In recent years, important research progress has been made on power grid fault diagnoses based on artificial neural networks. Bi et al. [22] conducted a series of explorations on the applications of traditional neural networks; here, they mainly explored radial basis function (RBF) networks for processing the status information of protection devices. Cardoso et al. [23] developed an improved RBF model coupled with a fuzzy network; dos Santos Fonseca et al. [24] successfully applied reset neural networks and multilayer perceptrons to model the protection action principle; Zou et al. [25] reported the conversion of protection action information to percentage inputs. With respect to deep learning, Wen et al. [26] reported a significant breakthrough by pioneering the use of convolutional neural networks (CNNs) to achieve fault-phase selection and discrimination between the inside and outside areas. Yang et al. [27] used signal-to-image conversion along with CNNs to greatly improve the feature extraction ability. The SR-CNN model proposed by Zhang et al. [28] effectively solved the processing problem of phasor measurement unit (PMU) time-series data. The P-CNN model developed by Alqudah et al. [29] showed promising performance in the positioning of DC transmission lines. Other researchers [30] have combined k-means clustering with recurrent neural networks (RNNs) to achieve accurate classification of fault types. The above studies provide new technical paths for intelligent power grid fault diagnoses through diverse data input methods like status signals, electrical quantities, and images as well as innovative network structure designs (parallel, coupled, etc.).

Although various scholars have achieved fruitful results with regard to power grid fault diagnoses, the existing methods are still constrained by the following main problems:

1. Rule-based fault diagnosis models rely on logical rules; thus, when constructing fault events with more complex logical relationships, the logical constraints are cumbersome, knowledge representation is difficult, and model complexity is high. In addition, large amounts of alarm information need to be screened and classified during the modeling process, which increases the modeling workload.

In response to the intelligent needs of power grid fault diagnoses, we propose a fault diagnosis method based on deep reinforcement learning. This method directly processes the alarm information text to significantly improve the intelligent diagnosis level and operational efficiency of the diagnostic system.

In the proposed fault diagnosis framework for power grid systems, the alarm text information is represented using a word vector matrix approach integrated with the DPCNN. This approach converts each alarm statement into a word vector matrix A ∈ ℝ^{M × N}, where M is the number of words and N is the dimension of each word vector, as detailed in Equation 1. Using pretrained word embeddings, such as word2vec, this approach captures the semantic relationships within alarm texts, enabling the DPCNN to extract the implicit fault features critical for identifying fault types like simple faults, switch failures, and protection failures. Then, sentence vectors are formed by concatenating the word vectors while preserving the temporal context and word order, which are essential for processing sequential alarm data from SCADA systems, as shown in the example alarm sequences in Supplementary Appendix Table S1.

Several alternative text representation methods have been applied in fault diagnosis and related natural language processing tasks. The term frequency inverse document frequency (TF-IDF) method represents text parts as sparse vectors based on the word frequency and corpus rarity, offering computational simplicity but failing to capture the semantic relationships or temporal dependencies, which limits its effectiveness for complex alarm texts in power grid applications. The bag-of-words (BoW) model, which counts word occurrences without regard to order or context, is similarly constrained and unsuitable for distinguishing nuanced fault types where sequences matter. Recent advances like the bidirectional encoder representations from transformers (BERT) method provide contextualized embeddings by modeling the bidirectional sentence context to achieve superior semantic understanding. However, the high computational demands of BERT and its need for large labeled datasets make it less practical for real-time fault diagnoses in resource-constrained environments, such as power grid control centers.

The word vector matrix approach with DPCNN offers distinct advantages over the aforementioned alternatives. By leveraging pretrained embeddings, it captures semantic nuances and enables robust feature extraction for diverse fault scenarios. Compared to BERT, its computational efficiency ensures suitability for real-time application, which is critical for rapid fault diagnosis in a power grid. Additionally, preserving the temporal context allows effective analysis of the sequential alarm data, unlike TF-IDF and BoW, which overlook the word order. However, the proposed approach relies on the quality of pretrained embeddings, which may not fully address domain-specific terminology in power grid systems; long alarm texts may also challenge the ability of the DPCNN to model extended dependencies. Thus, future research efforts could explore hybrid models combining the efficiency of DPCNN with the contextual depth of BERT to enhance domain-specific text representation, provided the computational constraints are mitigated.

In our proposed framework for power grid fault diagnosis, the DPCNN is employed to extract features from alarm text information. The DPCNN leverages a pyramid structure with residual connections to capture both local and long-range dependencies in the alarm texts, enabling robust extraction of implicit fault features, as represented by the word vector matrix in Equation 1. This approach achieves high classification accuracy and fault recognition range, as evidenced by the experimental results presented in Section 3 and the alarm sequence analysis given in Supplementary Appendix Table S1. The ability of the DPCNN to model sequential data from SCADA systems makes it particularly suitable for identifying complex fault types, such as switch failures and protection malfunctions, in real-time power grid applications.

Several alternative models have been explored for feature extraction in fault diagnosis models and related efforts. The long short-term memory (LSTM) network is a type of RNN designed to handle sequential data and has been applied to power grid fault diagnosis for its ability to capture temporal dependencies. However, LSTM suffers from high computational complexity and vanishing potential gradients, which can hinder performance with long alarm texts. BERT offers superior feature extraction by capturing the bidirectional context, which makes it effective for complex text analyses in fault diagnosis scenarios. Despite their high accuracies, BERT-based models require substantial computational resources and large labeled datasets, rendering them less feasible for real-time applications in resource-constrained power grid environments. Graph neural networks (GNNs) have emerged as a promising approach for modeling power grid topologies in terms of graphs by integrating alarm text with structural information. Although GNNs excel in topology-aware fault diagnoses, their focus on the graph-based features makes them less optimized for pure text feature extraction compared to DPCNN.

The DPCNN used herein outperforms the above alternatives in several key aspects; its pyramid structure enables efficient capture of both local and global text features to achieve balance between high accuracy and computational efficiency, as demonstrated in Section 3. Unlike LSTM, the DPCNN is not prone to gradient issues and processes lengthy sequences more effectively, making it suitable for real-time fault diagnosis. Compared to BERT, the DPCNN requires fewer computational resources, aligning with the practical constraints at power grid control centers. Although GNNs are advantageous for topology-based analyses, the focus of DPCNN on text feature extraction ensures superior performance for alarm text processing. However, DPCNN relies on pretrained embeddings, which may limit its ability to capture highly domain-specific fault patterns; further, its performance on extremely lengthy texts could be constrained by its receptive field. Future works could therefore explore integrating DPCNN with GNNs to combine text-based and topology-based features, potentially enhancing fault localization in complex power grid scenarios.

In the aftermath of a power grid failure, the protection and circuit breaker operation information related to the faulty equipment will be uploaded to the SCADA system in real time. Meanwhile, the SCADA system will receive additional alarm information, such as device failure and uncharged spring of the switch mechanism. These details are recorded in chronological order to generate the alarm information set at the time of the fault. Table 1 shows an example of the alarm information in a certain area. To facilitate computer processing, it is necessary to represent the text of the alarm information set in a numerical manner. The available text representation methods are mainly divided into discrete representation methods based on one-hot, BoW, and TF-IDF models or distributed representation methods based on n-gram, word2vec, and GloVe models. Among these, the word2vec model can map words from a high-dimensional space to a low-dimensional space while retaining the associations between words, thereby solving the problems of vector sparsity and loose semantic connections in the text representation process effectively. Therefore, we use the word2vec model to vectorize the alarm text information. The vectorization process entails the following steps:

1. Remove the time column in the alarm text information.

In the above equation, the symbol “⊕” represents vector concatenation, x_i is the sentence vector of the ith alarm message, and x_1:n is the concatenation of the sentence vectors of the first to nth alarm messages corresponding to the sentence vector matrix in Figure 1. x i = x i 1 ⊕ x i 2 ⊕ ⋯ ⊕ x i j ⊕ ⋯ ⊕ x i k (2)

In Equation 2, k is the dimension size of the sentence vector, and x_i ^j is the j-dimensional component of the sentence vector of the ith alarm information. For a single alarm message statement, steps (4) and (5) are omitted and the word vectors in the statement are directly spliced to form the word vector matrix, as shown in Figure 2.

The DPCNN constitutes a word-level-deep text classification network in which the convolutional and pooling layers are continuously deepened based on the CNN. The structure of the DPCNN model is shown in Figure 3 and is mainly composed of a text region embedding layer, two convolutional blocks, two pooling layers, and a fully connected layer. Among these, the convolutional blocks include residual modules to avoid the gradient vanishing problem caused by deepening of the network layers. Once the output from the region embedding layer passes through the operations of the convolutional blocks, the sequence length is halved and perceived text segments are doubled. If the model contains multiple pyramid convolutional modules, the alternating operations of the convolutional block and 1/2 pooling layer reduce the length of the text sequence exponentially, resulting in a “pyramid” shape as the number of modules increases. The pyramid convolutional module of the DPCNN model enables perception of long-distance text information and enhances the ability to extract global text features.

Given the global feature extraction ability of the DPCNN model for long-distance texts, we selected the DPCNN to extract features from the alarm text information. Given the alarm text A ∈ R^{n × k}, where n is the text sequence length and k is the sentence vector dimension, the regional features of the text are first extracted through the region embedding layer. Then, the convolutional operations are performed successively through the pyramid convolution module and convolution blocks, and the feature vector outputs of the convolutional blocks are max-pooled by the final pooling layer to obtain the global features of the alarm text. A convolution kernel parameterized by a weight matrix w ∈ R^{h × k} is used to perform the convolution operations in the direction of the text sequence length, and the convolution process is as shown in Equation 3: c i = f w · x i : j + h − 1 + b (3)

In the above equation, x i : j + h − 1 refers to the h sentence vector matrices contained in the convolution kernel window, b ∈ R is the bias vector, f is the non-linear activation function, and c_i is the feature value after convolution. A text of length n contains (n - h + 1) windows x 1 : h , x 2 : h + 1 , ⋯ , x n − h + 1 : n , and the convolution operation is performed on each window to generate the corresponding features. All the features are then concatenated to obtain the feature vector of the alarm information text, as shown in Equation 4: c r e g i o n = c 1 , c 2 , ⋯ , c n − h + 1 (4)

The above feature vector is input to the convolutional block for equal-length convolution and identity mapping, and the output feature vector is given as follows: c c o n v = c 1 ′ , c 2 ′ , ⋯ , c n − h + ! ′ (5)

In Equation 5, the dimension of c c o n v is the same as that of c r e g i o n .

The output from the first convolutional block is passed through the 1/2 pooling layer for downsampling before being input to the next convolutional block. The final pooling layer performs max pooling on the feature vector output from the convolutional block to produce the global features. Figure 4 shows the feature vector diagrams extracted by each of the convolutional and pooling layers in the pyramid convolutional module once an alarm text sequence of length 20 (seg_length = 20) is input to the DPCNN model (kernel_size = 3). The different colored areas represent feature values of different sizes; larger feature values are closer to the color white and have greater impacts on the text classification. It can be seen from the figure that the existence of the pyramid convolutional module changes the length of the feature vector in a pyramid shape.

We designed two fault classification models based on DPCNN for the alarm information set, as shown in Figure 5. Classification model 1 can be used to distinguish between non-fault, simple fault, complete protection operation failure, incomplete protection operation failure, switch operation failure, and developing fault conditions. The definitions of these fault types are as follows:

1. Simple fault: There is only one faulty equipment in the fault event, and both the protection and circuit breaker operate correctly.

Classification model 2 can be used to distinguish line, busbar, and transformer faults. The structures of the two classification models are identical, with only the inputs and outputs being different. The input layer of each model is the vector matrix A ∈ R n × k corresponding to the set of alarm information to be classified. The probability of each classification label is calculated by the fully connected layer, and the label with the highest probability is the considered as the fault classification result corresponding to the alarm text information.

We also designed a key information extraction model based on the DPCNN for a single alarm message in the alarm information set to extract key details like the alarm information of the protection and circuit breaker actions. The input layer of this model is shown in Figure 6 and is the word vector matrix A ∈ R n × k corresponding to the alarm information statement to be classified; here, m is the number of words in the statement, and k is the dimension of the word vector. The model output is the classification result of the alarm information statement, i.e., whether it contains key or non-key information. The alarm information set is input to the key information extraction model, and the classification result of each alarm information statement is output. If the alarm information statement is considered key information, then the equipment involved in this statement are defined as suspicious faulty equipment. The equipment involved in all the key information in the alarm information set are represented as the set of suspicious faulty equipment D. D = d 1 , d 2 , ⋯ , d i , ⋯ , d n (6)

In Equation 6, d i is the device involved in the ith key information in the alarm information set, and n is the number of suspicious faulty devices. To improve the accuracy of identifying faulty devices, the range of the set of suspicious faulty devices is further narrowed according to the fault classification results of suspected faulty transformers. The working mechanism of the key information extraction model for identifying “key information” from alarm texts is based on the DPCNN that extracts semantic features from alarm logs using a series of convolutional layers (filter size = 3) and max-pooling operations, followed by classification of text segments into “key” or “non-key” based on their relevance to fault diagnosis (e.g., equipment identifiers and fault timestamps) through the fully connected layer.

Once the set of suspicious faulty equipment is determined using the key information extraction model, it is necessary to identify the actual faulty equipment. In this work, we use an integrated fault-type and temporal-sequence-prioritized faulty equipment identification strategy to determine the temporal distribution characteristics of the alarm information from different fault types and obtain the final set of suspicious faulty devices D ′ . D ′ = d 1 , d 2 , ⋯ , d m (7)

In Equation 7, m ≤ n. Based on the classification results of model 2, D′ may be the set of suspected faulty lines, set of suspected faulty buses, or set of suspected faulty transformers. The faulty equipment identification strategy used in this work is shown in Figure 7.

1. When the classification result of model 1 is a simple fault, the equipment in the set D are identified as the faulty equipment.

The power grid fault diagnosis framework based on DPCNN is shown in Figure 8 and consists of two parts: model training and testing as well as online diagnosis. The steps of the model training and testing process are as follows:

1. Obtain the alarm information sets of historical fault events collected by the SCADA system as the sample set, and use the word2vec model to vectorize the alarm information sample set and single alarm information separately.

The trained models are used for online fault diagnosis, whose steps are as follows:

1. A sliding time window is established with a time step of 5 s and window size of 10 s to intercept the real-time alarm information received by the SCADA system and vectorize it as described above.

To demonstrate the applicability of the proposed models to alarm information having different description methods, we selected an actual power grid in a certain area as well as the TS2000 simulation system used to simulate the D5000 smart grid dispatching technology support system as the data sources to obtain samples. Here, the actual power grid contains 731 plants and substations, 5,382 lines, and 13 voltage levels ranging from 0.22 kV to 500 kV. The TS2000 simulation system contains 60 plants and substations, 825 lines, and 5 voltage levels ranging from 10 kV to 500 kV. The alarm information samples generated by these two power grids have the following characteristics. The alarm information generated by the actual power grid is the aggregate-level information and contains more non-fault information; this information has the characteristics of high noise and simplistic descriptions, as shown in Supplementary Appendix Table S1. The alarm information generated by the TS2000 simulation system has more detailed fault descriptions and less non-fault information; it the characteristics of low noise and high descriptiveness, as shown in Supplementary Appendix Table S2.

Based on the two fault classification models proposed previously, two alarm information sample sets and two alarm information statement sets each are obtained from the historical monitoring alarm information of the power grid in a certain area for 2019 and the TS2000 simulation system; these data are then divided into training and test sets in the ratio of 7:3. The compositions of the two alarm information sample sets are shown in Tables 2, 3. The classification performances of the two models are evaluated using the metrics shown in Equation 8–10. p r e c i s i o n = n T P n T P + n F P (8) r e c a l l = n T P n T P + n F N (9) F 1 = 2 · p r e c i s i o n · r e c a l l p r e c i s i o n + r e c a l l (10)

In these equations, n T P , n F N , n F P , and n T N respectively represent the numbers of samples of true positive (TP), false negative (FN), false positive (FP), and true negative (TN) event types, and these event relationships are shown in Table 4.

In this work, Python programming language was used to build the DPCNN model for fault diagnosis based on the deep-learning framework TensorFlow in the PyCharm development environment. To ensure that the model achieved the best classification performance and based on the samples shown in Table 3, the main parameters like convolution kernel size (kernel_size), number of convolution kernels (num_filters), and sequence length (seq_length) input to the model were changed to analyze the influences of different parameter values on the F1 value of the model. Accordingly, the convolution kernel size, number of convolution kernels, and input sequence length of the fault classification model were set to 3, 200, and 80, respectively, while the corresponding parameters of the key information extraction model were set to 3, 150, and 6. The other parameters of the models were set according to their performances and computing speeds, as shown in Table 5.