean absolute localization error, edge computing solar analytics