Power system graph recommendation system and method based on browsing path

Abstract

The invention provides a power system graph recommendation system and method based on a browsing path, and relates to the technical field of power system visualization. According to the method, operation behavior data of a user in power grid graph browsing software is collected, a graph set frequently accessed by the user and association rules are extracted through an association rule mining algorithm, and combined graph recommendation is generated by combining a key component extraction model. The method specifically comprises the steps of 1) capturing a user operation event based on a point burying technology and performing quantitative storage; 2) mining user preference graphs and association relationships thereof by adopting an improved Apriori algorithm; and 3) calculating the key probability of the components through the logistic regression model, extracting key components, and recombining and recommending the key components. According to the method, the problem that a traditional recommendation system cannot accurately capture an area of interest of the user is solved, automatic and accurate mining of the behavior of the power system graph browsing user is realized, the problem that a new power grid graph cannot be recommended and generated for the user in the traditional recommendation system is solved, and the accuracy and practicability of power grid graph recommendation are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system visualization, and in particular relates to a power system graphic recommendation system and method based on browsing paths. Background Art

[0002] Existing power grid graphic generation and recommendation technologies mostly use manual drawing or simple recommendations based on the type of graphic usage. They are unable to comprehensively mine and analyze large amounts of graphic data and make automated recommendations. The power grid graphics obtained by end users are of a single type and cannot efficiently support upper-level application systems or make decisions on power grid scheduling. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a power system graphic recommendation system and method based on browsing paths.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] The present invention provides a power system graphic recommendation system based on browsing path, which includes three modules: data acquisition module, graphic mining module and component extraction and combination module:

[0006] 1) The data acquisition module is used to collect and store a series of user operations on graphics. By embedding a script in the user's graphics browsing software, the user's operations on the graphics are recorded, including the name of the graphics clicked, the mouse position and the time, and these characteristic operations are quantified and stored in the database;

[0007] 2) The graph mining module performs association rule analysis on the collected data and uses the Apriori principle to find the graphs that users frequently visit as the graphs of interest to users;

[0008] 3) The Graphic Component Extraction and Combination Module uses collected data on user browsing behavior to calculate the probability that the component in the graphic of interest in step 2) is a key component. A higher probability indicates a greater likelihood of the component being a key component. Key components are extracted, representing the areas of interest to the user during browsing. These components are then reassembled into new graphics based on business needs and recommended to the user.

[0009] The present invention provides a recommendation method for a power system graphic recommendation system based on browsing paths, and the specific steps are as follows:

[0010] 1) Data collection and processing;

[0011] 2) Graphic browsing data collection;

[0012] 3) Quantification of user browsing behavior;

[0013] User browsing behaviors are divided into three categories: physiological behaviors, significant behaviors, and indirect behaviors. The collected user browsing data includes: browsing image name, browsing image duration, mouse events, keyboard events, and the data is stored in the corresponding data format.

[0014] 4) User browsing path transaction extraction;

[0015] A browse path transaction is defined as the sequence in which a user browses a set of graphics as follows:

[0016] <...,(p i ,t i ),(p i+1 ,t i+1 ),...(p i+k ,t i+k ),...>

[0017] Two-tuple (p i ,t i ) represents t i Always browse graphics i , and the sequence satisfies the following two conditions:

[0018] S1 satisfies t at any time t i< t i+1 ;

[0019] S2t i+k+1 -t i+k >T,t i -t i-1 >T, T is the time threshold, or there is no n∈i.....i+k such that p n =p i+k+1 ;

[0020] 5) Users prefer graph mining;

[0021] The preferred path mining algorithm based on the Apriori principle is used, the support and confidence thresholds are set to a high value, and the page browsing time weight is added to the support calculation;

[0022] In the user's browsing path, there are two indicators to evaluate whether the user cares about the graphics and whether there is a correlation between graphics: support and confidence. The support represents the probability of the graphics appearing in the browsing path, as shown in formula (1);

[0023]

[0024] Where count(A,B) represents the number of transactions in which both graph A and graph B appear in the browsing path transaction set, D represents the total number of all sample transactions, and the Time function calculates the duration of browsing the graph.

[0025] A confidence index based on conditional probability, i.e., the degree of dependence between graphs, is introduced, as shown in formula (2).

[0026]

[0027] Where P(A,B) represents the probability that graph A and graph B appear simultaneously in a browsing path transaction, and P(A) represents the probability that graph A appears in a browsing path transaction. According to the conditional probability formula, a greater confidence level indicates a higher degree of association between graphs. A group of frequently accessed and associated graphs is considered to be the user's preferred graphs.

[0028] Based on the above formulas and definitions, the mining process of user preference graphs is as follows:

[0029] S1 calculates the support of each power grid graph based on the user's browsing path transaction set Tr = {tr1,tr2…trn}, tri = <…(pl,tl),(pl+1,tl+1),…>, and selects the graph set whose support is greater than a specific threshold, which we call the frequent 1 subgraph set U1 = {pk,pk+1,…}.

[0030] S2 The frequent 1-subgraph U1 generated in the above step is combined with each other to calculate its support, and the frequent 2-subgraph set U2 is generated. k ,p k+1 ),(p k+3 ,p k+5 )…}.

[0031] S3 repeats step 2) until no new frequent subgraph meets the support threshold, and obtains all frequent subgraph sets U1, U2, ..., Un;

[0032] S4 Dependency Graph Mining: Perform association rule mining on the frequent subgraphs U2, ..., Un in sequence, select strong association rules with confidence greater than a preset threshold, and obtain a set of power grid graph browsing rules that the user is interested in and has strong dependency relationships. Each frequent graph in the rule is a user-preferred graph.

[0033] 6) Extraction and combination of key graphics components;

[0034] Extraction of key components;

[0035] Mining the user's preferred graphics set, extracting the key components of each graphic, and judging whether the component is a key component based on the user's operation characteristics on the graphic component;

[0036] The criterion for determining whether a graphic component is a key component is expressed by formula (3).

[0037]

[0038] In the formula, F represents the key indicator of the component, Wi represents the weight of the above i-th feature in the component, and x i represents the i-th eigenvalue of the component, n represents the total number of components contained in the user preference graph, x in It represents the ith eigenvalue of the nth component. The formula calculates the weighted average of each feature after normalization. The larger the index F is, the greater the user's attention to the component. The weight parameter Wi in formula (3) is obtained through the logistic regression model. The specific method is as follows:

[0039] Whether the S1 component is a key component is mapped to 1 and 0, 1 indicates that the component is a key component, and 0 indicates a non-key component;

[0040] The probability of whether a component is a critical component is defined by S2 and expressed by formula (4), where x i The i-th eigenvalue of the component representing the normalized component, y takes 0 or 1. When the probability P exceeds the threshold, it indicates that the component is a key component. The required parameter is W = {w1, w2, w3, w4};

[0041]

[0042] S3 uses maximum likelihood estimation to find the parameter Wi in formula (4) and defines the likelihood function as shown in formula (5);

[0043]

[0044] S4 inverts the likelihood function and uses the stochastic gradient descent method to find the minimum value of the likelihood function, which is the parameter W.

[0045] To summarize, the steps for extracting key components are as follows:

[0046] 1) According to the user operation record, calculate the characteristic value of each component feature in the power grid graph.

[0047] 2) Use maximum likelihood estimation to obtain feature weights;

[0048] 3) Calculate the key indicators of each component using formula (3);

[0049] 4) Select components whose index values are greater than a specific threshold, which are the key components of the graph.

[0050] Key component combinations;

[0051] Based on the key component description information and the association rules obtained previously, two rows of layouts are made with the components in the antecedent and consequent of the association rules to generate a new graph.

[0052] As a further improvement of the method of the present invention, the data collection in step 1) includes embedding points in the graphic browsing software, capturing user operation events, and quantifying and storing user browsing paths.

[0053] As a further improvement to the method of the present invention, the specific steps of step 2) collecting graphic browsing data are as follows:

[0054] The graphic browsing software used by users adopts a B / S architecture and uses JavaScript script code to embed the tracking technology on the graphic browsing software client. Custom code is implanted into the graphic browsing software client to monitor and record the user's operation details on the browser, including keyboard input, mouse clicks, mouse scrolling and mouse movement, and output the time when the action occurs and the relative position on the power grid graphic canvas to the database. At the same time, the name and logo of each graphic browsed by the user are stored in the database.

[0055] As a further improvement to the method of the present invention, the data format of step 3) is specifically as follows, expressed in XML form, wherein the page node represents the total record of browsing graphic data once; the time node represents the duration of browsing the graphic, and the start and end attributes represent the timestamps of starting and ending browsing; the operations node describes a series of operations performed by the user on the graphic; the event node is the user event captured by the embedding script, which corresponds one-to-one to the event specified in the JavaScript language, the type attribute is the event type, the timestamp is the timestamp of the triggering event, the x and y attributes represent the relative position of the mouse on the graphic, and the content of the event node indicates which component of the graphic is triggered, specifically the component ID. Since each component in the CIM / G graphic specification must contain coordinates and ID, the specific x, y and component ID in the event node can all be calculated by the JavaScript script and the CIM / G file.

[0056] As a further improvement to the method of the present invention, the component features extracted from the key components in step S6) include the following features 1 to 4, specifically as follows:

[0057] Feature 1. Number of times a user has logged in to a component.

[0058] Feature 2. The total duration of the user's mouse movement on the component;

[0059] Feature 3. The total distance the wheel event scrolls up and down on the component;

[0060] Feature 4. Evaluation speed of up and down scrolling events on the component.

[0061] As a further improvement of the method of the present invention, in step S6) key component extraction, there are two methods for determining key components through index F. One is to sort the components in descending order according to index F and select the components at the top as key components. The other is to determine that the component is a key component when index F exceeds a certain threshold.

[0062] As a further improvement to the method of the present invention, the properties of the key component combination in step S6) are as follows:

[0063] Among them, title represents the title of the component, source represents the name of the component from the graphic, content represents the content of the component in the original CIM / G file, datasource represents the data source associated with the component, type represents the component type in the CIM / G specification, and target represents the key indicator.

[0064] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0065] 1) Abstract the user's browsing behavior characteristics on power grid graphics and automatically mine the graphical data that users are interested in;

[0066] 2) Extract grid graph components according to weights, combine them to generate new grid graphs, and provide accurate recommendations to users;

[0067] 3) Reduce the cost of grid graphics drawing and support new applications such as power system dispatching, power system monitoring and decision analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is the browsing path mining and recommendation framework of the present invention;

[0069] Figure 2 This is the user browsing behavior data storage format of the present invention;

[0070] Figure 3 is the number of times the user of the present invention accesses the graph;

[0071] Figure 4 is the stochastic gradient descent algorithm of the present invention;

[0072] Figure 5 is the average click rate of the present invention. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0074] Overall framework

[0075] The power system graph recommendation system based on browsing path includes three modules: data acquisition module, graph mining module and component extraction and combination module.

[0076] 1) The data acquisition module is used to collect and store a series of user operations on graphics. By implanting scripts in the user's graphics browsing software, the user's operations on the graphics are recorded, including the name of the clicked graphic, the mouse position and time, etc., and these characteristic operations are quantified and stored in the database.

[0077] 2) The graph mining module performs association rule analysis on the collected data and uses the Apriori principle to find out the graphs that users frequently visit as the graphs that users are interested in.

[0078] 3) The Graphic Component Extraction and Combination Module uses collected data on user browsing behavior to calculate the probability that the component in the graphic of interest in step 2) is a key component. A higher probability indicates a greater likelihood of the component being a key component. Key components (i.e., the areas of interest to the user during browsing) are extracted, reassembled into new graphics based on business needs, and recommended to the user.

[0079] The process of the entire recommendation framework is as follows Figure 1 shown.

[0080] Key technologies

[0081] The key to a browsing path-based graph recommendation system is to mine the graphs that users are interested in and extract the key components in the graphs. The following is a detailed introduction to this.

[0082] Data acquisition and processing

[0083] Data collection includes embedding points in graphic browsing software, capturing user operation events, and quantifying and storing user browsing paths.

[0084] Graphical browsing data collection

[0085] The graphics browsing software used by users adopts a B / S architecture. Therefore, this application uses tracking technology in the form of JavaScript script code to track the graphics browsing software client. Custom code is embedded in the graphics browsing software client to monitor and record user operation details on the browser, including keyboard input, mouse clicks, mouse scrolling, and mouse movement. The time when the action occurred and the relative position on the power grid graphics canvas are output to the database. At the same time, the name and ID of each graphic viewed by the user are stored in the database.

[0086] Quantification of user browsing behavior

[0087] User browsing behavior is divided into three categories: physiological behavior, significant behavior, and indirect behavior. In order to facilitate quantification and collection, and in combination with the business of power grid graphics, the user browsing data collected in this application includes: browsing graphic name, browsing graphic duration, mouse events, and keyboard events. At the same time, the data is stored in Figure 2 In format.

[0088] Figure 2 The data format is expressed in XML, where the page node represents the total record of a single graphic data browse; the time node represents the duration of browsing the graphic, with the start and end attributes representing the timestamps of the start and end of browsing; the operations node describes a series of user operations on the graphic; and the event node represents user events captured by the embedding script, corresponding one-to-one with events specified in the JavaScript language. The type attribute is the event type, the timestamp is the timestamp of the triggering event, the x and y attributes represent the relative position of the mouse on the graphic, and the content of the event node indicates which component of the graphic was triggered, specifically the component ID. Because each component in the CIM / G graphic specification must contain coordinates and an ID, the specific x, y, and component ID in the event node can be calculated by the JavaScript script and the CIM / G file.

[0089] User browsing path transaction extraction

[0090] Define a browsing path transaction as a sequence of a user browsing a set of graphics <..., (p i ,t i ),(p i+1 ,t i+1 ),...(p i+k ,t i+k ),...>, tuple (p i ,t i ) represents t i Always browse graphics i , and the sequence satisfies the following two conditions:

[0091] 1) At any time t, t i< t i+1 ;

[0092] 2)t i+k+1 -t i+k >T,t i -t i-1 >T (T is the time threshold), or there is no n∈i.....i+k such that p n =p i+k+1 ;according to Figure 1 Given the data storage format and browsing path definition, the user browsing path transaction is represented as shown in Table 1.

[0093] Table 1 Browsing path transactions

[0094]

[0095] Users prefer graph mining;

[0096] A preference path mining algorithm based on the Apriori principle is adopted. In order to prevent the generation of too many key components and the combination of the new graph into a large image that is inconvenient for users to browse, the support and confidence thresholds are set to a higher value, and the page browsing time weight is added to the support calculation.

[0097] In the user's browsing path, there are two indicators to evaluate whether the user cares about the graphics and whether there is a correlation between graphics: support and confidence. The support represents the probability of the graphics appearing in the browsing path, as shown in formula (1).

[0098]

[0099] In the formula, count(A,B) represents the number of transactions in the browsing path transaction set where both graphics A and graphics B appear simultaneously, D represents the total number of all sample transactions, and the Time function calculates the duration of browsing graphics. Considering that in transaction set D, although a certain graphic is frequently visited, it is not the graphic that the user really cares about, such as an intermediate navigation graphic. Therefore, a browsing time calculation function is introduced in formula (1). The greater the support, the more frequently the graphic is browsed and the higher the user's attention. Using a single support index to judge preferred graphics cannot accurately achieve graphic combination recommendation, because the user may frequently browse two unrelated graphics. In this case, combining the key components of these two graphics is obviously meaningless. To this end, a confidence index based on conditional probability is introduced, that is, the degree of dependence between graphics, as shown in formula (2).

[0100]

[0101] Where P(A,B) represents the probability that shapes A and B appear simultaneously in a browsing path transaction, and P(A) represents the probability that shape A appears in a browsing path transaction. According to the conditional probability formula, a higher confidence level indicates a higher degree of association between shapes. A group of frequently accessed and associated shapes is considered to be a user's preferred shape.

[0102] Based on the above formulas and definitions, the mining process of user preference graphs is as follows:

[0103] 1) Based on the user's browsing path transaction set Tr = {tr1, tr2...trn}, tri = <...(pl,tl),(pl+1,tl+1),...>, calculate the support of each power grid graph separately and select the graph set whose support is greater than a specific threshold, which we call the frequent 1 subgraph set U1 = {pk,pk+1,...}.

[0104] 2) In the above steps, the support of the two combinations of the frequent 1-subgraph U1 is calculated to generate a frequent 2-subgraph set U2 = {(p k ,p k+1 ),(p k+3 ,p k+5 )…}.

[0105] 3) Repeat step 2) until no new frequent subgraph meets the support threshold, and obtain all frequent subgraph sets U1, U2, ..., Un.

[0106] 4) Dependency graph mining: Association rules are mined for the frequent subgraphs U2, ..., Un in sequence. Strong association rules with confidence greater than a preset threshold are selected to obtain a set of power grid graph browsing rules that users are interested in and have strong dependencies. Each frequent graph in the rule is a user-preferred graph.

[0107] Extraction and combination of key components of graphics

[0108] Key component extraction

[0109] Mining the user's preferred graphics set, extracting the key components of each graphic, this application determines whether a component is a key component based on the user's operation characteristics on the graphic component. Component characteristics include the following features 1 to 4.

[0110] Feature 1. Number of times a user has logged into a component.

[0111] Feature 2. The total duration that the user moves the mouse over the component.

[0112] Feature 3. The total distance scrolled up and down by the scroll wheel event on the component.

[0113] Feature 4. Evaluation speed of up and down scrolling events on the component.

[0114] Based on the above characteristics, the judgment index of whether a graphic component is a key component is expressed by formula (3).

[0115]

[0116] In the formula, F represents the key indicator of the component, Wi represents the weight of the above i-th feature in the component, and x irepresents the i-th eigenvalue of the component, n represents the total number of components contained in the user preference graph, x in It represents the ith eigenvalue of the nth component. The formula calculates the weighted average of each feature after normalization. The larger the index F is, the greater the user's attention to the component. There are two ways to determine the key components based on the index F. One is to sort the components in descending order according to the index F and select the components with the highest number as the key components. The other is to determine the component as the key component when the index F exceeds a certain threshold. The second method is used here to determine the key components. The weight parameter Wi in formula (3) is obtained through the logistic regression model. The specific method is as follows:

[0117] 1) Whether a component is a key component is mapped to 1 and 0, where 1 indicates that the component is a key component and 0 indicates a non-key component.

[0118] 2) The probability of defining whether a component is a critical component is expressed by formula (4), where x i The i-th eigenvalue of the component representing the normalized component, y takes 0 or 1. When the probability P exceeds the threshold, it indicates that the component is a key component. The required parameter is W = {w1, w2, w3, w4}.

[0119]

[0120] 3) Use maximum likelihood estimation to find the parameter Wi in formula (4). Define the likelihood function as shown in formula (5).

[0121]

[0122] 4) The solution space of the likelihood function is obtained by inverting the likelihood function and using the stochastic gradient descent method to find the minimum value of the likelihood function, which is the parameter W.

[0123] To summarize, the steps for extracting key components are as follows:

[0124] 3) According to Figure 1 The user operation record in the grid is used to calculate the characteristic value of each component characteristic 1-4 in the grid graph.

[0125] 4) Use maximum likelihood estimation to obtain feature weights.

[0126] 3) Calculate the key indicators of each component using formula (3).

[0127] 4) Select components whose index values are greater than a specific threshold, which are the key components of the graph.

[0128] Key component combination

[0129] The extracted key components are represented by the attributes shown in Table 2.

[0130] Table 2 Key component attributes

[0131]

[0132] Based on the key component description information shown in Table 2 and the association rules obtained above, two rows of components in the antecedent and consequent parts of the association rules are laid out to generate a new graph.

[0133] Experimental verification;

[0134] This application uses historical user operation data from a web-based graphical browsing system in a power dispatching center's next-generation dispatching system for verification. System simulation experiments were conducted using a Windows 10 operating system, 8GB of RAM, an Intel Core i3-9100F CPU @ 3.60GHz, and an NVIDIA GeForce GTX 1650 GPU.

[0135] Data collection

[0136] A user browsing information tracking script was developed using JavaScript and embedded in the next-generation dispatching system's web-based graphical browsing system during its deployment. The script collected user browsing behavior data for power grid system diagrams (graphic file names ending in .sys.pic.g) from June 20, 2024, to July 20, 2024, and stored it in a Mongo database as semi-structured data. During this period, 176 power grid diagrams were collected, 477 user browsing transactions were recorded, and a total of 2,655 diagram visits were made.

[0137] We mapped the 176 graphs to serial numbers and manually labeled the serial numbers of key components within the graphs. In subsequent representations of browsing path transactions, frequent graphs, and association rules, serial numbers were used instead of specific graph names. The data is presented in Tables 3 and 4.

[0138] Table 3 Graphic number mapping (partial)

[0139]

[0140] Table 4 User browsing graphic path transactions (partial)

[0141]

[0142] Draw the graphic number and the number of visits into a line graph (such as Figure 3 ), it is observed that although there are many graphics involved in the dataset, there are relatively few graphics that are frequently accessed by users. Based on this, the support and confidence thresholds of the graphics can be estimated, which is convenient for subsequent experimental verification.

[0143] Data Modeling

[0144] User browsing interest modeling

[0145] According to formula (2) and formula (3), the graph dependency rules of frequent subgraphs and strong associations in browsing path transactions are calculated. Under different confidence and support thresholds, the number of graphs contained in the maximum frequent items generated is different. In order to provide users with a better graph browsing experience, the number of items in the maximum frequent item set mined should not be too large, and the number of frequent subgraphs in the rule antecedents and rule consequents in the extracted rules should not be too many. After multiple experiments, the confidence threshold is set to 55%, the support threshold is set to 65%, and the number of graphs contained in the maximum frequent item generated is 4.

[0146] Table 6 shows some of the mined rules. As can be seen, for Rule 2, after accessing the real-time monitoring overview, users will specifically check for power flow and voltage violations. For Rule 6, when users need to view the current load situation in the East China Power Grid, they will simultaneously view the load plan, load forecast, and power generation and transmission plan charts. These rules align with the expected user operational process.

[0147] Table 6 Strong association rules (partial)

[0148]

[0149] Key component extraction modeling

[0150] When using maximum likelihood estimation to determine the weights of component features, the stochastic gradient descent algorithm was adopted. In the test dataset, each component in 176 graphics was manually marked as a key component, of which 140 graphics were used as training sets and the remaining 36 graphics were used as test sets. Figure 4 The figure shows the number of iterations required for different step sizes when the loss accuracy is set to 0.001. As can be seen from the figure, when the step size is 0.3, the algorithm requires the lowest number of iterations. Finally, the feature weights in formula (3) are obtained, as shown in Table 7.

[0151] Table 7 Component feature weights

[0152]

[0153] According to the key component weights in Table 7, verification is performed on the test data set, with accuracy A, precision P, and recall R as verification indicators. The calculation methods are formula (6), formula (7), and formula (8), respectively. In the formula, D represents the total number of test data sets, TN represents the number of non-key components correctly predicted by the model, TP represents the number of key components correctly predicted by the model, FP represents the number of key components incorrectly predicted by the model, and FN represents the number of non-key components incorrectly predicted by the model.

[0154]

[0155] The evaluation metrics of the key component extraction model are shown in Table 8.

[0156] Table 8 Model evaluation indicators

[0157]

[0158] System Verification

[0159] Graphics are recommended to users based on the key component attributes (source) shown in Table 3. When a user browses to a graphic, the user is presented with a graphic from the combined generated graphics whose key component sources contain the currently browsed graphic. This recommendation system is implemented as a JavaScript plug-in within the next-generation scheduling system's web graphics browser. The effectiveness of the proposed graphics recommendation framework is determined by the average click-through rate (CTR) of recommended graphics over a period of time (the number of times a user clicks on a recommended graphic divided by the number of times a user visits the graphic). Figure 5 The average click-through rate of all graphics and the average click-through rate of access to the power balance graphics are shown every week for a consecutive month.

[0160] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. The power system graph recommendation system based on browsing path includes three modules: data acquisition module, graph mining module and component extraction and combination module. Its characteristics are: 1) The data acquisition module is used to collect and store a series of user operations on graphics. By embedding a script in the user's graphics browsing software, the user's operations on the graphics are recorded, including the name of the graphics clicked, the mouse position and the time, and these characteristic operations are quantified and stored in the database; 2) The graph mining module performs association rule analysis on the collected data and uses the Apriori principle to find the graphs that users frequently visit as the graphs of interest to users; 3) The Graphic Component Extraction and Combination Module uses collected data on user browsing behavior to calculate the probability that the component in the graphic of interest in step 2) is a key component. A higher probability indicates a greater likelihood of the component being a key component. Key components are extracted, representing the areas of interest to the user during browsing. These components are then reassembled into new graphics based on business needs and recommended to the user.

2. A recommendation method using the browsing path-based power system graphic recommendation system according to claim 1, characterized in that: The specific steps are as follows: 1) Data collection and processing; 2) Graphic browsing data collection; 3) Quantification of user browsing behavior; User browsing behaviors are divided into three categories: physiological behaviors, significant behaviors, and indirect behaviors. The collected user browsing data includes: browsing image name, browsing image duration, mouse events, keyboard events, and the data is stored in the corresponding data format. 4) User browsing path transaction extraction; A browse path transaction is defined as the sequence in which a user browses a set of graphics as follows: <...,(p i ,t i ),(p i+1 ,t i+1 ),...(p i+k ,t i+k ),...>; Two-tuple (p i ,t i ) represents t i Always browse graphics i , and the sequence satisfies the following two conditions: S1 satisfies t at any time t i< t i+1 ; S2t i+k+1 -t i+k >T,t i -t i-1 >T, T is the time threshold, or there is no n∈i.....i+k such that p n =p i+k+1 ; 5) Users prefer graph mining; The preferred path mining algorithm based on the Apriori principle is used, the support and confidence thresholds are set to a high value, and the page browsing time weight is added to the support calculation; In the user's browsing path, there are two indicators to evaluate whether the user cares about the graphics and whether there is a correlation between graphics: support and confidence. The support represents the probability of the graphics appearing in the browsing path, as shown in formula (1); Where count(A,B) represents the number of transactions in which both graph A and graph B appear in the browsing path transaction set, D represents the total number of all sample transactions, and the Time function calculates the duration of browsing the graph. A confidence index based on conditional probability, i.e., the degree of dependence between graphs, is introduced, as shown in formula (2). Where P(A,B) represents the probability that graph A and graph B appear simultaneously in a browsing path transaction, and P(A) represents the probability that graph A appears in a browsing path transaction. According to the conditional probability formula, a greater confidence level indicates a higher degree of association between graphs. A group of frequently accessed and associated graphs is considered to be the user's preferred graphs. Based on the above formulas and definitions, the mining process of user preference graphs is as follows: S1 calculates the support of each power grid graph based on the user's browsing path transaction set Tr = {tr1,tr2…trn}, tri = <…(pl,tl),(pl+1,tl+1),…>, and selects the graph set whose support is greater than a specific threshold, which we call the frequent 1 subgraph set U1 = {pk,pk+1,…}. S2 The frequent 1-subgraph U1 generated in the above step is combined with each other to calculate its support, and the frequent 2-subgraph set U2 is generated. k ,p k+1 ),(p k+3 ,p k+5 )…}. S3 repeats step 2) until no new frequent subgraph meets the support threshold, and obtains all frequent subgraph sets U1, U2, ..., Un; S4 Dependency Graph Mining: Perform association rule mining on the frequent subgraphs U2, ..., Un in sequence, select strong association rules with confidence greater than a preset threshold, and obtain a set of power grid graph browsing rules that the user is interested in and has strong dependency relationships. Each frequent graph in the rule is a user-preferred graph. 6) Extraction and combination of key graphics components; Extraction of key components; Mining the user's preferred graphics set, extracting the key components of each graphic, and judging whether the component is a key component based on the user's operation characteristics on the graphic component; The criterion for determining whether a graphic component is a key component is expressed by formula (3). In the formula, F represents the key indicator of the component, Wi represents the weight of the above i-th feature in the component, and x i represents the i-th eigenvalue of the component, n represents the total number of components contained in the user preference graph, x in It represents the ith eigenvalue of the nth component. The formula calculates the weighted average of each feature after normalization. The larger the index F is, the greater the user's attention to the component. The weight parameter Wi in formula (3) is obtained through the logistic regression model. The specific method is as follows: Whether the S1 component is a key component is mapped to 1 and 0, 1 indicates that the component is a key component, and 0 indicates a non-key component; The probability of whether a component is a critical component is defined by S2 and expressed by formula (4), where x i The i-th eigenvalue of the component representing the normalized component, y takes 0 or 1. When the probability P exceeds the threshold, it indicates that the component is a key component. The required parameter is W = {w1, w2, w3, w4}; S3 uses maximum likelihood estimation to find the parameter Wi in formula (4) and defines the likelihood function as shown in formula (5); S4 inverts the likelihood function and uses the stochastic gradient descent method to find the minimum value of the likelihood function, which is the parameter W. To summarize, the steps for extracting key components are as follows: 1) According to the user operation record, calculate the characteristic value of each component feature in the power grid graph. 2) Use maximum likelihood estimation to obtain feature weights; 3) Calculate the key indicators of each component using formula (3); 4) Select components whose index values are greater than a specific threshold, which are the key components of the graph. Key component combinations; Based on the key component description information and the association rules obtained previously, two rows of layouts are made with the components in the antecedent and consequent of the association rules to generate a new graph.

3. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2 is characterized in that: The data collection in step 1) includes embedding points in the graphic browsing software, capturing user operation events, and quantifying and storing user browsing paths.

4. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2 is characterized in that: The specific steps of step 2) collecting graphic browsing data are as follows: The graphic browsing software used by users adopts a B / S architecture and uses JavaScript script code to embed the tracking technology on the graphic browsing software client. Custom code is implanted into the graphic browsing software client to monitor and record the user's operation details on the browser, including keyboard input, mouse clicks, mouse scrolling and mouse movement, and output the time when the action occurs and the relative position on the power grid graphic canvas to the database. At the same time, the name and logo of each graphic browsed by the user are stored in the database.

5. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2 is characterized in that: The data format of step 3) is as follows, which is expressed in XML format, wherein the page node represents the total record of a browsing graphic data; the time node represents the duration of browsing the graphic; the start and end attributes represent the timestamps of starting and ending browsing; The operations node describes a series of user operations on the graph; the event node is the user event captured by the tracking script, which corresponds one-to-one to the events specified in the JavaScript language. The type attribute is the event type, the timestamp is the timestamp of the triggering event, and the x and y attributes represent the relative position of the mouse on the graph. The content of the event node indicates which component of the graph is triggered, specifically the component ID. Since each component in the CIM / G graphics specification must contain coordinates and ID, the specific x, y and component ID in the event node can be calculated by the JavaScript script and the CIM / G file.

6. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2, characterized in that: The component features extracted from the key components in step S6) include the following features 1 to 4, specifically as follows: Feature 1. Number of times a user has logged in to a component. Feature 2. The total duration of the user's mouse movement on the component; Feature 3. The total distance the wheel event scrolls up and down on the component; Feature 4. Evaluation speed of up and down scrolling events on the component.

7. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2, characterized in that: In the step S6), there are two methods for extracting key components by using the index F. One is to sort the components in descending order according to the index F and select the components at the top as key components. The other is to determine that the component is a key component when the index F exceeds a certain threshold.

8. The recommendation method of the power system graphic recommendation system based on browsing path according to claim 2 is characterized in that: The attributes of the key component combination in step S6) are as follows: Among them, title represents the title of the component, source represents the name of the component from the graphic, content represents the content of the component in the original CIM / G file, datasource represents the data source associated with the component, type represents the component type in the CIM / G specification, and target represents the key indicator.

Images