Solar energy is a renewable and clean source that has become integral to daily life. Solar photovoltaic power stations are characterized by simple design, short installation cycles, no noise, durability, and easy maintenance. The solar photovoltaic array generates direct current, which is converted into usable grid power via solar inverters. When solar energy is abundant, indoor systems can rely on solar-generated electricity; during insufficient solar power, automatic switching to grid power occurs. In this process, solar inverters play a critical role, making their fault diagnosis essential. Faults in solar inverters often manifest briefly and involve complex circuitry, categorized into soft faults (e.g., parametric degradation like component performance decline) and hard faults (e.g., physical damage to switching elements altering topology). Common diagnostic methods include Self-Organizing feature Map (SOM) and Back Propagation (BP) neural networks. However, BP neural networks suffer from poor generalization, require large datasets, and exhibit slow real-time performance, while SOM networks need extensive training and cannot represent results vectorially. To address these limitations, this study employs a Simulink-based simulation model of a three-phase half-bridge solar inverter, simulating 25 soft fault scenarios by varying capacitor degradation levels. The performance of BP, SOM, and SOM-BP neural networks as diagnostic tools is compared to identify the optimal approach for improving fault diagnosis accuracy in solar inverters.
Neural networks are pivotal in diagnosing faults in solar inverters due to their ability to handle nonlinear patterns. The BP neural network is a multilayer feedforward network that utilizes error backpropagation for training, enhancing accuracy. It consists of an input layer, one or more hidden layers, and an output layer. The BP neural network’s topology involves forward propagation of inputs and backward propagation of errors to adjust weights and thresholds, minimizing the difference between predicted and expected outputs. The algorithm proceeds as follows: network initialization, computation of hidden layer outputs, output layer prediction, error calculation, weight and threshold updates, and iterative refinement until convergence. The hidden layer output for node j is given by:
$$H_j = f\left(\sum_{i=1}^{n} w_{ij} x_i – a_j\right)$$
where f is the activation function, x_i denotes input values, w_{ij} represents weights, and a_j is the threshold for hidden node j. The output layer prediction for node k is:
$$O_k = \sum_{j=1}^{l} H_j w_{jk} – b_k$$
with w_{jk} as weights and b_k as the output layer threshold. The error e_k is computed as:
$$e_k = Y_k – O_k$$
where Y_k is the expected output. Weight updates follow:
$$w_{ij} = w_{ij} + \eta H_j (1 – H_j) x_i \sum_{k=1}^{m} w_{jk} e_k$$
and
$$w_{jk} = w_{jk} + \eta H_j e_k$$
with η as the learning rate. Threshold updates are:
$$a_j = a_j + \eta H_j (1 – H_j) \sum_{k=1}^{m} w_{jk} e_k$$
and
$$b_k = b_k + e_k$$
Despite its utility, BP neural networks have drawbacks: sensitivity to initial values leading to local minima, overfitting or underfitting due to improper architecture, and dependence on representative samples. These issues can hinder the diagnostic reliability for solar inverters.
In contrast, the SOM neural network is an unsupervised, self-organizing model with input and competitive layers, capable of learning input data distribution and topology. Neurons in the competitive layer inhibit surrounding neurons, with the winning neuron adapting to input vectors. The algorithm includes initialization, input vector introduction, distance calculation, weight learning, output computation, and iteration. The distance d_j between input vector X and weight vector W_j for neuron j is:
$$d_j = \| \mathbf{X} – \mathbf{W}_j \| = \sqrt{\sum_{i=1}^{m} (x_i(t) – w_{ij}(t))^2}$$
Weight learning updates the winning neuron and its neighbors:
$$\Delta w_{ij} = w_{ij}(t-1) – w_{ij}(t) = \phi(t) (x_i(t) – w_{ij}(t))$$
where φ(t) is the learning rate that decreases over time. The output o_k is:
$$o_k = f(\min \| \mathbf{X} – \mathbf{W}_j \|)$$
SOM networks face challenges in balancing learning speed and stability, potentially resulting in “dead” neurons that never win competitions, limiting their effectiveness for solar inverter fault diagnosis.
To overcome these limitations, the SOM-BP串联 neural network combines SOM as the primary network for clustering and BP as the secondary network for refinement. This hybrid approach reduces input nodes and training samples while enhancing diagnostic accuracy for solar inverters. The structure includes an input layer, competitive layer, hidden layer, and output layer. Algorithm steps involve: normalizing sample data, training the SOM network for preliminary classification, normalizing SOM outputs, initializing and training the BP network, and forming a classification model for testing. This integration mitigates the need for large datasets in BP and the vector representation issue in SOM, making it ideal for solar inverter applications.
For simulation, a three-phase half-bridge solar inverter model was developed in Simulink. Line voltage signals at the midpoint of the three-phase bridge arms served as fault indicators. Capacitors C1 and C2 were set to five degradation levels (N: 0–10%, A: 10–20%, B: 20–30%, C: 30–40%, D: 40–50%), simulating 25 soft fault modes in solar inverters. The fault types are summarized in the table below:
| Fault Name | C1 Degradation Range (%) | C2 Degradation Range (%) |
|---|---|---|
| NN | 0–10 | 0–10 |
| AN | 10–20 | 0–10 |
| BN | 20–30 | 0–10 |
| CN | 30–40 | 0–10 |
| DN | 40–50 | 0–10 |
| NA | 0–10 | 10–20 |
| NB | 0–10 | 20–30 |
| NC | 0–10 | 30–40 |
| ND | 0–10 | 40–50 |
| AA | 10–20 | 10–20 |
| AB | 10–20 | 20–30 |
| AC | 10–20 | 30–40 |
| AD | 10–20 | 40–50 |
| BA | 20–30 | 10–20 |
| BB | 20–30 | 20–30 |
| BC | 20–30 | 30–40 |
| BD | 20–30 | 40–50 |
| CA | 30–40 | 10–20 |
| CB | 30–40 | 20–30 |
| CC | 30–40 | 30–40 |
| CD | 30–40 | 40–50 |
| DA | 40–50 | 10–20 |
| DB | 40–50 | 20–30 |
| DC | 40–50 | 30–40 |
| DD | 40–50 | 40–50 |
From 1,250 signal groups, 1,100 were used for training and 150 for testing. The neural network configurations were as follows: BP with 5 input nodes, 32 hidden nodes, and 25 output nodes; SOM with 25 competitive neurons in a 5×5 grid; and SOM-BP with 64 competitive neurons (8×8 grid), 32 hidden nodes, and 25 output nodes. The SOM-BP network demonstrated superior performance, with training and testing prediction errors illustrated in the results. The diagnostic accuracy comparison is presented in the table below:
| Network Type | Training Set Accuracy | Test Set Accuracy |
|---|---|---|
| SOM-BP | 0.931 | 0.920 |
| BP | 0.914 | 0.900 |
| SOM | 0.817 | 0.820 |
The results indicate that the SOM-BP串联 neural network achieves higher accuracy in diagnosing soft faults in solar inverters, reducing the need for extensive samples and input nodes while improving reliability. The training set prediction error for SOM-BP showed a convergence to minimal values, with an accuracy of 93.18%, and the test set accuracy reached 92%, outperforming BP and SOM networks. This enhancement is crucial for real-world applications where solar inverters are subject to varying operational conditions and require rapid fault identification.
In conclusion, the SOM-BP串联 neural network offers a robust solution for soft fault diagnosis in solar inverters, addressing the shortcomings of individual BP and SOM networks. By leveraging the clustering capabilities of SOM and the predictive power of BP, this hybrid model enhances diagnostic precision and efficiency. The experimental validation confirms its practical value in maintaining the reliability and performance of solar energy systems, ensuring that solar inverters operate optimally under diverse fault scenarios. Future work could explore adaptive learning rates and integration with real-time monitoring systems to further advance solar inverter diagnostics.