Construction and intelligent analysis of power grid physical data knowledge graph based on Internet of Things for power system

A standardised knowledge representation language is implemented. In RDF, knowledge always comes in the form of triples. The subject is an entity, and the predicate is an attribute. An attribute can connect two entities, or connect an entity and its attribute value. If the subject and object of the triple are regarded as the nodes of the graph, the RDF knowledge base can be viewed as a graph model. The Neo4j database can also be shared by multiple departments and users, meeting the needs of interactive infrastructure engineering data between departments.

The information is extracted from the text data of the power grid objects in the power system statistics under the Internet of Things. Triples get stored in a list format. But in the Neo4j database, different parts of the property graph are stored separately in different files, that is, entities and relationships need to be stored in different formats. The process of importing the power grid objects extracted from the information into the Neo4j database to generate a knowledge map library is shown in Figure 2.

First, entity data in text data and table data is read based on Python.

Fig. 2

NER aims to recognise entities of a specified category in text.

Entity extraction for tabular data: For the Excel file extraction entity of the equipment inventory class, it is only necessary to estimate the column where the node is placed and store it separately. By numbering, they can correspond to their respective attribute relationships to form a knowledge map.

Fig. 3

First, the original data obtained in the text are transformed and encoded into a vector form suitable for computer processing before NLP-related algorithms can be applied. Traditional word mapping has the problem of sparse data. Therefore, this study adopts the skip-gram model to optimise the word vector matrix. The model learns an accurate word vector representation for each word. Given any n tuple (w, C) = w_i−n ⋯w_i ⋯ w_i+n, the model directly uses the word vector e(w_i) of the central word to forecast the probability of the tth word w_t in the context: (1) P(wt|wi)=exp(e(wi)⋅e(wt))∑k=1|V|exp(e(wi)⋅e(wk)) P({w_t}|{w_i}) = {{\exp (e({w_i}) \cdot e({w_t}))} \over {\sum\limits_{k = 1}^{|V|} \exp (e({w_i}) \cdot e({w_k}))}}

Here, w_i represents the head word, e(w_i) represents the w_i corresponding d dimensional word vector and C indicates the background window size vocabulary size. The objective function of the model is to optimise the word vector matrix to maximise the log-likelihood of all context words: (2) L*=argmaxL∑wiV∑wi(wi,C)log2P(wt|wi) {L^*} = \arg \mathop {\max }\limits_L \sum\limits_{{w_i}}^V \sum\limits_{{w_i}}^{({w_i},C)} {\log _2}P({w_t}|{w_i})

After model training, the optimised word vector matrix L^* is achieved, which contains a distributed vector representation of all words in the vocabulary V.

Then, given a sequence of Chinese characters X = x₀, x₁,…, x_T, the word vector e_i corresponding to each Chinese character x_i is searched for in the trained word vector table, where d₁ represents the vector dimension. LSTM is regulated by three gates and one storage memory unit. (3) ii=σ(Wie(wi−1)+Uihi−1+Vici−1+bi) {i_i} = \sigma \left({{W_i}e({w_{i - 1}}) + {U_i}{h_{i - 1}} + {V_i}{c_{i - 1}} + {b_i}} \right) (4) fi=σ(Wfe(wi−1)+Ufhi−1+Vfci−1+bf) {f_i} = \sigma \left({{W_f}e({w_{i - 1}}) + {U_f}{h_{i - 1}} + {V_f}{c_{i - 1}} + {b_f}} \right) (5) oi=σ(Woe(wi−1)+Uohi−1+Voci−1+bo) {o_i} = \sigma \left({{W_o}e({w_{i - 1}}) + {U_o}{h_{i - 1}} + {V_o}{c_{i - 1}} + {b_o}} \right) (6) σ(x)=11+e−x \sigma (x) = {1 \over {1 + {e^{ - x}}}} (7) ci˜=tanh(Wce(wi−1)+Uchi−1+bi) \widetilde {{c_i}} = \tanh \left({{W_c}e({w_{i - 1}}) + {U_c}{h_{i - 1}} + {b_i}} \right) (8) ct=fi⊙ct−1+ii⊙ci˜ {c_t} = {f_i} \odot {c_{t - 1}} + {i_i} \odot \widetilde {{c_i}} (9) hi=oi⊙tanh(ci) {h_i} = {o_i} \odot \tanh \left({{c_i}} \right) (10) tanh(x)=1−e−2x1+e−2x \tanh (x) = {{1 - {e^{ - 2x}}} \over {1 + {e^{ - 2x}}}} where i_i, f_i and o_i represent the input gate, forget gate and output gate, respectively; c_i represents a memory unit; σ(x) is the sigmoid activation function; and ⊙ represents the dot product.

At the same time, the forward LSTM obtains the hidden layer representation hi← \overleftarrow {{h_i}} (d₂ represents the number of hidden layer neurons) corresponding to each input text character. In the same way, backward LSTM gets another hidden layer representation hi→ \overrightarrow {{h_i}} . hi→ \overrightarrow {{h_i}} can capture e(i) and contextual information e₀ …e_i−1e_i on the left. hi→ \overrightarrow {{h_i}} can capture e(i) and contextual information on the right. Bi-LSTM concatenates the global features of hi← \overleftarrow {{h_i}} , and hi→ \overrightarrow {{h_i}} centres to get the label sequence Y. The conditional probability P(Y |X) is modelled as follows: (11) P(Y|X)=exp(∑k∑i=1T−1λkfk(yi+1,yi,X,i)) P(Y|X) = \exp \left({\sum\limits_k \sum\limits_{i = 1}^{T - 1} {\lambda _k}{f_k}({y_{i + 1}},{y_i},X,i)} \right)

Finally, the decoding of the model uses the Viterbi algorithm, maintaining two sets of variables, δ_t(y) and φ_t(y). δ_t(y) record the maximum probability corresponding to the path ending with label y up to time t. φ_t(y) record the label corresponding to the time t − 1 of the path δ_t(y): (12) δt(y)=max{δt−1(y')P(y|y')P(xt|y)} {\delta _t}(y) = \max \left\{ {{\delta _{t - 1}}({y^{'}})P(y|{y^{'}})P({x_t}|y)} \right\} (13) φt(t)=argmax{δt−1(y')P(y|y')P(xt|y)} {\varphi _t}(t) = \arg \max \left\{ {{\delta _{t - 1}}({y^{'}})P(y|{y^{'}})P({x_t}|y)} \right\} where P(y|y^′) is the state transition probability and P(x_t|y) is the emission probability. When the T-th character is calculated, the label corresponding to the T-th character can be obtained by using the following equations: (14) yT=argmaxy{δt(y)} {y_T} = \arg \mathop {\max }\limits_y \{ {\delta _t}(y)\} (15) yt=φt+1(yt+1) {y_t} = {\varphi _{t + 1}}({y_{t + 1}})

Knowing the weight vector w of the model, the process of finding the dependency tree with the largest weight corresponding to the sentence is called decoding. Second-order decoding is performed using the descendant and parent–child information (2o-carreras). Assuming that the score of each input text is related to all characters, a set of possible dependencies H is determined for each character, and all combinations are obtained according to the score. Taking the input text data as an example, the dependent syntax structure obtained after the model operation.

The KG and intelligent analysis framework of power grid physical data based on the Internet of Things constructed in this study were pilot tested in a city company. The data are obtained from a new construction of a 110 kV substation. Among them, the unstructured data are the design specification, audit report and equipment test report (Word file) of the project construction. The semi-structured data are the equipment inventory (Excel file) in the completion and acceptance stages of the project. The experimental data obtained after data cleaning and other preprocessing are 4983 Chinese text strings and 32 lists with a length of 187. Information extraction and intelligent analysis are carried out, respectively.

The evaluation standard of intelligent analysis is that natural language is encoded and converted into a word vector by the skip-gram model as the input of the Bi-LSTM-CRF model, and three tasks of Chinese word segmentation, part-of-speech tagging and NER are performed at the same time. Using cross-entropy as the loss function, the operation effect of the model is measured by the accuracy rate, recall rate and F value. Taking Chinese word segmentation as an example, the calculation equation of the indicator is as follows: (16) R=TPTP+FN×100% R = {{TP} \over {TP + FN}} \times 100\% (17) P=TPTP+FP×100% P = {{TP} \over {TP + FP}} \times 100\% (18) F=2PRP+R×100% F = {{2PR} \over {P + R}} \times 100\%

Dependency analysis is performed on text strings, and attribute relationships between entities are extracted. The graph-based dependency syntax accuracy evaluation index selects the labelled attachment score (LAS) that considers the type of dependency (represented by LLAS) and the indicator that does not consider the type of dependency unlabelled attachment score (UAS) (represented by UUAS). The equation is as follows: (19) Uuas=Number of correct words in core nodestotal×100% {U_{uas}} = {{{\rm{Number}}{\kern 1pt} {\rm{of}}{\kern 1pt} {\rm{correct}}{\kern 1pt} {\rm{words}}{\kern 1pt} {\rm{in}}{\kern 1pt} {\rm{core}}{\kern 1pt} {\rm{nodes}}} \over {{\rm{total}}}} \times 100\% (20) Llas=Number of times the core node and relationship are correcttotal×100% {L_{las}} = {{{\rm{Number}}{\kern 1pt} {\rm{of}}{\kern 1pt} {\rm{times}}{\kern 1pt} {\rm{the}}{\kern 1pt} {\rm{core}}{\kern 1pt} {\rm{node}}{\kern 1pt} {\rm{and}}{\kern 1pt} {\rm{relationship}}{\kern 1pt} {\rm{are}}{\kern 1pt} {\rm{correct}}} \over {{\rm{total}}}} \times 100\%

This study compares the BiLSTM-CRF model with the SVM model. A total of 102,794 words are segmented by the model in this study, and 8,825 words are segmented by the SVM model (Table 1).

It can be seen from the results that the performance is significantly better than that of the SVM model, and the reason for the accuracy of NER is slightly lower than that of Chinese word segmentation because the recognition effect of design technical entities needs to be improved.

This study uses the 2o-carreras decoding algorithm to analyse dependencies with traditional first-order decoding (lo). Compared with second-order decoding with descendant information (20-sib), the algorithm has the optimal substructure required by the dynamic programming algorithm, with stronger expressive ability and higher accuracy (Table 2).

It can be observed that the BiLSTM-CRF intelligent analysis model used in this study can achieve more accurate results than the traditional model in the knowledge map of power grid physical data based on the Internet of Things. After model operation, a total of 429 entity nodes and 461 relational edges between entities are extracted from 5 categories of single project name, installation address, design technology, equipment name and equipment purchase price, forming 566 knowledge triples.