Data availability 数据可用性
所使用的数据是保密的。
Table 1. Main specifications of the selected PV array.
表 1.所选光伏阵列的主要规格。
Parameter 参数 | Description 说明 |
---|---|
Module technology 模块技术 | Mono-crystalline (mc-Si) 单晶硅(mc-Si) |
PV array nominal power 光伏阵列额定功率 | 3.18 kWp |
Inverter type and size 逆变器类型和尺寸 | IG30 Fronius single-phase, 2.5 kW IG30 Fronius 单相,2.5 千瓦 |
Modules per inverter 每个逆变器的模块 | 30 |
Modules in series (Ns) 串联模块 (Ns) | 15 |
Strings in parallel (Np) 并行字符串 (Np) | 2 |
Tilt - Azimuth 倾斜 - 方位 | 35° − 10° West 西经 35° - 10° |
Table 2. Electrical characteristics of the considered PV module.
表 2.所考虑的光伏组件的电气特性。
Parameter 参数 | Value 价值 |
---|---|
Pmp (W) | 106 |
ISC (A) | 6.54 |
VOC (V) | 21.6 |
Imp (A) | 6.10 |
Vmp (V) | 17.4 |
βVOC (%/°C) | −0.36 -0.36 |
αISC (%/°C) | 0.06 |
Fig. 1. The operating conditions of curves 1, 2, and 3 are interpolated to obtain the operating conditions of Curves 4 and 0.
图 1.对曲线 1、2 和 3 的运行条件进行内插,得到曲线 4 和 0 的运行条件。
Fig. 2. MGWO algorithm flowchart.
图 2.MGWO 算法流程图。
Fig. 3. Flowchart of the proposed fault detection and diagnosis strategy.
图 3.建议的故障检测和诊断策略流程图。
Fig. 4. Failure types considered in the proposed methodology (#1 partial shading, open-circuit fault#2, #3 short-circuit fault, and #4 Line-to-Line fault).
图 4.建议方法中考虑的故障类型(#1 部分遮挡、#2 开路故障、#3 短路故障和 #4 线对线故障)。
Fig. 5. DC output power of the grid-connected PV system within various fault scenarios.
图 5.各种故障情况下并网光伏系统的直流输出功率。
Fig. 6. The general structure of the deployed RF model.
图 6.部署的射频模型的总体结构。
Fig. 7. The grid search algorithm's principle.
图 7 网格搜索算法原理网格搜索算法的原理。
Table 3. Defined classes and their corresponding fault type.
表 3.定义的类别及其对应的故障类型
Phase 阶段 | Class 班级 | Corresponding fault type 相应的故障类型 |
---|---|---|
Detection 检测 | 0 | Healthy 健康 |
1 | Faulty 故障 | |
Diagnosis 诊断 | 0 | #2: Open-circuit fault #2:开路故障 |
1 | #1: Partial Shading #1:局部遮光 | |
3 | #3: Short-circuit fault #3:短路故障 | |
9 | #4: Line-to-line fault #4: 线对线故障 |
Table 4. Details of the detection and diagnosis database construction.
表 4.检测和诊断数据库建设详情。
Phase 阶段 | Class 班级 | Test data set (25 %) 测试数据集(25) | Train data set (75 %) 训练数据集(75) | Total 总计 |
---|---|---|---|---|
Detection 检测 | 0 | 12,145 | 36,433 | 242,890 |
1 | 48,578 | 145,734 | ||
Diagnosis 诊断 | 0 | 12,145 | 36,433 | 194,312 |
1 | 12,144 | 36,434 | ||
3 | 12,145 | 36,433 | ||
9 | 12,144 | 36,434 |
Fig. 8. Predicted I-V curve at STC (Curve 0) using the current–voltage translation method.
图 8.使用电流-电压转换法预测的 STC(曲线 0)I-V 曲线。
Table 5. Extracted ODM parameters at STC.
表 5.在 STC 提取的 ODM 参数。
Parameter 参数 | Value 价值 |
---|---|
Rp (Ω) | 42.9633 |
RS(Ω) | 0.2212 |
Io (A) | 4.344 10-7 |
n | 45.1606 |
Iph (A) | 6.8378 |
RMSE | 0.0122 |
Fig. 9. PV array model validation under a) T = 28.1, G = 749, b) T = 28.2, G = 800.
图 9.a) T = 28.1,G = 749,b) T = 28.2,G = 800 下的光伏阵列模型验证。
Fig. 10. Dynamic validation of the PV array model under different weather conditions.
图 10.光伏阵列模型在不同天气条件下的动态验证。
Table 6. Optimal hyperparameters.
表 6.最佳超参数。
Hyperparameter 超参数 | RF Detection model 射频检测模型 | RF Diagnosis model 射频诊断模式 |
---|---|---|
max_depth 最大深度 | 45 | 85 |
n_estimators | 65 | 35 |
Criterion 标准 | gini 基尼 | entropy 熵 |
Bootstrap | True 正确 | True 正确 |
Min_samples_leaf 最小样本叶片 | 1 | 1 |
Min_sample_split 最小样本分割 | 2 | 2 |
Max_features 最大特征 | 6 | 6 |
Table 7. Classification report of RF detection model.
表 7.射频检测模型的分类报告。
Empty Cell | Precision 精度 | Recall 回顾 | F1score | Samples number 样本数量 |
---|---|---|---|---|
Class0 0 级 | 1.00 | 0.970 | 0.985 | 12,145 |
Class1 1 级 | 0.993 | 1.000 | 0.996 | 48,578 |
Macro avg 宏观平均值 | 0.996 | 0.985 | 0.991 | 60,723 |
Weighted avg 加权平均数 | 0.994 | 0.994 | 0.994 | 60,723 |
Accuracy (%) 准确度(%) | 99.4 | 60,723 |
Table 8. Classification report of RF diagnosis model.
表 8.射频诊断模型的分类报告。
Empty Cell | Precision 精度 | Recall 回顾 | F1score | Samples number 样本数量 |
---|---|---|---|---|
Class0 0 级 | 0.978 | 1.000 | 0.989 | 12,145 |
Class1 1 级 | 1.000 | 0.974 | 0.987 | 12,144 |
Class3 3级 | 0.999 | 1.000 | 1.000 | 12,145 |
Class9 9 级 | 0.998 | 1.000 | 0.999 | 12,144 |
Macro avg 宏观平均值 | 0.994 | 0.994 | 0.994 | 48,578 |
Weighted avg 加权平均数 | 0.994 | 0.994 | 0.994 | 48,578 |
Accuracy (%) 准确度(%) | 99.4 | 48,578 |
Fig. 11. Normalized Confusion matrix of RF detection model.
图 11.射频检测模型的归一化混淆矩阵。
Fig. 12. Normalized Confusion matrix of RF diagnosis model.
图 12.射频诊断模型的归一化混淆矩阵。
Fig. 13. Fault detection results.
图 13.故障检测结果。
Fig. 14. Fault diagnosis results.
图 14.故障诊断结果。
Table 9. Comparative Analysis between SVM, KNN, DT, SGDC, MLP, and RF trained and tested using the same data set.
表 9.使用相同数据集训练和测试 SVM、KNN、DT、SGDC、MLP 和 RF 的比较分析。
Phase 阶段 | Indicator 指标 | label 标签 | SVM | MLP Classifier MLP 分类器 | KNN | DT | SGDC | RF |
---|---|---|---|---|---|---|---|---|
Detection 检测 | Precision 精度 | 0 | 0.979 | 0.913 | 0.979 | 0.988 | 0.000 | 1.000 |
1 | 0.984 | 0.990 | 0.984 | 0.981 | 0.796 | 0.993 | ||
Recall 回顾 | 0 | 0.939 | 0.960 | 0.939 | 0.927 | 0.000 | 0.970 | |
1 | 0.995 | 0.977 | 0.995 | 0.997 | 1.000 | 1.000 | ||
F1score | 0 | 0.959 | 0.936 | 0.959 | 0.956 | 0.000 | 0.985 | |
1 | 0.990 | 0.983 | 0.990 | 0.989 | 0.886 | 0.996 | ||
Accuracy (%) 准确度(%) | 84.5 | 97.3 | 98.3 | 98.3 | 79.6 | 99.4 | ||
Diagnosis 诊断 | Precision 精度 | 0 | 0.958 | 0.992 | 0.923 | 0.992 | 0.851 | 0.978 |
1 | 1.000 | 1.000 | 0.996 | 0.997 | 0.876 | 1.000 | ||
3 | 0.933 | 0.974 | 0.998 | 0.972 | 0.986 | 0.999 | ||
9 | 0.964 | 0.964 | 0.997 | 0.971 | 0.908 | 0.998 | ||
Recall 回顾 | 0 | 0.928 | 0.975 | 0.997 | 0.972 | 0.967 | 1.000 | |
1 | 0.951 | 0.972 | 0.905 | 0.975 | 0.930 | 0.974 | ||
3 | 0.968 | 0.981 | 0.997 | 0.985 | 0.697 | 1.000 | ||
9 | 1.000 | 1.000 | 1.000 | 0.997 | 1.000 | 1.000 | ||
F1score | 0 | 0.943 | 0.983 | 0.959 | 0.982 | 0.905 | 0.989 | |
1 | 0.975 | 0.986 | 0.948 | 0.986 | 0.902 | 0.987 | ||
3 | 0.951 | 0.978 | 0.997 | 0.979 | 0.816 | 1.000 | ||
9 | 0.981 | 0.982 | 0.998 | 0.984 | 0.952 | 0.999 | ||
Accuracy (%) 准确度(%) | 96.1 | 98.2 | 97.5 | 98.3 | 89.8 | 99.4 |
Although the presented models in (Dhimish & Tyrrell, 2023) and (Badr et al., 2021) have considered both a small dataset and a number of critical conditions in fault detection, the downside is the models insufficient accuracies which noticeably reduces the models reliability. Moreover, authors in (Amiri, Oudira, et al., 2024) have presented a seemingly accurate model which considers a number of critical conditions in fault detection and is trained using a relatively small dataset. However, the model is incomprehensive since the capability of fault detection is only limited to a few kinds of faults in PV arrays.
In this context, deep learning models, as a significant branch of artificial intelligence, have gained widespread exploration and application in the specific scenario of PV panel anomaly detection due to their powerful capabilities in feature extraction and pattern recognition [17–19]. These models can accurately capture anomalous patterns during PV panel operation, enabling efficient and precise fault prediction and localization [4,20–25]. Li et al. [26] utilized aerial images obtained from drones for defect pattern recognition.