Comprehensive energy knowledge recommendation method and device

Abstract

The invention relates to the technical field of energy knowledge maps, in particular to a comprehensive energy knowledge recommendation method and device, and the method comprises the steps: obtaining an energy information keyword in a question of a user, and obtaining a vector representation of the energy information keyword; forming the vectors of the energy information keywords into time sequence vectors according to a time sequence; according to the time sequence vector, predicting to obtain an interest keyword of the user; nodes similar to the interest keywords are determined in the comprehensive energy knowledge graph, and recommended content is determined. According to the recommendation method based on the comprehensive energy knowledge graph, personalized recommendation is rapidly and accurately performed on the user, the structure of the user and the comprehensive energy information can be reserved to the maximum extent, and then prediction and recommendation are accurately performed.

Description

Technical Field

[0001] The present application relates to the technical field of energy knowledge graphs, and in particular to a comprehensive energy knowledge recommendation method and device. Background Art

[0002] With the development of economy and society, energy demand is increasing day by day, and the integrated energy system has emerged as a new type of integrated energy system. It integrates multiple energy sources in the region, such as oil, coal, natural gas and electricity, through advanced technology and management models, and realizes coordinated planning, interactive response and mutual complementarity among multiple heterogeneous energy subsystems, thereby meeting diversified energy demand, effectively improving energy utilization efficiency and promoting sustainable development of energy.

[0003] However, the integrated energy system contains a large amount of multi-source heterogeneous data. These data are not only complex and huge, but traditional methods have been unable to cope with the need to quickly obtain the required information. In particular, when faced with issues such as how to improve search efficiency and achieve intelligent recommendations, existing technical means seem to be unable to cope with the situation. For example, traditional search methods cannot effectively integrate and utilize the large-scale data accumulated in the integrated energy system, resulting in inaccurate or insufficiently personalized recommendation results, which cannot fully meet the specific needs of users.

[0004] In order to solve the above problems, improve the search efficiency of the comprehensive energy knowledge graph and use it for recommendation, it is urgent to develop a comprehensive energy knowledge recommendation method that can provide more accurate personalized recommendations. Summary of the invention

[0005] In view of this, it is necessary to provide a comprehensive energy knowledge recommendation method and device to solve the technical problem that the personalized recommendation results in the prior art are not accurate enough.

[0006] A comprehensive energy knowledge recommendation method, comprising:

[0007] Obtaining energy information keywords in the user's question and obtaining vector representations of the energy information keywords;

[0008] Forming a time series vector of the energy information keyword vector in chronological order;

[0009] Predicting the user's interest keywords based on the time series vector;

[0010] Nodes similar to the interest keywords are determined in the comprehensive energy knowledge graph, and recommended content is determined.

[0011] Preferably, predicting the user's interest keywords based on the time series vector includes: predicting the user's interest keywords based on the time series vector by using a TCN-BiLSTM-GAT algorithm.

[0012] Preferably, the method further includes: using a sigmoid function to predict the user's unique keyword of interest through a fully connected layer.

[0013] Preferably, determining nodes similar to the interest keywords in the comprehensive energy knowledge graph and determining recommended content includes:

[0014] The breadth-first search algorithm is used to search for nodes similar to the interest keywords in the comprehensive energy knowledge graph, and the recommended content is determined based on the degree of similarity.

[0015] Preferably, it also includes: if a breadth-first search algorithm is used to search for multiple nodes similar to the interest keyword in the comprehensive energy knowledge graph, the Euclidean distance measurement method is used to measure the similarity between the interest keyword and the similar nodes, and the most similar node is determined.

[0016] Preferably, if a keyword without a direct path connection is found in the comprehensive energy knowledge graph using a breadth-first search algorithm, a PRA path reasoning method is used to find potential connection paths, and the paths are sorted according to their weights to determine the most similar nodes.

[0017] Preferably, obtaining energy information keywords in the user's question and obtaining vector representations of the energy information keywords includes:

[0018] Get the user's question content;

[0019] The BERT+BiLSTM+CRF model is used to identify energy information keywords in the user's question content, and the XLM-R model is used to convert the energy information keywords into vector representations.

[0020] Preferably, it also includes establishing a comprehensive energy knowledge graph:

[0021] Acquire multi-source heterogeneous data generated by integrated energy systems;

[0022] Preprocessing the multi-source heterogeneous data;

[0023] Extract entity relationships from preprocessed data;

[0024] Performing knowledge fusion based on the entity relationship;

[0025] Constructing a hierarchical graph of a comprehensive energy knowledge graph; wherein the hierarchical graph includes an energy production layer, an energy management and control layer, an energy power generation layer, and a power grid supply layer;

[0026] Establish a comprehensive energy knowledge graph based on a top-down hierarchical structure.

[0027] Preferably, it also includes: using a Neo4j graph database to store the comprehensive energy knowledge graph.

[0028] Compared with the prior art, the beneficial effect of the present application is that through a recommendation method based on a comprehensive energy knowledge graph, fast and accurate personalized recommendations for users are achieved, the structure of user and comprehensive energy information can be retained to the greatest extent, and accurate predictions and recommendations can be made. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the comprehensive energy knowledge recommendation method provided in this application.

[0030] Figure 2 This is a schematic diagram of the algorithm flow of TCN-BiLSTM-GAT.

[0031] Figure 3 It is a structural diagram of TCN.

[0032] Figure 4 This is a schematic diagram of the process of establishing a comprehensive energy knowledge graph provided by this application.

[0033] Figure 5 It is a hierarchical structure diagram of the comprehensive energy knowledge graph provided by this application.

[0034] Figure 6 It is a schematic diagram of the comprehensive energy pretreatment process provided by this application.

[0035] Figure 7 It is a schematic diagram of the entity relationship extraction model of the comprehensive energy knowledge graph provided in this application.

[0036] Figure 8 It is a structural diagram of the comprehensive energy knowledge recommendation device 300 provided in this application. DETAILED DESCRIPTION

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] like Figure 1 As shown, Figure 1 : is a flowchart of the comprehensive energy knowledge recommendation method provided by this application, and the comprehensive energy knowledge recommendation method includes the following steps:

[0039] S101: Obtain energy information keywords in a user's question and obtain vector representations of the energy information keywords.

[0040] User questions can also include historical search data of the user's questions, and keywords can be extracted from the user's historical search records. The BERT+BiLSTM+CRF model can be used to identify keywords in the user's questions, and the XLM-R model can be used to obtain the embedded word vector representation of the keywords.

[0041] Specifically, first, questions, queries, and search statements are processed by the BERT embedding layer to convert characters into word vectors, and the attention mechanism is used to extract the information features of the characters.

[0042] Subsequently, the word vector is passed to the bidirectional semantic feature extraction layer (bidirectional long short-term memory network, BiLSTM) to form a score matrix reflecting the entity tag corresponding to each character in the sentence. Finally, the conditional random field (CRF) entity sequence information annotation layer is used to adjust the correlation of adjacent tags to obtain the final prediction result, and it is labeled according to the preset annotation scheme. In this way, the entity recognition part can accurately identify the entities in the input question and extract energy information keywords. XLM-R is used to train the model from a large corpus to obtain a pre-trained word vector representation, that is, the energy information keyword sequence is represented by XLM-R, and the energy information keyword vector (R is the vector space) is obtained through XLM-R training as shown below:

[0043]

[0044] S102: The vectors of the energy information keywords are converted into time series vectors in chronological order.

[0045] The keywords are organized in chronological order to form a time series vector representation. Furthermore, if there are multiple keywords in the same time period, the path representation within the time granularity is formed through vector fusion. The formation of a time series vector is to arrange the keywords in the user's historical behavior in chronological order to form an ordered time series vector. This time series vector is the basis for subsequent predictions. It contains the user's activity trajectory in a certain time window in the past and can be used to capture the evolution of the user's interests.

[0046] S103: Predict the user's interest keywords based on the time series vector.

[0047] The predicted keywords can also be said to be information points that users may be interested in. This information point can be understood as a certain aspect or theme in the field of integrated energy that users may pay attention to in the future.

[0048] Further, such as Figure 2 and Figure 3 As shown, Figure 2 This is a schematic diagram of the algorithm flow of TCN-BiLSTM-GAT. Figure 3 This is a schematic diagram of the structure of TCN. The TCN-BiLSTM-GAT algorithm can be used to predict the user's keywords of interest. The algorithm combines the time domain convolutional network (TCN), the bidirectional long short-term memory network (BiLSTM) and the graph attention network (GAT). TCN uses dilated causal convolution to expand the receptive field, BiLSTM makes up for the shortcoming of TCN that it cannot "look back", and GAT helps to explore deep user preferences; thus improving the accuracy and robustness of prediction.

[0049] The prediction here refers to the use of TCN (Temporal Convolutional Network) to capture the characteristics of time series based on the user's past behavior patterns (i.e., keywords in historical search records), BiLSTM (Bidirectional Long Short-Term Memory) to supplement the forward and backward contextual information in the time series, and GAT (Graph Attention Networks) to focus on the key elements in user behavior through the attention mechanism. By combining these models, it is possible to predict the keywords that users may pay attention to in the future, and then provide users with personalized recommendations.

[0050] Furthermore, the fully connected layer uses the sigmoid function to predict the user's unique keyword of interest. The sigmoid function can output a continuous value between 0 and 1, which helps to accurately predict the user's keyword of interest, ensure the uniqueness of the predicted keyword, and reduce the repetition of the recommended content.

[0051] Furthermore, if the breadth-first search algorithm is used to find keywords without direct path connections in the comprehensive energy knowledge graph, the PRA path reasoning method is used to find potential connection paths, and sort them according to the weights of the paths to determine the most similar nodes. When there is no direct path connection between keywords, the PRA path reasoning method can help find potential connection paths, and sort them by weight through the PRA path reasoning method to ensure the rationality of the recommended content.

[0052] S104: Determine nodes similar to the interest keywords in the comprehensive energy knowledge graph and determine recommended content.

[0053] Based on the prediction results, the breadth-first search algorithm is used in the knowledge graph to find the most similar nodes, and the recommendation reasons are given according to the path to improve the interpretability of the recommendation. This means that the system will recommend the content that is closest to the user's past behavior, while also taking into account the user's interest evolution trend. Among them, the breadth-first search algorithm is a graph search algorithm. Its main idea is to start with an initial node, marked as 0, and after visiting the initial node, update the access flag to 1, and then search layer by layer in a radial manner. The flag is 0 before the node that has not been visited, and the node flag is 1 after the visit, ensuring that the node is only visited once. In the recommendation system, the breadth-first search algorithm is used to find the nodes most relevant to the energy information keywords extracted from the user's question in the comprehensive energy knowledge graph. It can help the system determine which nodes are most relevant to the keywords entered by the user, and ensure the order of node access through the "queue principle", thereby improving the accuracy and efficiency of the search.

[0054] The following is an introduction to the comprehensive energy knowledge recommendation method provided by this application in conjunction with a specific embodiment:

[0055] In the knowledge graph, the higher the similarity between energy information keywords and nodes, the greater the access probability of the node. The access probability of the node is defined as:

[0056]

[0057] Where V = (v1, v2, … v m ) represents the node vector representation in the knowledge graph, and the L2 norm is used to calculate the probability of visiting the adjacent points.

[0058] Interest keywords have different path connections on the comprehensive energy knowledge graph, which leads to different representations in the vector space and different amounts of information obtained by users. Therefore, it is necessary to select a suitable path as a multivariate time series representation. In particular, there is not necessarily a path connection for the above energy information keywords, so the PRA path is used for reasoning, and its main steps are as follows:

[0059] Extract paths of different lengths from the knowledge graph. A path is a sequence of nodes (entities) and edges (relationships). Each path represents a certain relationship from one entity to another. Weights are calculated for the extracted paths. PRA assigns a weight to each path, usually adjusting these weights by known relationships in the training data. The weights are usually calculated based on the frequency of the path, the complexity of the path, and the performance of the path in a specific task. Path ranking: Rank all paths according to the calculated weights. Paths with high weights are considered more useful for reasoning tasks. Use the top-ranked paths for knowledge reasoning. Specifically, for a given entity pair (e.g. (entity A, entity B)), PRA uses high-weighted paths to predict possible relationships between them.

[0060] Different energy sources are substitutable to a certain extent, but with the changes in external factors such as user demand and energy cost, the complementary relationship between energy sources will be affected. In the comprehensive energy knowledge graph, with the change of each step of the path, the degree of information represented on the knowledge graph also changes, so the substitution coefficient is used to represent the degree of information. The substitution coefficient is expressed as follows:

[0061]

[0062] Among them, D KL represents the KL distance, v m Indicates the current node.

[0063]

[0064] Among them, D KL (p|q) represents the asymmetric measure of the difference between p and q, that is, the information loss when q is used to replace the true vector p. Specifically, D KL (V|v m ) measures the degree of deviation between the current node and the target node, and α represents the degree of information reduction generated by the alternative energy node compared to the original energy node after replacement. KL The larger it is, the greater the degree of information loss and the smaller the substitution coefficient. Under the influence of the degree of information loss and nodes, the probability of energy triples in the comprehensive energy knowledge graph is:

[0065] P=α m ·v m

[0066] Among them, for the i-th path, its probability is P i .

[0067] Calculate the maximum probability of the energy information keyword path and use it as the keyword path representing the hidden user information:

[0068]

[0069] in is the sigmoid function, Represents a keyword vector.

[0070] Taking hours as the time granularity, the keyword paths obtained above are divided in time order to form a time series vector representation. For multiple keyword path representations in the same time period, vector fusion is performed by addition to form a path representation within the time granularity. In the formula, add represents vector addition fusion.

[0071]

[0072] Furthermore, for multiple nodes in the same path, the head node and the tail node are selected to form the path representation formed above. The triplet, h represents the head entity, t represents the tail entity, and forms a multivariate time series.

[0073] After obtaining the above multivariate time series, the TCN-BiLSTM-GAT algorithm is used for prediction, that is, the h and Predict t.

[0074] TCN-BiLSTM-GAT algorithm For the multivariate time series obtained above, in order to facilitate subsequent calculations, in this embodiment, the embedding layer is first used to obtain a unified dimensional feature of 512 dimensions.

[0075] The time domain convolutional network (TCN) is a network structure based on the improvement of the convolutional neural network. The causal convolution inside it is a unique structure of TCN, which ensures the causality of the input time series, which means that the elements in the output sequence can only depend on the previous elements in the input sequence. In order to reduce information loss, the dilated convolution method is adopted. TCN uses dilated convolution to increase the receptive field. Dilated convolution can reduce the depth of causal convolution, and also ensure that the input and output sizes are the same by skipping input values ​​in a given time step. TCN includes two hole causal convolutions, two instance regularizations, two ReLU activation functions, and two Dropout operations. In order to capture the bidirectional information of the time series, BiLSTM can not only look forward (that is, rely on past inputs), but also look backward (that is, take into account future inputs), so as to better capture the overall characteristics of the time series.

[0076] Further, the time feature is extracted and the data is input into BiLSTM. Specifically, BiLSTM is composed of the forward and reverse LSTM networks, which respectively calculate the values ​​of the forward input and reverse input, obtain the hidden layer state output, and then concatenate the vectors of the forward and reverse hidden layer state outputs to obtain the final output of the BiLSTM network layer. Among them, H is the final output of the BiLSTM network layer; is the positive hidden layer state output; is the reverse hidden layer state output.

[0077] At the same time, in order to better explore the interests and preferences of users, the graph attention network is used in parallel with TCN. For the multivariate time series obtained above, the definition is given:

[0078] Temporal Graph Attention: All timestamps are considered as a complete graph. Each timestamp in the graph is considered as an independent node, and the edges between nodes represent the dependencies between different timestamps. Specifically, the graph attention network introduces a self-attention mechanism in the propagation process, and the hidden state of each node is calculated by paying attention to its neighboring nodes.

[0079]

[0080] h j Represents node h i Neighbor node input features, h′ i is the node output feature, δ represents the sigmoid activation function, α ij is the attention score between node i and its neighbor node j, L is the number of all neighbor nodes of node i, and the feature attention score only considers all first-order neighbors of node i. ij The calculation of is as follows:

[0081]

[0082] e(h i ,h j ) is the scoring function, which is defined as:

[0083] e(h i ,h j )=LeakyReLU(μ T [Wh i ∣∣Wh j ])

[0084] Where LeakyReLU is a nonlinear activation function, μ∈R 2m is the learnable weight vector, m is the feature vector of each node. W is the linear transformation weight matrix applied to each node. || represents the concatenation operation.

[0085] Then, the multivariate time series representation after unified dimension is input into the TCN model. In this embodiment, the number of convolution layers is set to 2, the number of convolution kernels in the first layer is 64, the convolution kernel size is 3, and the expansion rate is 1; the number of convolution kernels in the second layer is 128, the convolution kernel size is 5, the expansion rate is 2, the Dropout rate is set to 0.1, and the number of neurons in the hidden layer is 256. The output feature K is obtained and input into the BiLSTM layer.

[0086] For the BiLSTM layer, the internal hidden layer units of the forward and backward LSTM layers are set to feature dimension 512. The output feature K is obtained ′ , the features after the embedding layer is unified in dimension are input into the graph attention network, the number of hidden layers is set to 3, the multi-head attention mechanism uses 8 attention heads, the number of neurons in the input layer and the output layer is set to the feature dimension, and the output feature is G.

[0087] The output feature K and the output feature G are fused to obtain a new feature R = concat (K, G), and the new feature R is input into the fully connected layer. In order to prevent overfitting, the fully connected layer has 4 hidden layers, one input layer, and one output layer. The sigmoid function is used for the output layer to obtain a unique prediction value.

[0088] For the predicted value t ′ ,The breadth-first search algorithm is used again in the knowledge graph to find similar nodes, and the similarity between them is measured by Euclidean distance. The node with the largest similarity level is taken as the recommendation target, and the reason for the recommendation is given by the above path, which greatly improves the interpretability of the recommendation.

[0089] Furthermore, the comprehensive energy knowledge recommendation method provided in this application also includes establishing a comprehensive energy knowledge graph, such as Figure 4 The schematic diagram of the process of establishing a comprehensive energy knowledge graph shown in FIG. 1 includes the following steps:

[0090] S201: Acquire multi-source heterogeneous data generated by the integrated energy system.

[0091] S202: Preprocessing the multi-source heterogeneous data;

[0092] S203: extracting entity relationships from the preprocessed data;

[0093] S204: performing knowledge fusion based on the entity relationship;

[0094] S205: constructing a hierarchical graph of a comprehensive energy knowledge graph; wherein the hierarchical graph includes an energy production layer, an energy management and control layer, an energy power generation layer, and a power grid power supply layer;

[0095] S206: Establish a comprehensive energy knowledge graph based on a top-down hierarchical structure.

[0096] For details, please refer to Figures 5 to 7 , Figure 5 It is a hierarchical diagram of the comprehensive energy knowledge graph; Figure 6 It is a schematic diagram of the comprehensive energy pretreatment process diagram; Figure 7 It is a schematic diagram of the entity relationship extraction model of the comprehensive energy knowledge graph.

[0097] The specific process of the above steps is as follows: Data preprocessing: The integrated energy system has accumulated a large amount of multi-source heterogeneous data in its long-term operation, which brings great challenges to the subsequent establishment of the integrated energy knowledge graph and its query search and intelligent recommendation. Therefore, appropriate data preprocessing is carried out for the massive structured and unstructured multi-scale and multi-dimensional spatiotemporal data of the integrated energy system to make the data format consistent, the data information refined, and the data storage centralized.

[0098] Then create entity relationships, including entity labeling: Based on the preprocessed data and the expert's background knowledge, use "head entity-relationship-tail entity" to label entities. Entity recognition: Use natural language processing technology to identify entities. Knowledge extraction: On the basis of entity recognition, extract pre-defined entity relationships from the data to obtain the relationship between entities, providing a basis for building a knowledge graph. Knowledge fusion: There is a problem of description diversity in the data in the integrated energy system, and knowledge fusion is required; first, entity disambiguation is performed, and then knowledge fusion is performed by combining knowledge representation learning algorithms and reinforcement learning methods. Hierarchical graph: According to the integrated energy application process, its structural levels are divided, including energy production layer, energy management and control layer, energy power generation layer and power grid supply layer, so as to form a comprehensive energy knowledge graph;

[0099] Among them, the construction of the knowledge graph of the integrated energy system adopts a top-down and bottom-up approach, and the ontology layer is constructed according to the hierarchical division of the above-mentioned integrated energy knowledge graph. Through various types of data after data processing, the pattern layer is instantiated, the ontology is improved through different extraction tasks, and the generated ontology is used to guide the extraction of knowledge, thereby obtaining the ontology from knowledge extraction.

[0100] Furthermore, after obtaining the comprehensive energy knowledge graph, it can also be stored in the Neo4j graph database. Neo4j is a database designed specifically for graph data, which can efficiently execute complex graph queries, especially for path queries and connection queries in the graph, which is more efficient than traditional relational databases or NoSQL databases.

[0101] In summary, this application proposes an intelligent recommendation method based on a comprehensive energy knowledge graph, which aims to achieve accurate personalized recommendations through multiple steps such as keyword acquisition in user questions, time series prediction, node exploration in the graph, and recommended content generation. First, the keywords in the user's question are identified by the BERT+BiLSTM+CRF model, and converted into vector representations using the XLM-R model. Then, the TCN-BiLSTM-GAT algorithm is used to predict the keyword vectors arranged in chronological order to obtain the user's interest keywords. Then, in the comprehensive energy knowledge graph, the breadth-first search algorithm is used to find the nodes most similar to the interest keywords, and the most similar nodes are determined by the Euclidean distance metric. For keywords without direct path connections, the PRA path reasoning method is used to find potential paths. Finally, the recommended content is determined based on the prediction results and path weights, which improves the accuracy and interpretability of the recommendation. In addition, this application also describes in detail the construction process of the comprehensive energy knowledge graph, including data preprocessing, entity relationship extraction, knowledge fusion and graph construction, and uses the Neo4j graph database for storage to support efficient query and recommendation.

[0102] Based on an inventive concept, the present application also provides a comprehensive energy knowledge recommendation device 300, such as Figure 8 The structure diagram of the comprehensive energy knowledge recommendation device 300 shown in FIG. 3 includes:

[0103] The data acquisition module 301 is used to acquire energy information keywords in the user's question and obtain vector representations of the energy information keywords;

[0104] The data processing module 302 is used to form a time series vector of the energy information keyword in time sequence;

[0105] A keyword prediction module 303 is used to predict the user's interest keywords based on the time series vector;

[0106] The content recommendation module 304 is used to determine nodes similar to the interest keywords in the comprehensive energy knowledge graph and determine recommended content.

[0107] The specific implementation of each functional module in this embodiment can refer to the introduction in the above-mentioned comprehensive energy knowledge recommendation method embodiment, and this embodiment will not be repeated. Figure 8 As shown, the comprehensive energy knowledge recommendation device can be logically divided into multiple modules, each module can have different functions, and the function of each module can be realized by the processor in the computing device reading and executing the instructions in the memory. It should be understood that Figure 8It is only a structural schematic diagram of a comprehensive energy knowledge recommendation device 300 exemplarily shown, and the present application does not limit the division of functional modules in the comprehensive energy knowledge recommendation device.

Claims

1. A comprehensive energy knowledge recommendation method, characterized in that: include: Obtaining energy information keywords in the user's question and obtaining vector representations of the energy information keywords; Forming a time series vector of the energy information keyword vector in chronological order; Predicting the user's interest keywords based on the time series vector; Nodes similar to the interest keywords are determined in the comprehensive energy knowledge graph, and recommended content is determined.

2. The method according to claim 1, characterized in that Predicting the user's interest keywords according to the time series vector includes: predicting the user's interest keywords according to the time series vector by using a TCN-BiLSTM-GAT algorithm.

3. The method according to claim 2, characterized in that Also includes: The sigmoid function is used through the fully connected layer to predict the user’s unique keyword of interest.

4. The method according to claim 1, characterized in that: Determine nodes similar to the interest keywords in the comprehensive energy knowledge graph and determine recommended content, including: The breadth-first search algorithm is used to search for nodes similar to the interest keywords in the comprehensive energy knowledge graph, and the recommended content is determined based on the degree of similarity.

5. The method according to claim 4, characterized in that Also includes: If a breadth-first search algorithm is used to search for multiple nodes similar to the interest keyword in the comprehensive energy knowledge graph, the Euclidean distance measurement method is used to measure the similarity between the interest keyword and the similar nodes, and the most similar node is determined.

6. The method according to any one of claims 4 or 5, characterized in that: If the breadth-first search algorithm is used to find keywords without direct path connections in the comprehensive energy knowledge graph, the PRA path reasoning method is used to find potential connection paths, and the paths are sorted according to their weights to determine the most similar nodes.

7. The method according to claim 1, characterized in that Obtaining energy information keywords in the user's question and obtaining vector representations of the energy information keywords includes: Get the user's question content; The BERT+BiLSTM+CRF model is used to identify energy information keywords in the user's question content, and the XLM-R model is used to convert the energy information keywords into vector representations.

8. The method according to claim 1, characterized in that It also includes building a comprehensive energy knowledge graph: Acquire multi-source heterogeneous data generated by integrated energy systems; Preprocessing the multi-source heterogeneous data; Extract entity relationships from preprocessed data; Performing knowledge fusion based on the entity relationship; Constructing a hierarchical graph of a comprehensive energy knowledge graph; wherein the hierarchical graph includes an energy production layer, an energy management and control layer, an energy power generation layer, and a power grid supply layer; Establish a comprehensive energy knowledge graph based on a top-down hierarchical structure.

9. The method according to claim 8, characterized in that Also includes: The Neo4j graph database is used to store the comprehensive energy knowledge graph.

10. A comprehensive energy knowledge recommendation device, characterized in that: include: A data acquisition module, used to acquire energy information keywords in user questions and obtain vector representations of the energy information keywords; A data processing module, used for forming a time series vector from the vectors of the energy information keywords in chronological order; A keyword prediction module, used to predict the user's interest keywords based on the time series vector; The content recommendation module is used to determine nodes similar to the interest keywords in the comprehensive energy knowledge graph and determine recommended content.

Images