Power grid data standardization reconstruction method, system, device and storage medium

Abstract

The application discloses a power grid data standardization reconstruction method, system, device and storage medium. The method comprises the following steps: acquiring power grid multi-source heterogeneous data, wherein the multi-source heterogeneous data comprises structured data, semi-structured data and unstructured data; performing feature extraction on the structured data, the semi-structured data and the unstructured data respectively to obtain attribute initial semantic features; dividing the attribute initial semantic features into an attribute static feature set and a dynamic feature set with time sequence relationship from the angles of static and time sequence dynamic features; based on the dynamic feature set, performing fusion and extraction on the attribute time sequence feature set in the time dimension; and cascading the attribute static feature set and the attribute time sequence feature set to obtain initial standardization reconstruction features.

Description

Technical Field

[0001] This invention relates to the field of power distribution network big data mining technology, specifically to a method, system, equipment, and storage medium for standardized reconstruction of power distribution network data. Background Technology

[0002] With socio-economic development and the improvement of people's living standards, the main contradiction in the current electricity market has shifted to the conflict between users' demands for high power supply capacity and quality and the reality of low power supply capacity and service quality. The distribution network is a crucial component of the power system, and its planning and real-time monitoring directly impact the system's service capacity and quality for power supply to users.

[0003] With the increasing scale of power distribution networks and the significant growth in the number of users, higher demands are placed on the safe and reliable operation and use of these networks. Traditional methods relying solely on the large-scale power data collected by the power grid are no longer sufficient to meet the needs for fine-grained and real-time accurate fault diagnosis and analysis. Simultaneously, with the widespread adoption of the internet, which has reached every household and driven the development of various industries, user network data represents a completely new and untapped area in power distribution network planning. Therefore, the main challenge facing power grid companies is to overcome development bottlenecks by integrating user-side network equipment data into their transformation efforts.

[0004] This raises another issue: the sources of traditional power big data are becoming increasingly diverse. The construction and transformation of power grids involve various data formats and multiple data sources. Now, data information from network-side devices is being introduced. Although traditional power big data research already had a standardized format for power data, the data from network-side devices that need to be introduced now involves different formats from PMS systems, electricity consumption information collection systems, and user-side network devices, making the data more multidimensional and breaking the traditional fusion methods. Therefore, a more robust and compatible deep fusion method is needed to achieve the standardized reconstruction of multi-source heterogeneous data and integrate it into a unified system, which can improve the value of data utilization and establish a reliable and practical system.

[0005] Because these data come from different systems and the modeling standards followed by each system are not uniform, it is difficult to integrate the data involved in the current system, resulting in poor data fusion effect of multi-source heterogeneous data in the distribution network.

[0006] In the related technology, Chinese invention patent application document with publication number CN110674311A discloses a method for fusion of heterogeneous power asset data based on knowledge graphs. It extracts features only through simple machine learning methods such as POS-CBOW association model and regularization strategy, which is suitable for feature extraction of data containing only one power source.

[0007] Chinese invention patent application CN115456080A discloses a two-stage fusion method and system for multi-source data. This method involves acquiring multi-source heterogeneous data, including data from communication networks and power grids; extracting initial standardized reconstruction features from the multi-source heterogeneous data; and performing data fusion based on these initial standardized reconstruction features. However, this scheme only considers the extraction of shallow and deep static features in the feature extraction stage, neglecting the extraction of dynamic features. Since the power grid state changes over time, the data fusion result cannot accurately reflect the power grid state. Summary of the Invention

[0008] The technical problem to be solved by this invention is how to achieve the initial standardized reconstruction of heterogeneous power grid data.

[0009] The present invention solves the above-mentioned technical problems through the following technical means:

[0010] On the one hand, this invention proposes a method for standardizing and reconstructing power grid data, the method comprising the following steps:

[0011] Acquire multi-source heterogeneous data of the power grid, wherein the multi-source heterogeneous data includes structured data, semi-structured data and unstructured data;

[0012] Feature extraction is performed on the structured data, the semi-structured data, and the unstructured data respectively to obtain initial semantic features of the attributes;

[0013] From the perspective of static and temporal dynamic features, the initial semantic features of the attributes are divided into a set of static attributes and a set of dynamic features with temporal relationships.

[0014] Based on the dynamic feature set, a time-series feature set of attributes is extracted by fusion in the time dimension;

[0015] The static feature set and the temporal feature set of the attributes are concatenated to obtain the initial standardized reconstructed features.

[0016] Furthermore, the acquisition of multi-source heterogeneous data from the power grid includes:

[0017] The multi-source heterogeneous data is obtained from a communication network and a power grid, wherein the communication network includes user-side network equipment and the power grid includes a power management system and a user electricity consumption information system.

[0018] Further, the step of extracting features from the structured data, the semi-structured data, and the unstructured data to obtain initial semantic features of the attributes includes:

[0019] The structured data is used to extract features using a one-hot text semantic model to obtain the initial semantic features of the first user attribute;

[0020] The word2vec model is used to extract features from the semi-structured data to obtain the initial semantic features of the second user attributes;

[0021] The BERT model is used to extract features from the unstructured data to obtain the initial semantic features of the third user attribute;

[0022] The initial semantic features of the attributes are obtained based on the first initial semantic features of user attributes, the second initial semantic features of user attributes, and the third initial semantic features of user attributes.

[0023] Furthermore, the word2vec model includes an input layer, a hidden layer, and an output layer connected in sequence;

[0024] The input layer includes 2m nodes, which are used to encode the input semi-structured data and generate 2m word vectors, where m is the window size.

[0025] The input layer to the hidden layer is used to average 2m word vectors. The hidden layer includes n neurons, and the output of the hidden layer is... Let V represent the intermediate features extracted from the input semi-structured data, where V represents the parameter matrix and x represents the intermediate features extracted from the input semi-structured data. t+i This represents the vector of the (t+i)th word;

[0026] The hidden layer to the output layer comprises |V| neuron nodes, and the output of the output layer is: U is a |V|×n dimensional matrix. The value at the i-th position indicates that the central word is word w. i The probability of this is used as the initial semantic feature of the second user attribute.

[0027] Furthermore, the step of using the BERT model to extract features from the unstructured data to obtain the initial semantic features of the third user attribute includes:

[0028] A [CLS] tag is inserted before each sentence in the unstructured data, and a [SEP] tag is inserted after each sentence. The [CLS] tag is used to aggregate the feature vector of a sentence or a pair of sentences.

[0029] For sentence i, specify an embedding interval, provided that i is either odd or even.

[0030] The feature vector Ti, which is aggregated from the i-th CLS marker, will be used as the representation of sentence i.

[0031] Based on the representation of i, dense features Y are predicted through the summarization layer. i This serves as the initial semantic feature of the third user attribute.

[0032] Furthermore, the step of fusing and extracting temporal features of attributes based on the dynamic feature set in the time dimension includes:

[0033] The dynamic feature set is fused and extracted using an LSTM recurrent neural network in the time dimension.

[0034] Furthermore, the initial standardized reconstructed features are obtained by concatenating the static feature set and the temporal feature set of the attributes, as expressed by the formula:

[0035] F = S (u,s) +S (u,t)+ S (e,s)+ S (e,s)

[0036] In the formula: S (u,s) S is the set of static attribute features of user-side network device data. (u,t) S is the set of time-series characteristics of user-side network device data attributes. (e,s) S is the set of static attribute features of power grid data. (e,s) It is a set of time-series features of the attributes of power grid data.

[0037] Furthermore, this invention also proposes a power grid data standardization and reconstruction system, the system comprising:

[0038] The data acquisition module is used to acquire multi-source heterogeneous data of the power grid, including structured data, semi-structured data and unstructured data.

[0039] The feature extraction module is used to extract features from the structured data, the semi-structured data, and the unstructured data respectively to obtain initial semantic features of the attributes;

[0040] The feature segmentation module is used to divide the initial semantic features of the attribute into a set of static attribute features and a set of dynamic features with temporal relationships from the perspectives of static and temporal dynamic features.

[0041] The temporal feature extraction module is used to extract a set of attribute temporal features by fusing the dynamic feature set in the time dimension.

[0042] The feature concatenation module is used to concatenate the static feature set of attributes and the temporal feature set of attributes to obtain initial standardized reconstructed features.

[0043] Furthermore, the present invention also proposes a power grid data standardization and reconstruction device, the device including a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the power grid data standardization and reconstruction method as described above.

[0044] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the power grid data standardization and reconstruction method as described above.

[0045] The advantages of this invention are:

[0046] (1) Since the multi-source heterogeneous data of the power grid comes from different systems and the modeling standards followed by each system are not uniform, the difficulty of data fusion is increased, resulting in poor fusion effect of multi-source heterogeneous data of the distribution network. This invention addresses the problem of standardized reconstruction of multi-source heterogeneous data of the power grid. Based on the degree of structuring, it divides the data into three types of processable data: structured, semi-structured, and unstructured. It then performs differentiated processing on the three types of data to achieve feature extraction of the three types of data and reduce the differences of heterogeneous data. Furthermore, it divides the data from the perspective of static and time-series dynamic features. For the dynamic feature set with time-series relationship, it further extracts time-series related attribute features and finally maps the multi-source heterogeneous data to the attribute static feature set and attribute time-series feature set. The attribute static feature set and attribute time-series feature set are then concatenated to achieve the purpose of initial standardized reconstruction of heterogeneous data. In addition to considering static characteristics, this invention further analyzes and integrates time-series dynamic characteristics. Considering time-series dynamic characteristics can provide more comprehensive and accurate information on power grid characteristics, helping to understand the evolution trend of the power grid system, detect anomalies and faults, and support power grid prediction and optimization, thereby improving the reliability, stability and operating efficiency of the power grid. Because the power grid state changes over time, by considering time-series characteristics, it is possible to provide better, more real-time and accurate judgments and decisions on the current power grid state.

[0047] (2) By using deep networks to extract and fuse static and temporal features of attributes, structured data, semi-structured data and unstructured data from different systems are mapped through deep mapping, so that the system source of the data is not distinguished and all are unified to the feature fusion from the perspective of static and dynamic, thereby further achieving standardization.

[0048] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the power grid data standardization and reconstruction method proposed in this invention;

[0050] Figure 2 This is a schematic diagram of the initial standardized reconstruction technology structure for heterogeneous data from power grid data and user-side network equipment in this invention;

[0051] Figure 3 This is a schematic diagram illustrating the generation principle of the initial standardized reconstructed features in this invention;

[0052] Figure 4 This is a schematic diagram of the one-hot encoding process in this invention;

[0053] Figure 5 This is a schematic diagram illustrating the process of feature extraction from unstructured data in this invention;

[0054] Figure 6 This is a schematic diagram of the Bert model in this invention;

[0055] Figure 7 This is a schematic diagram of the LSTM recurrent neural network structure in this invention;

[0056] Figure 8 This is a schematic diagram of the power grid data standardization and reconstruction system proposed in this invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1 As shown, the first embodiment of the present invention proposes a method for standardizing and reconstructing power grid data, the method comprising the following steps:

[0059] S10. Obtain multi-source heterogeneous data of the power grid, wherein the multi-source heterogeneous data includes structured data, semi-structured data and unstructured data;

[0060] S20. Extract features from the structured data, the semi-structured data, and the unstructured data respectively to obtain initial semantic features of the attributes;

[0061] S30. From the perspective of static and temporal dynamic features, the initial semantic features of the attribute are divided into a set of static attribute features and a set of dynamic features with temporal relationships.

[0062] It should be noted that since the features are extracted from various attributes or relationships provided by the power grid or user-side network, if the source attribute data contains time information, the extracted features are dynamic features with time-series relationships; if the source attribute data does not contain time information, the extracted features are fixed static features.

[0063] S40. Based on the dynamic feature set, perform fusion extraction of the attribute temporal feature set in the time dimension;

[0064] S50. Concatenate the static feature set of attributes and the temporal feature set of attributes to obtain the initial standardized reconstructed features.

[0065] This embodiment addresses the standardization and reconstruction problem of multi-source heterogeneous data in power grids. Based on the degree of structuring, it categorizes the data into three types: structured, semi-structured, and unstructured. Differential processing is then applied to each of these three types to extract their features and reduce the differences between heterogeneous data. Furthermore, based on static and time-series dynamic features, time-series-related attribute features are extracted from the dynamic feature set with time-series relationships. Finally, the multi-source heterogeneous data is mapped to a static attribute feature set and a time-series attribute feature set. These two sets are then concatenated to achieve the initial standardization and reconstruction of the heterogeneous data.

[0066] Since static features only reflect the fixed state of the power grid over a long period of time, while the state of the power grid changes over time, this embodiment further considers the fusion of time-series features, which has the following advantages:

[0067] Revealing the evolution trend of the system: Time-series dynamic characteristics can reflect the evolution trend of the power grid system over time. By analyzing time-series data, we can discover the periodic changes, trend evolution, and seasonal fluctuations of various factors in the power grid, thereby gaining a better understanding of the development status of the power grid system.

[0068] Detecting anomalies and faults: Time-series dynamic characteristics can help monitor abnormal conditions and fault events in the power grid. By analyzing time-series data, abnormal fluctuations, sudden changes, or trends deviating from the normal range can be identified, allowing for the early detection of potential fault points in the power grid and the implementation of corresponding repair or prevention measures.

[0069] Supporting Prediction and Optimization: Time-series dynamic characteristics are crucial for power grid prediction and optimization. Modeling and analyzing historical time-series data enables tasks such as power grid load forecasting, renewable energy generation forecasting (wind power, solar power, etc.), and power grid stability assessment. This provides decision support for power grid operators, optimizing grid dispatch and resource allocation.

[0070] In one embodiment, step S10, acquiring multi-source heterogeneous data of the power grid, specifically includes:

[0071] The multi-source heterogeneous data is obtained from a communication network and a power grid, wherein the communication network includes user-side network equipment and the power grid includes a power management system and a user electricity consumption information system.

[0072] It should be noted that due to the diverse data formats and multiple data sources involved in power grid construction and renovation, power grid data itself contains a large amount of heterogeneous content. Although there are traditional standardization mechanisms for power data in power grid data research, such as those used in PMS systems and user electricity information systems, the introduction of user network-side equipment data to enhance real-time service and power supply quality in the distribution network makes the data more granular and multidimensional, significantly increasing the difficulty of data feature standardization and reconstruction, and breaking the traditional integration methods. This embodiment starts with the structured differentiation of the original data, and uses targeted models to differentiate them, further reducing the differences between heterogeneous data from different sources, while also improving the semantic discrimination capability of data features. By utilizing real-time network data stream feedback and combining it with traditional power data research, it is integrated into a unified system, which can improve the value of data utilization, establish a reliable and practical system, and further strengthen the planning of the distribution network.

[0073] In one embodiment, step S20, which involves extracting features from the structured data, the semi-structured data, and the unstructured data to obtain initial semantic features of the attributes, specifically includes the following steps:

[0074] S21. Use a one-hot text semantic model to extract features from the structured data to obtain the initial semantic features of the first user attribute;

[0075] S22. Use the word2vec model to extract features from the semi-structured data to obtain the initial semantic features of the second user attributes;

[0076] S23. Use the BERT model to extract features from the unstructured data to obtain the initial semantic features of the third user attribute;

[0077] S24. The initial semantic features of the first user attribute, the initial semantic features of the second user attribute, and the initial semantic features of the third user attribute are concatenated to obtain the initial semantic features of the attribute.

[0078] It should be noted that this embodiment addresses text data structures originating from different user-side network devices and power grid systems by uniformly classifying them into structured, semi-structured, and unstructured raw text data based on their level of data standardization. Structured data is generally stored in relational databases; semi-structured data refers to data with certain characteristics to a certain extent, such as power grid equipment technical supervision work standard libraries, work plans, evaluation reports, network activity records, annual and monthly reports, and early warning notices; unstructured data refers to text data conforming to natural language standards, such as defect and fault reports, expert knowledge and experience, etc. By utilizing these three data type classifications and selectively employing three levels of pre-trained networks to extract relevant discriminative features, the differences between heterogeneous data can be further reduced.

[0079] Furthermore, for standard structured text data, this embodiment can utilize the underlying semantic extraction model to simply represent it as a one-hot encoded representation, such as... Figure 4 As shown, this approach is simple and feasible for operating on standard structured raw data, while reducing computational overhead.

[0080] For semi-structured data involving complex and difficult-to-convert words, a simple and feasible word2vec pre-trained model can be used as the initial feature extractor. For unstructured raw text data with more complex natural language structure features, the BERT pre-trained model, which is widely used in natural language processing, can be used as a good feature extraction model for complex text structures.

[0081] In one embodiment, in step S21: using a one-hot text semantic model to extract features from the structured data to obtain the initial semantic features of the first user attribute, the structured text data is simply represented as a one-hot encoding using the underlying semantic extraction model; all text in the corpus is segmented into words, all words (vocabulary) are collected and numbered to construct a vocabulary list, the text is segmented into words and each word is converted into a vector; or the text is segmented into characters and each character is converted into a vector.

[0082] In one embodiment, the word2vec model includes an input layer, a hidden layer, and an output layer connected in sequence;

[0083] The input layer includes 2m nodes, which are used to encode the input semi-structured data and generate 2m word vectors, where m is the window size.

[0084] The input layer to the hidden layer is used to average 2m word vectors. The hidden layer includes n neuron nodes, and the parameter matrix from the input layer to the hidden layer is V. n×|V|The output of the hidden layer is Where i∈{-m,-m+1,…,m-1,m}\{0}, Let V represent the intermediate features extracted from the input semi-structured data, where V represents the parameter matrix and x represents the intermediate features extracted from the input semi-structured data. t+i This represents the vector of the (t+i)th word;

[0085] The hidden layer to the output layer comprises |V| neuron nodes, and the output of the output layer is: U is a |V|×n dimensional matrix. The value at the i-th position indicates that the central word is word w. i The probability of this is used as the initial semantic feature of the second user attribute.

[0086] Specifically, for the input semi-structured data of 2m words, one-hot encoding can yield 2m word vectors, while x t+i It is the word vector generated by the (t+i)th word; V represents the learnable parameter matrix; The new feature obtained by multiplying the x vector generated from all 2m words and the learnable parameter matrix V is represented as the intermediate feature extracted from the input semi-structured data.

[0087] It should be noted that this embodiment is for semi-structured text data, which involves some complex and difficult-to-convert words. This embodiment uses a pre-trained word2vec model as the initial feature extractor, as shown in the figure, to extract features from the semi-structured data. This method is simple and feasible. The vector output by the output layer of the word2vec model... The vector of the input layer is x t+i Although the dimensions are the same, It is not a one-hot vector, and the vector Each element in the matrix is ​​meaningful; the value at the i-th position indicates that the head word is word w. i The probability of.

[0088] In one embodiment, step S23: using the BERT model to extract features from the unstructured data to obtain initial semantic features of the third user attribute, specifically includes the following steps:

[0089] A [CLS] tag is inserted before each sentence in the unstructured data, and a [SEP] tag is inserted after each sentence. The [CLS] tag is used to aggregate the feature vector of a sentence or a pair of sentences.

[0090] For sentence i, specify an embedding interval, provided that i is either odd or even.

[0091] The feature vector Ti, which is aggregated from the i-th CLS marker, will be used as the representation of sentence i.

[0092] Based on the representation of i, dense features Y are predicted through the summarization layer. i This serves as the initial semantic feature of the third user attribute.

[0093] It should be noted that this embodiment is more suitable for unstructured raw text data with more complex natural language structure features. By utilizing the BERT pre-trained model, which is widely used in natural language processing, it can serve as a good feature extraction model for complex structured text.

[0094] like Figure 5 As shown, a [CLS] token is inserted before each sentence, and a [SEP] token is inserted after each sentence. In the model, [CLS] is used as a symbol to aggregate the features of a sentence or a pair of sentences. We learn the model by using multiple [CLS] tokens to obtain the features of sentences in ascending order of symbols. Even a single sentence can have relevant features extracted using this model.

[0095] This embodiment uses interval segment embedding to distinguish multiple sentences in a document. For sentence i, we will specify an embedding E. A or E B The segment is defined if i is either odd or even. For example, for [sent1, sent2, sent3, sent4, sent5], we will assign [E... A E B E A E B E A The vector Ti, which is the vector of the i-th (CLS) symbol from the top BERT layer, will be used as the representation of sentence i.

[0096] After obtaining sentence vectors from BERT, several summarization-specific layers are built on top of the BERT output to capture document-level features for summarization. For each sentence i, the final predicted dense feature Yi is computed through the summarization layer. This is then divided into attribute static features Y. (,s) i And the dynamic characteristics Y with time sequence relations (,t) i (t=T).

[0097] Furthermore, the BERT model framework, such as Figure 6As shown, through iterative multi-layer encoder attention structures, residual connections directly add the input and output of the module as the final output. A regularization layer standardizes a node of a certain neural network layer with 0 mean and 1 variance. Linear transformations are then applied twice to the enhanced semantic vector of each character to enhance the expressive power of the entire model.

[0098] In one embodiment, step S40: based on the dynamic feature set, performing fusion extraction of the attribute temporal feature set in the time dimension, specifically includes:

[0099] The dynamic feature set is fused and extracted using an LSTM recurrent neural network in the time dimension.

[0100] It should be noted that, after extracting the relevant attribute initialization features in this embodiment, the attribute initialization features are then divided from the perspectives of static and temporal dynamic features. This is for the attribute feature set C = {Y} which contains temporal relationships. t i (t=1),Y t i (t=2),···,Y t i (t=n-1),Y t i (t=n)}, using an LSTM recurrent neural network to fuse and extract dynamic temporal features H in the time dimension. t i .

[0101] Furthermore, the LSTM structure is as follows: Figure 7 As shown, the operation involves three gate structures: an input gate, a forget gate, and an output gate. The forget gate is the output H of the previous unit. t i (t=T-1) and the input Y of this unit t i (t=T) is the input sigmoid function, which is H t i Each term in (t=T-1) generates a value in [0,1] (which can be seen as a probability value) to control the degree to which the state of the previous unit is forgotten:

[0102] f t =σ(W f ·[H t-1 ,Y t ]+b f )

[0103] The input gate, in conjunction with a tanh function, controls what new information is added. The tanh function generates a new candidate vector, and the input gate generates a value in the range [0,1] for each term, controlling the amount of new information added. This yields the output f of the forget gate. t Used to control the degree to which the previous unit is forgotten, resulting in the output i of the input gate. t This is used to control how much new information is added, and then update the cell state of this memory unit. The formula is expressed as:

[0104] i t =σ(W i ·[H t-1 ,Y t ]+b i )

[0105] C t =tanh(W C ·[H t-1 ,Y t ]+b C )

[0106] The output gate controls how much of the current cell state is filtered out. First, the cell state is activated, and then the output gate generates a value in the range [0,1] for each item, controlling the degree to which the control state is filtered.

[0107] o t =σ(W o [H t-1 ,Y t ]+b o )

[0108] H t =o t *tanh(C t )

[0109] In the formula: f t The output timing characteristics of the forget gate; W f H is the learnable weight matrix for the forget gate; t-1 The attribute dynamic temporal feature of the previous time step; Y t b represents the dynamic characteristics of the time step with temporal relationships. f W represents the learnable bias parameters for the forget gate. i W is the learnable weight matrix of the input gate; C b is the learnable weight matrix for candidate memory temporal states; C Learnable bias parameters for candidate memory temporal states; i t The output timing characteristics of the input gate; C t Candidate memory temporal states; o t The output attributes of the output gate are dynamic timing characteristics; Wo b is the learnable weight matrix of the output gate; o The learnable bias parameters for the output gate; o t For the dynamic timing characteristics of the output attributes of the output gate; C t These are candidate memory temporal states.

[0110] This embodiment introduces an external user-side network data source, which further increases the difficulty of feature extraction. Based on the characteristics of the three data structures, three deep feature extraction models are used respectively: one-hot, word2Vec, and BERT model to extract features, which can further explore semantic features.

[0111] In one embodiment, step S50 involves concatenating the static feature set and the temporal feature set of the attributes to obtain initial standardized reconstructed features, expressed by the formula:

[0112] F = S (u,s) +S (u,t)+ S (e,s)+ S (e,s)

[0113] In the formula: S (u,s) S is the set of static attribute features of user-side network device data. (u,t) S is the set of time-series characteristics of user-side network device data attributes. (e,s) S is the set of static attribute features of power grid data. (e,s) It is a set of time-series features of the attributes of power grid data.

[0114] Furthermore, for user-side network device data, the set of static attribute features and the set of time-series attribute features of user-side network device data can be concatenated; for power grid data, the set of static attribute features and the set of time-series attribute features of power grid data can also be concatenated.

[0115] It should be noted that, as Figures 2 to 3 As shown, the initial standardized feature extraction framework based on heterogeneous data from power grid data and user-side network equipment consists of three parts. For different structured text data from user-side network equipment and the power grid system, targeted pre-trained models are used to extract corresponding attribute features; then, these features are divided into static attribute features Y. (,s) i And the dynamic characteristics Y with time sequence relations (,t) i (t=T), for the attribute feature set C={Y} containing temporal relationships. t i (t=1),Y t i (t=2),···,Yt i (t=n-1),Y t i (t=n)}, using an LSTM recurrent neural network to fuse and extract dynamic temporal features H in the time dimension. t i Thus, the static feature set S of user-side network device data attributes is obtained respectively. (u,s) ={Y (u,s) 1 ,Y (u,s) 2 ,···,Y (u,s) n-1 ,Y (u,s) n User-side network device data attribute time-series feature set S (u,t) ={H (u,t) 1 H (u,t) 2 ,···,H (u,t) n-1 H (u,t) n}、Static feature set S of power grid data attributes (e,s) ={Y (e,s) 1 ,Y (e,s) 2 ,···,Y (e,s) n-1 ,Y (e,s) n} and the time series feature set S of power grid data attributes (e,t) ={H (e,t) 1 H (e,t) 2 ,···,H (e,t) n-1 H (e,t) n Finally, through feature cascading at the attribute level, the initial standardized reconstruction features F for the heterogeneous data of power grid data and user-side network equipment are obtained.

[0116] In addition, such as Figure 8 As shown, the second embodiment of the present invention also proposes a power grid data standardization and reconstruction system, the system comprising:

[0117] The data acquisition module 10 is used to acquire multi-source heterogeneous data of the power grid, including structured data, semi-structured data and unstructured data.

[0118] The feature extraction module 20 is used to extract features from the structured data, the semi-structured data, and the unstructured data respectively to obtain initial semantic features of the attributes;

[0119] The feature segmentation module 30 is used to segment the initial semantic features of the attribute into a set of static attributes and a set of dynamic features with temporal relationships from the perspectives of static and temporal dynamic features.

[0120] The temporal feature extraction module 40 is used to extract a set of attribute temporal features by fusing the dynamic feature set in the time dimension.

[0121] The feature concatenation module 50 is used to concatenate the static feature set of attributes and the temporal feature set of attributes to obtain initial standardized reconstructed features.

[0122] This embodiment addresses the standardization and reconstruction problem of multi-source heterogeneous data in power grids. Based on the degree of structuring, it categorizes the data into three types: structured, semi-structured, and unstructured. Differential processing is then applied to each of these three types to extract their features and reduce the differences between heterogeneous data. Furthermore, based on static and time-series dynamic features, time-series-related attribute features are extracted from the dynamic feature set with time-series relationships. Finally, the multi-source heterogeneous data is mapped to a static attribute feature set and a time-series attribute feature set. These two sets are then concatenated to achieve the initial standardization and reconstruction of the heterogeneous data.

[0123] In one embodiment, the data acquisition module 10 is specifically used to acquire the multi-source heterogeneous data from a communication network and a power grid. The communication network includes user-side network equipment, and the power grid includes a PMS system and a user electricity consumption information system.

[0124] In one embodiment, the feature extraction module 20 specifically includes:

[0125] The first feature extraction unit is used to extract features from the structured data using a one-hot text semantic model to obtain the initial semantic features of the first user attribute.

[0126] The second feature extraction unit is used to extract features from the semi-structured data using the word2vec model to obtain the initial semantic features of the second user attributes.

[0127] The third feature extraction unit is used to extract features from the unstructured data using the BERT model to obtain the initial semantic features of the third user attribute.

[0128] The feature combination unit is used to obtain the initial semantic features of the attribute based on the first initial semantic features of the user attribute, the second initial semantic features of the user attribute, and the third initial semantic features of the user attribute.

[0129] In one embodiment, the word2vec model includes an input layer, a hidden layer, and an output layer connected in sequence;

[0130] The input layer includes 2m nodes, which are used to encode the input semi-structured data and generate 2m word vectors, where m is the window size.

[0131] The input layer to the hidden layer is used to average 2m word vectors. The hidden layer includes n neurons, and the output of the hidden layer is... Let V represent the intermediate features extracted from the input semi-structured data, where V represents the parameter matrix and x represents the intermediate features extracted from the input semi-structured data. t+i This represents the vector of the (t+i)th word;

[0132] The hidden layer to the output layer comprises |V| neuron nodes, and the output of the output layer is: U is a |V|×n dimensional matrix. The value at the i-th position indicates that the central word is word w. i The probability of this is used as the initial semantic feature of the second user attribute.

[0133] In one embodiment, the third feature extraction unit is specifically used for:

[0134] A [CLS] tag is inserted before each sentence in the unstructured data, and a [SEP] tag is inserted after each sentence. The [CLS] tag is used to aggregate the feature vector of a sentence or a pair of sentences.

[0135] For sentence i, specify an embedding interval, provided that i is either odd or even.

[0136] The feature vector Ti, which is aggregated from the i-th CLS marker, will be used as the representation of sentence i.

[0137] Based on the representation of i, dense features Y are predicted through the summarization layer. i This serves as the initial semantic feature of the third user attribute.

[0138] In one embodiment, the temporal feature extraction module 40 is specifically used to: use an LSTM recurrent neural network to fuse and extract the temporal features of the dynamic feature set in the time dimension.

[0139] In one embodiment, the feature concatenation module 50 is specifically used to concatenate the static attribute feature set and the temporal attribute feature set to obtain initial standardized reconstructed features, expressed by the formula:

[0140] F = S (u,s) +S (u,t)+ S (e,s)+ S (e,s)

[0141] In the formula: S (u,s) S is the set of static attribute features of user-side network device data. (u,t) S is the set of time-series characteristics of user-side network device data attributes. (e,s) S is the set of static attribute features of power grid data. (e,s) It is a set of time-series features of the attributes of power grid data.

[0142] It should be noted that other embodiments or implementation methods of the power grid data standardization and reconstruction system described in this invention can refer to the above-described method embodiments, and will not be repeated here.

[0143] Furthermore, the third embodiment of the present invention also proposes a power grid data standardization and reconstruction device, the device including a memory and a processor; wherein, the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the power grid data standardization and reconstruction method described in the first embodiment above.

[0144] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the power grid data standardization and reconstruction method as described in the first embodiment above.

[0145] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0146] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0147] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for standardizing and reconstructing power grid data, characterized in that, The method includes: Acquire multi-source heterogeneous data of the power grid, wherein the multi-source heterogeneous data includes structured data, semi-structured data and unstructured data; Feature extraction is performed on the structured data, semi-structured data, and unstructured data respectively to obtain initial semantic features of attributes. This includes using a one-hot text semantic model to extract features from the structured data to obtain first initial semantic features of user attributes; using a word2vec model to extract features from the semi-structured data to obtain second initial semantic features of user attributes; using a BERT model to extract features from the unstructured data to obtain third initial semantic features of user attributes; and obtaining the initial semantic features of attributes based on the first, second, and third initial semantic features of user attributes. Specifically, the step of using a BERT model to extract features from the unstructured data to obtain the third initial semantic features of user attributes includes: inserting a [CLS] tag before each sentence in the unstructured data and a [SEP] tag after each sentence. The [CLS] tag is used to aggregate the feature vectors of a sentence or a pair of sentences. For a sentence... i Specify an embedded interval, with the following conditions: i Is it an odd or even number; will the first i The feature vector T aggregated from CLS tags i will be used as a sentence i The representation; based on the i The representation is that dense features Y are predicted through the summary layer. i As the initial semantic feature of the third user attribute; From the perspective of static and temporal dynamic features, the initial semantic features of the attributes are divided into a set of static attributes and a set of dynamic features with temporal relationships. Based on the dynamic feature set, a time-series feature set of attributes is extracted by fusion in the time dimension; The static feature set and the temporal feature set of the attributes are concatenated to obtain the initial standardized reconstructed features.

2. The power grid data standardization and reconstruction method as described in claim 1, characterized in that, The acquisition of multi-source heterogeneous data from the power grid includes: The multi-source heterogeneous data is obtained from a communication network and a power grid, wherein the communication network includes user-side network equipment and the power grid includes a power management system and a user electricity consumption information system.

3. The power grid data standardization and reconstruction method as described in claim 1, characterized in that, The word2vec model comprises an input layer, a hidden layer, and an output layer connected in sequence. The input layer includes 2m nodes, which are used to encode the input semi-structured data and generate 2m word vectors, where m is the window size. The input layer to the hidden layer is used to perform averaging on 2m word vectors. The hidden layer includes... n The hidden layer has 100 neurons and its output is 100. , This represents the intermediate features extracted from the input semi-structured data. V Represents the parameter matrix, Indicates the ( t + i ) word vectors; The hidden layer to the output layer includes | V | neurons, and the output of the output layer is , U It is a | V |× n A 3D matrix The Middle i The numerical values ​​at each position indicate that the central word is a word. The probability of this is used as the initial semantic feature of the second user attribute.

4. The power grid data standardization and reconstruction method as described in claim 1, characterized in that, The step of fusing and extracting temporal features of attributes based on the dynamic feature set in the time dimension includes: The dynamic feature set is fused and extracted using an LSTM recurrent neural network in the time dimension.

5. The power grid data standardization and reconstruction method as described in claim 2, characterized in that, The initial standardized reconstructed features are obtained by concatenating the static feature set and the temporal feature set of the attributes, as expressed by the formula: F = S (u,s) + S (u,t) + S (e,s) + S (e,s) In the formula: S (u,s) S is the set of static attribute features of user-side network device data. (u,t) S is the set of time-series characteristics of user-side network device data attributes. (e,s) S is the set of static attribute features of power grid data. (e,s) It is a set of time-series features of the attributes of power grid data.

6. A power grid data standardization and reconstruction system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source heterogeneous data of the power grid, including structured data, semi-structured data and unstructured data. The feature extraction module is used to extract features from the structured data, the semi-structured data, and the unstructured data respectively to obtain initial semantic features of attributes. This includes using a one-hot text semantic model to extract features from the structured data to obtain a first initial semantic feature of user attributes; using a word2vec model to extract features from the semi-structured data to obtain a second initial semantic feature of user attributes; using a BERT model to extract features from the unstructured data to obtain a third initial semantic feature of user attributes; and obtaining the initial semantic feature of the attribute based on the first, second, and third initial semantic features of user attributes. Specifically, the step of using a BERT model to extract features from the unstructured data to obtain the third initial semantic feature of user attributes includes: inserting a [CLS] tag before each sentence in the unstructured data and a [SEP] tag after each sentence. The [CLS] tag is used to aggregate the feature vector of a sentence or a pair of sentences. For a sentence... i Specify an embedded interval, with the following conditions: i Is it an odd or even number; will the first i The feature vector T aggregated from CLS tags i will be used as a sentence i The representation; based on the i The representation is that dense features Y are predicted through the summary layer. i As the initial semantic feature of the third user attribute; The feature segmentation module is used to divide the initial semantic features of the attribute into a set of static attribute features and a set of dynamic features with temporal relationships from the perspectives of static and temporal dynamic features. The temporal feature extraction module is used to extract a set of attribute temporal features by fusing the dynamic feature set in the time dimension. The feature concatenation module is used to concatenate the static feature set of attributes and the temporal feature set of attributes to obtain initial standardized reconstructed features.

7. A power grid data standardization and reconstruction device, characterized in that, The device includes a memory and a processor; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, so as to implement the power grid data standardization and reconstruction method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power grid data standardization and reconstruction method as described in any one of claims 1-5.

Images