3.2.2. Outcomes

This study considers 37 varied scenarios for events. In order to determine the need for various models (fault analysis), we performed some comparative experiments according to various PMU kinds. In one group, properties of localization/segmentation are sent to the related DML model in order to train, and in the other one, whole features are sent to various machine learning models. Moreover, it is shown in Table 4 that data can be effectively split according to the PMU resources. Splitting the data can enhance the accuracy of classification models as well as reduce data dimensions and enhance training speed and minimize computing sources. The score of the significant features is shown in Figure 4.

**Table 4.** Transfer diverse characteristics to the layout for comparison.

**Figure 4.** Significance features score.

Several corresponding experiments are conducted on various ways of replacing abnormal values in data. Table 5 shows the outcomes. The replacement method is shown in the left column, and the suggested approach is represented by *log*\_*mean*. Zero shows a process to replace NAN and INF with zero values, and mean shows a process to replace with the mean value. The AWV model is utilized as a trial model, and the accuracy is adopted as the assessment metrics, that is, the right column in Table 5.

**Table 5.** Diverse methods to procedure Inf and Nan.


Applying the *log*\_*mean* technique for replacing the unusual amount in the data is intuitively the best approach. According to the outcome, the suggested process in order to process abnormal values has proven successful.

Comparison experiments are also conducted to verify feature selection. First, the significance of the original features is determined, and afterward, they are arranged based on significance. A variety of mixtures of features has been selected for training, and Table 6 shows these outcomes.

The approach was verified practically through a comparative test. The test extracts the test group and training group from 15 multiclass data sets in a 9:1 ratio at random, and afterward, these data sets have been combined into 1 training group. The training group has been transferred to the layout to train and learn. Table 7 presents the outcomes of 15 test sets transferred to the model for practically simulating the efficiency of the model applications. It is apparent that the model's accuracy has decreased. It is because data interaction would occur by increasing the number of data resulting in changing the model, and whenever whole data has been combined, there would unavoidably be abnormal

points and noises. Due to the fact that such noises and anomalies have not been separated in training, the model's indexes alter, and the robustness decreases.


**Table 6.** Assessment of characteristics chosen.

**Table 7.** Layout accuracy on 15 trail sets in the actual simulation.


Firstly, the efficacy of the features created from the feature construction engineering in the model is determined by sorting the significance of features. Model interpretability can be determined by determining the significance of features. Weight, gain, cover, and so on are general indicators of feature significance. In the XGBoost method [30], the number of times a property appears in a tree has been shown by weight, the mean gain of the slot using the property has been represented by the gain, and the mean coverage of the slot using the property is shown by the cover. According to Figure 4, weight calculates feature significance. The abscissa indicates the names of the beat 45 properties, and the ordinate indicates the assessment score. The origin features are shown by the gray part. The features derived from feature construction engineering are represented by the red mark. It is evident that each of the 16-making properties is in the best 45.

The test trains 15 sets of multiclass classification data sets and tests respectively and uses accuracy as an assessment metric. The accuracy of the trail data sent to the layout before and after optimization based on the main 128 properties is shown in Figure 5. The classification accuracy of the trail group on various layouts with default variables is shown in Figure 5a, and the accuracy of the trail group on the layout applying optimized variables is represented in Figure 5b. For a more intuitive visualization of the variation in accuracy after layouts are optimized, Figure 5a and b are combined, and the mean of the accuracy values for whole sets are adopted, i.e., Figure 5c. Figure 5 shows that the SVM layout with default variables has an accuracy of approximately 0.30, but after optimization, it grows to 0.85, which represents a near 200% advancement. Other models have improved significantly in accuracy after optimization as well. The best accuracy of the suggested AWV model is 0.9217.

Table 3 shows that every data set has about 5000 segments of data; therefore, the CNN layout cannot be used. The semantic relationships among features might also be ignored by several neural networks, such as CNN and long-short-term memory (LSTM) layouts. Thus, in several cases, statistical features according to the manual design could positively affect model accuracy as well. Moreover, the tree-based algorithm outperforms KNN and SVM.

The test set had better performance on the model suggested in this study in comparison to the conventional DML and CNN, as shown in Figure 5.

**Figure 5.** *Cont.*

**Figure 5.** Proficiency comparisons of variables through applying 128 properties (**a**); (**b**) precision over 15 data sets through applying optimum variables; (**c**) mean accuracy comparison.
