In recent years, the escalating depletion of fossil fuels has intensified global environmental pollution and energy consumption, leading to a growing crisis in both energy supply and ecological sustainability. Consequently, the development and application of renewable energy systems have garnered significant attention from researchers worldwide. Solar power systems, as a critical component of the新能源 industry, are increasingly integrated into power grids, with solar inverters playing a pivotal role in ensuring stable, safe, and efficient operation by converting and managing energy. As technological advancements drive the expansion of solar system capacities, the demand for higher efficiency and lower costs in solar inverters has become more pronounced. Among various inverter topologies, the three-level solar inverter stands out due to its superior performance compared to traditional two-level inverters. It exhibits reduced harmonic distortion, enhanced power handling capabilities due to increased voltage levels, lower voltage stress on internal switching devices, and higher operational efficiency, making it highly suitable for photovoltaic applications. However, the increased number of switching components in three-level solar inverters compromises circuit reliability, as any single device failure can halt operation and potentially cause secondary faults, leading to significant losses in the power system. Thus, research on fault diagnosis for three-level solar inverters is crucial for maintaining system stability.
Current studies on three-level solar inverters primarily focus on three aspects: fault modes involving single-device open-circuit failures, diagnostic methods for multi-device faults, and intelligent algorithms for fault detection. Traditional approaches often rely on output voltage and inductor current as detection signals, but the slow variation of inductor currents can prolong diagnosis times due to the need for threshold adjustments. Intelligent diagnostic techniques, such as artificial neural networks (ANNs) and fuzzy logic, are gaining traction; however, methods like backpropagation (BP) neural networks face limitations in parameter tuning, convergence speed, and local optima. In contrast, extreme learning machines (ELMs) offer improved generalization and faster training by determining hidden layer parameters autonomously and solving linear equations, though their application in solar inverter fault diagnosis remains underexplored. This paper addresses these gaps by proposing a rapid diagnostic method for three-level solar inverters based on wavelet packet energy spectrum and ELM, focusing on bridge arm voltage and power device voltages as detection signals to enhance anti-interference and accuracy.
The selection of detection signals is paramount in fault diagnosis for solar inverters. In operational states, current signals are commonly used but are influenced by circuit conditions, such as load variations, which can delay fault detection. For three-level solar inverters, voltage-based signals—specifically bridge arm voltage, upper switch voltage, and lower switch voltage—are more effective as they are less affected by load changes, thereby accelerating fault identification. In this study, we concentrate on the neutral-point-clamped (NPC) topology of three-level solar inverters, leveraging its symmetry to analyze fault modes in phase A as a representative case. The typical fault modes include single-device open-circuit faults (e.g., f1, f2, f5) and multi-device faults (e.g., f7, f8, f9, f10), each altering the operational states of the solar inverter. For instance, faults like f2 and f7 exhibit similar bridge arm voltage waveforms, necessitating additional signals like upper switch voltage to distinguish between them accurately. The detection signals and their corresponding fault modes are summarized in Table 1.
| Detection Signals | Fault Modes |
|---|---|
| Vao | f1, (f2 or f7), (f3 or f12), f4–f6, f8–f11 (10 types) |
| Vao, VSa1 | f1, f2, (f3 or f12), f4, f5–f11 (11 types) |
| Vao, VSa4 | f1, (f2 or f7), f3–f6, f8–f12 (11 types) |
| Vao, VSa1, VSa4 | f1–f12 (12 types) |
Fault feature extraction is performed using wavelet packet energy spectrum analysis, which decomposes voltage signals into multiple frequency bands to capture distinctive energy patterns. The wavelet packet decomposition involves splitting a signal node into two sub-nodes iteratively, reconstructing the signal across 2^j frequency bands for a j-level decomposition. The energy in each band i is computed as the sum of squared coefficients, forming an energy vector that characterizes the fault. For a solar inverter, the energy spectrum feature vector is derived from three voltage signals, resulting in a 3×2^j-dimensional vector. To reduce dimensionality and enhance computational efficiency, principal component analysis (PCA) is applied, selecting the most significant components to form a compact fault feature vector. The energy calculation for each frequency band i is given by:
$$E_i = \sum_{k=1}^{N} |d_{j,k}^m|^2, \quad i = 1, 2, \ldots, 2^j$$
where \(E = [E_1, E_2, \ldots, E_{2^j}]\) represents the energy vector, and \(d_{j,k}^m\) denotes the wavelet coefficients. The combined energy spectrum feature vector for the solar inverter is expressed as:
$$E_p = [E_{q,1}, \ldots, E_{q,2^j}, E_{s,1}, \ldots, E_{s,2^j}, E_{x,1}, \ldots, E_{x,2^j}]$$
PCA transforms this vector into a lower-dimensional space by identifying principal components that explain the maximum variance, as illustrated in Table 2 for a 24-dimensional energy spectrum. For example, the first principal component may reflect features from Es1 and Ex1, enabling efficient fault representation.
| Principal Component | Comprehensive Reflection Indicators |
|---|---|
| 1 | Es1, Ex1 |
| 2 | Eq1, Ex2 |
| 3 | Eq2, Es3 |
| 4 | Es7, Ex3 |
| 5 | Es6, Ex6 |
| 6 | Eq3, Eq6 |
| 7 | Eq3, Eq6 |
| 8 | Eq4, Eq6, Eq8 |
| 9 | Eq5, Ex4 |
| 10 | Es2, Es8, Ex6 |
| 11 | Ex5, Ex7 |
| 12 | Ex4, Ex5, Ex7 |
| 13 | Eq8, Ex8 |
The fault diagnosis method for solar inverters involves a step-by-step process: (1) Establishing fault conditions and sampling signals from measurement points in the solar inverter. (2) Selecting a wavelet function (e.g., db3) and determining the decomposition level to perform wavelet packet decomposition, obtaining energy spectrum feature vector samples for each fault mode. (3) Applying PCA to reduce the dimensionality of these vectors, extracting salient fault features. (4) Training an ELM model using the reduced-dimensional samples. (5) Integrating measured error features into the trained ELM for real-time fault detection. The ELM, as a single-hidden-layer feedforward neural network, autonomously sets hidden layer parameters and computes output weights via least squares solution, enabling rapid training and high generalization. The output function of the ELM for a solar inverter fault diagnosis can be represented as:
$$f(\mathbf{x}) = \sum_{i=1}^{L} \beta_i g(\mathbf{w}_i \cdot \mathbf{x} + b_i)$$
where \(g(\cdot)\) is the activation function, \(\mathbf{w}_i\) and \(b_i\) are the input weights and biases, \(\beta_i\) are the output weights, and L is the number of hidden nodes. This approach significantly reduces computational complexity compared to iterative methods like BP neural networks.
To validate the proposed method, simulations were conducted on an NPC three-level solar inverter under 13 different operational modes, with an output voltage of 200 V and a 20-cycle sampling period. A total of 650 electrical signal parameters were collected from various measurement points under nominal inductive load conditions, with a 10% load variation to test robustness. Using db3 wavelet packet decomposition with j=3 (resulting in 8 frequency bands), a 24-dimensional energy feature vector was generated for each sample. PCA was then applied to reduce this to 13 principal components, forming the fault feature vector. For training, 390 error vector samples were used, and the ELM was configured with hidden layer nodes determined through cross-validation. The performance was compared against traditional BP neural networks and least squares support vector machines (LSSVM), as summarized in Table 3. The ELM-based method achieved a diagnosis accuracy of 97.1%, with training and testing times of 0.002881 s and 0.001414 s, respectively, outperforming both BPNN (85.7% accuracy, 6.236422 s training) and LSSVM (92.3% accuracy, 0.033916 s training). This demonstrates the superior efficiency and precision of the ELM for online fault diagnosis in solar inverters, as it requires only a single parameter input and linear equation solution.
| Diagnostic Method | Diagnosis Rate (%) | Training Time (s) | Testing Time (s) |
|---|---|---|---|
| ELM | 97.1 | 0.002881 | 0.001414 |
| BPNN | 85.7 | 6.236422 | 0.139401 |
| LSSVM | 92.3 | 0.033916 | 0.135021 |
The wavelet packet energy spectrum effectively captures fault characteristics in solar inverters, with energy distribution patterns varying significantly across fault types. For instance, under single-device faults like f1 or f2, specific frequency bands exhibit heightened energy levels, whereas multi-device faults such as f7 or f8 show distinct energy shifts. The PCA reduction ensures that these patterns are retained in a compact form, facilitating rapid ELM processing. The overall diagnosis process for solar inverters can be summarized mathematically as: given a voltage signal \(\mathbf{v}(t)\) from the solar inverter, the wavelet packet transform yields coefficients \(\mathbf{d}_{j,k}\), from which the energy vector \(\mathbf{E}\) is computed. After PCA projection to a feature vector \(\mathbf{F}\), the ELM model outputs a fault class \(\hat{y}\) using:
$$\hat{y} = \text{ELM}(\mathbf{F})$$
This streamlined approach enables the diagnosis of multiple fault types in solar inverters within seconds, making it highly suitable for real-time applications.
In conclusion, the increasing reliance on clean energy systems underscores the importance of reliable solar inverters in power grids. The proposed method, combining wavelet packet energy spectrum and ELM, offers a robust solution for multi-fault diagnosis in three-level solar inverters. By leveraging voltage-based detection signals and PCA for feature reduction, it overcomes the limitations of traditional methods, such as slow response and low accuracy. Simulation results on NPC solar inverters confirm that this approach achieves near 98% diagnosis accuracy within 3 seconds, significantly reducing fault diagnosis costs and enhancing system reliability. Future work could explore adaptive wavelet functions and deep learning extensions to further improve the performance of solar inverter fault diagnosis systems.