We present a comprehensive study on intelligent fault diagnosis technology for solar panel modules, addressing the complex challenges in modern photovoltaic power stations. Our work focuses on a 10MW distributed solar panel plant that encountered severe composite faults including hot spots, power degradation, and poor contact. By integrating deep learning models with digital twin visualization, we developed a closed-loop operation and maintenance system that significantly improves diagnostic accuracy and operational efficiency. This paper details the fault characteristics, the hybrid CNN-LSTM-Attention model, quantitative results, and the practical outcomes achieved.
1. Fault Characteristics and Root Cause Analysis
In our 10MW distributed solar panel station, comprising 41,600 solar panel modules, routine monitoring revealed a persistent 18% decline in power generation for inverter No.3 compared to historical levels. The inverter frequently triggered branch current imbalance alarms. Initial manual inspections found no obvious cable damage or junction box faults, but the anomalies exhibited dynamic fluctuations with varying irradiance and ambient temperature. We thus suspected composite faults of solar panel modules. On-site thermal imaging and electrical parameter measurements confirmed three coexisting fault types: hot spots, power degradation, and poor contact, forming a coupled “hot spot + power degradation + poor contact” scenario.
1.1 Identification of Fault Types
Based on infrared thermography and time-series electrical data, we classified the faults into three fundamental types and their composite forms:
| Fault Type | Primary Cause | Key Signature |
|---|---|---|
| Hot spot | Partial shading (e.g., sand/dust deposition) causing local heating | Local temperature rise 15–25°C above normal; dynamic position with sun angle |
| Power degradation | Encapsulation aging, backsheet yellowing, cell microcracks | Sustained output power >15% below rated; simultaneous drop in open-circuit voltage and short-circuit current |
| Poor contact | Diode failure, connector resistance increase, terminal loosening | Irregular intermittent current fluctuations up to ±12% of rated value |
In our station, these three faults overlapped in the same area, creating a dynamic, coupled composite fault with much higher diagnostic complexity than single faults.
1.2 Characteristics of Composite Faults
Through comprehensive analysis, we identified three core challenges in diagnosing composite faults of solar panel modules:
Feature coupling and superposition: Different fault signatures interact physically and electrically. For instance, local heating from a hot spot accelerates encapsulation aging, triggering power degradation. Poor contact induces current fluctuations that alter the operating point, affecting hot spot formation. In data, this appears as simultaneous abnormal branch current, power curve decline, and temperature peaks—single-dimension features cannot isolate the dominant source.
Dynamic environmental dependence: Fault manifestation is strongly modulated by environment. Shading-caused hot spots move with the sun; poor contact may improve due to thermal expansion at high temperature or worsen under vibration. This leads to non-stationary, time-varying features, rendering static threshold methods prone to false positives or missed detection.
Ambiguity in diagnosis and localization: For example, a reduced branch current could stem from one severely degraded solar panel in the string or from poor contact at the end of the string—current alone cannot pinpoint the faulty module. An infrared hot spot could be either a genuine hot spot or a hot connector from poor contact, requiring combined analysis of electrical time-series data.
2. The Intelligent Diagnosis Model
To tackle these challenges, we constructed a hybrid diagnostic model integrating Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Attention mechanism. This model achieves deep fusion of spatial and temporal heterogeneous data and adaptive focusing on key fault features.
2.1 Spatial-Temporal Feature Extraction
The model uses a dual-channel parallel architecture. The spatial channel takes infrared thermal images of solar panels as input, passing through multiple CNN convolutional layers to extract spatial features such as hot spot shape, location, and temperature gradient. The convolution operation is expressed as:
$$ \mathbf{F}^{(l)} = \sigma \left( \mathbf{W}^{(l)} * \mathbf{F}^{(l-1)} + \mathbf{b}^{(l)} \right) $$
where \(\mathbf{F}^{(l)}\) is the output feature map at layer \(l\), \(\mathbf{W}^{(l)}\) is the convolution kernel, \(\mathbf{b}^{(l)}\) is the bias, \(\sigma\) is the ReLU activation, and \(*\) denotes convolution.
The temporal channel receives time-series electrical parameters (current, voltage, power) from the same monitoring point and feeds them into an LSTM network. LSTM units extract temporal dependencies such as current fluctuation trends and power decay slopes through gating mechanisms (input gate \(i_t\), forget gate \(f_t\), output gate \(o_t\)):
$$ \begin{aligned}
f_t &= \sigma_g (W_f x_t + U_f h_{t-1} + b_f) \\
i_t &= \sigma_g (W_i x_t + U_i h_{t-1} + b_i) \\
\tilde{C}_t &= \tanh(W_c x_t + U_c h_{t-1} + b_c) \\
C_t &= f_t \odot C_{t-1} + i_t \odot \tilde{C}_t \\
o_t &= \sigma_g (W_o x_t + U_o h_{t-1} + b_o) \\
h_t &= o_t \odot \tanh(C_t)
\end{aligned} $$
where \(x_t\) is input at time \(t\), \(h_{t-1}\) is previous hidden state, \(C_t\) is cell state, \(\tilde{C}_t\) is candidate state, \(W,U,b\) are weights and biases, \(\odot\) denotes element-wise multiplication, and \(\sigma_g\) is the sigmoid function. The LSTM generates deep temporal features that capture the non-stationary behavior of solar panel faults.
2.2 Attention Mechanism for Feature Enhancement
After fusing the deep features from CNN and LSTM, we introduce an attention mechanism to assign adaptive weights to different features. This strengthens features with high diagnostic contribution (e.g., hot spot temperature difference, current harmonic distortion) and suppresses environmental noise. The attention weights are computed as:
$$ \alpha_i = \frac{\exp\left( v_a^T \tanh(W_a h_i + b_a) \right)}{\sum_{j=1}^{N} \exp\left( v_a^T \tanh(W_a h_j + b_a) \right)} $$
and the weighted feature vector is:
$$ \mathbf{F}_{att} = \sum_{i=1}^{N} \alpha_i \mathbf{h}_i $$
Here, \(\alpha_i\) is the attention weight for the \(i\)-th element \(\mathbf{h}_i\) of the fused feature vector, \(\mathbf{h}_j\) is the \(j\)-th element, \(N\) is the number of elements, and \(W_a, b_a, v_a\) are trainable parameters. The resulting \(\mathbf{F}_{att}\) emphasizes the most informative features for fault diagnosis.
2.3 Fault Classification and Localization
The attention-weighted feature vector is fed into a fully connected layer followed by a Softmax classifier to output the probability distribution over fault types, along with a confidence score. A fault localization module maps the activation regions in the CNN feature maps to the digital twin of the solar panel plant, achieving component-level localization (e.g., inverter No.4, combiner box No.3, string No.10, solar panel No.5).
2.4 Model Training and Performance Evaluation
We trained the model on 12,000 samples (40% composite faults, 60% single faults) from historical data and simulated solar panel scenarios, using Adam optimizer with learning rate \(\eta=0.001\) for 100 epochs. The test set consisted of 2,000 real field measurements. The fault recognition accuracy comparison is shown below.
| Model | Composite Fault Accuracy | Single Fault Accuracy | Localization Error | Diagnosis Time |
|---|---|---|---|---|
| SVM | 78.3% | 89.5% | 3–5 solar panels | 12 s/sample |
| CNN only | 85.6% | 94.1% | 2–3 solar panels | 8 s/sample |
| CNN-LSTM | 93.2% | 97.5% | 1–2 solar panels | 5 s/sample |
| CNN-LSTM+Attention | 96.8% | 98.2% | ≤1 solar panel | 4 s/sample |
The results demonstrate that our proposed model outperforms conventional SVM and single CNN by significant margins, achieving 96.8% composite fault accuracy and localization error within one solar panel, with a diagnosis time of only 4 seconds per sample—meeting the real-time requirements of field operation.
3. Fault Handling and Effectiveness Analysis
3.1 Digital Twin-Assisted Maintenance Solution
We integrated visualization techniques into the fault handling process, forming a closed-loop operation and maintenance system with four layers: perception, diagnosis, action, and review. Based on the precise layout of the solar panel station, we created a 1:1 digital twin model. This model offers several core functions:
Fault visualization: Uses three colors (red for severe, orange for moderate, yellow for mild) to label faulty solar panel components. Clicking on any label reveals a detailed diagnostic report including fault type, confidence, and severity.
Intelligent path planning: Automatically calculates and displays the optimal inspection route based on fault distribution and solar panel station terrain, annotating key points and safety precautions.
Standardized action guidance: Pushes tailored action plans for each fault type: e.g., for hot spots, “clean shading or replace solar panel”; for poor contact, “tighten connectors and test circuit”; and includes simulation of the expected effect after action.
Historical trend analysis: Links historical fault records to visualize high-frequency fault zones and recurrence trends, supporting preventive maintenance decisions.
This solution deeply integrates intelligent diagnosis with visual digital twin, significantly improving the precision and efficiency of fault localization and repair for solar panel systems.
3.2 System Deployment and Site-Wide Inspection
During deployment, we installed three edge servers in the solar panel station’s control center, running lightweight diagnostic models for real-time data preprocessing and alarm generation, with average response time ≤10 seconds. Simultaneously, we set up a cloud data center and visualization platform, completing full-system data integration for real-time monitoring and fault visualization.
After deployment, we conducted the first site-wide intelligent inspection. Traditional manual inspection of all 41,600 solar panels took about 48 hours. In contrast, the intelligent system, using drones with infrared cameras and multi-channel electrical parameter collectors, completed full scanning in only 8 hours—an 83.3% efficiency improvement. The inspection automatically identified 87 fault points, including 32 composite faults (20 hot spot + power degradation, 10 hot spot + poor contact, 6 power degradation + poor contact) and 55 single faults. The inspection report automatically generated a detailed list with fault location, type, confidence, and recommended actions, providing precise guidance for subsequent review and repair.
3.3 Effectiveness Analysis
Through the application of the intelligent diagnosis and digital twin maintenance system, we achieved remarkable operational and economic benefits. Key indicators before and after deployment are compared below.
| Indicator | Before Deployment | After Deployment | Improvement |
|---|---|---|---|
| Annual power generation efficiency | Baseline (faulty area –18%) | Restored to design level | Significant increase |
| Number of operation staff | 8 persons | 5 persons | −37.5% |
| Full-site inspection cycle | 1 time/month | 1 time/quarter | 3× longer interval |
| Annual operation cost | Baseline | Reduced by 42.3% | Substantial savings |
| Fault occurrence rate | Baseline | Reduced by >85% | Greatly improved reliability |
All key operational metrics improved significantly. The annual power generation efficiency recovered fully, with the previously affected area reaching historical benchmark levels. In terms of maintenance management, the intelligent inspection mode dramatically increased efficiency. While achieving full-site coverage, the inspection cycle extended from once a month to once a quarter, substantially reducing manual labor while ensuring coverage and depth. The operation team downsized from eight to five persons, and the annual operation cost dropped by 42.3%. Moreover, the fault occurrence rate decreased by over 85%, demonstrating a major leap in solar panel station reliability.
4. Conclusion
In this work, we developed and deployed an intelligent fault diagnosis system for solar panel modules in a 10MW distributed photovoltaic station. By combining CNN, LSTM, and Attention mechanism, our model achieves 96.8% accuracy for composite faults and component-level localization, outperforming traditional methods. The integration with a digital twin visualization platform enables a complete closed-loop operation and maintenance system—from perception, diagnosis, to action and review. After deployment, we observed an 83.3% reduction in inspection time, 37.5% reduction in staff, 42.3% cost savings, and over 85% decrease in fault occurrence. The system has proven highly effective for solar panel composite fault diagnosis. Future work will focus on further optimizing the model for adaptability across different solar panel types and environmental conditions, as well as incorporating edge-cloud collaborative intelligence to enhance real-time performance.