Linkage calculation method and system for combined power scenario semantic model
Through the interactive power scene semantic model, query record data is extracted, twin comparisons and business directional identification recognition are performed, special model combination scheme is retrieved, and computing performance is optimized, which solves the problem of low response speed of linkage calculations and achieves more efficient Resource configuration and response rate.
Patent Information
- Application Number
- CN202411218986.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2044-09-02
AI Technical Summary
In the prior art, the allocation of linkage computing resources is not in line with the actual situation, which leads to the problem of low response speed of linkage computing.
Through the interactive power scene semantic model, query record data is extracted, twin comparisons and business directional identification are performed, special model combination schemes are retrieved, fitness screening and timing analysis are performed, linkage optimization scheduling module is activated, and computational performance configuring results are optimized. , improve response speed.
The response rate of linked calculations in complex scenarios is improved, and the resource configuration of special models is optimized to meet actual needs and improve computing efficiency.
Smart Images
Figure CN118734004B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and in particular to a linkage calculation method and system for a combined power scenario semantic model. Background Art
[0002] The power system is a complex field, which is affected by various factors, including weather, load changes, market demand, etc. In order to analyze various power scenario data and semantic information, so as to model and understand the scenarios and scenes in the power field, a combined power scenario semantic model under various power scenarios is obtained.
[0003] The combined power scenario semantic model can provide more comprehensive power system information and context, while the specialized model can provide more in-depth technical details and analysis results. However, as the complexity of power scenarios increases, the response speed of specialized models cannot meet the needs because resource allocation is not closely aligned with the actual situation. Summary of the invention
[0004] The present application provides a linkage calculation method and system for a combined power scenario semantic model, which is used to solve the technical problem in the prior art that the linkage calculation resource allocation is not highly consistent with the actual situation, resulting in a low linkage calculation response speed.
[0005] In view of the above problems, the present application provides a linkage calculation method and system for a combined power scenario semantic model.
[0006] In a first aspect of the present application, a linkage calculation method using a power scenario semantic model is provided, the method comprising:
[0007] Interacting with the power scenario semantic model, extracting inquiry records in a preset inquiry window, and generating a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times;
[0008] Performing twin comparisons on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business-oriented identifiers;
[0009] Retrieving a set of multiple dedicated model combination schemes called by the multiple business orientation identifiers in historical time, wherein a dedicated model combination scheme includes a called dedicated model type, a called dedicated model quantity, and a computing performance allocation result;
[0010] Perform fitness screening on the plurality of dedicated model combination scheme sets respectively to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers;
[0011] Determining multiple earliest inquiry start times of the multiple inquiry record clusters according to the multiple inquiry start times, and performing linkage calculation timing analysis on the multiple inquiry record clusters according to the multiple earliest inquiry start times to obtain multiple inquiry record cluster processing time series;
[0012] Perform overlapping call time analysis of dedicated models of the multiple target dedicated model combination solutions according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, activate the linkage optimization scheduling module according to the analysis results, and optimize the calculation performance allocation results of the multiple target dedicated model combination solutions by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination solutions;
[0013] Based on the multiple optimization target dedicated model combination schemes, the multiple query record clusters are linked calculated to obtain multiple query record results.
[0014] A second aspect of the present application provides a linkage computing system using a power scenario semantic model, the system comprising:
[0015] An inquiry record data generating unit, used to interact with the power scenario semantic model, extract inquiry records in a preset inquiry window, and generate a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times;
[0016] An inquiry record cluster obtaining unit, configured to perform twin comparison on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business orientation identifiers;
[0017] A combination scheme set calling unit, used to call a plurality of special model combination scheme sets called by the plurality of business orientation identifiers in historical time, wherein a special model combination scheme includes a calling special model type, a calling special model quantity and a computing performance allocation result;
[0018] A target-specific model combination scheme obtaining unit is used to perform fitness screening on the plurality of special model combination scheme sets respectively to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers;
[0019] a processing time series obtaining unit, configured to determine a plurality of earliest inquiry start times of the plurality of inquiry record clusters according to the plurality of inquiry start times, and to perform linkage calculation time series analysis on the plurality of inquiry record clusters according to the plurality of earliest inquiry start times to obtain a plurality of inquiry record cluster processing time series;
[0020] An optimization scheme obtaining unit is used to perform overlapping call time analysis of the dedicated models of the multiple target dedicated model combination schemes according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, activate the linkage optimization scheduling module according to the analysis result, and optimize the computing performance allocation results of the multiple target dedicated model combination schemes by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination schemes;
[0021] The query record result cluster obtaining unit is used to perform linkage calculation on the multiple query record clusters based on the multiple optimization target dedicated model combination schemes to obtain multiple query record results.
[0022] One or more technical solutions provided in this application have at least the following technical effects or advantages:
[0023] The present application uses an interactive power scenario semantic model to extract inquiry records in a preset inquiry window and generate multiple inquiry record data, wherein the multiple inquiry record data include multiple inquiry dialogue information and multiple inquiry start times, and then respectively perform twin comparison on the multiple inquiry dialogue information to obtain multiple inquiry record clusters, wherein the multiple inquiry record clusters have multiple business orientation identifiers, and then call up multiple special model combination scheme sets called by the multiple business orientation identifiers in historical time, wherein a special model combination scheme includes calling a special model type, calling a special model number and a computing performance allocation result, and then respectively perform fitness screening on the multiple special model combination scheme sets to obtain multiple target special model combination schemes, wherein the multiple target special model combination schemes include multiple A linkage calculation time identifier is used to determine the multiple earliest query start times of multiple query record clusters according to the multiple query start times, and a linkage calculation timing analysis is performed on the multiple query record clusters according to the multiple earliest query start times to obtain multiple query record cluster processing time series, and then a dedicated model overlapping call time analysis of multiple target dedicated model combination schemes is performed according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, and a linkage optimization scheduling module is activated according to the analysis results, and the calculation performance allocation results of the multiple target dedicated model combination schemes are optimized by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination schemes, and then multiple query record clusters are linked to calculate based on the multiple optimized target dedicated model combination schemes to obtain multiple query record results. The technical effect of improving the response rate of linkage calculation in complex scenarios is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0025] Figure 1 A schematic diagram of a linkage calculation method for a power scenario semantic model provided in an embodiment of the present application;
[0026] Figure 2 A schematic diagram of the structure of a linkage computing system for a power scenario semantic model provided in an embodiment of the present application.
[0027] Explanation of the reference numerals: inquiry record data generating unit 11, inquiry record cluster obtaining unit 12, combination scheme set retrieving unit 13, target-specific model combination scheme obtaining unit 14, processing time series obtaining unit 15, optimization scheme obtaining unit 16, inquiry record result cluster obtaining unit 17. DETAILED DESCRIPTION
[0028] The present application provides a linkage calculation method and system for a combined power scenario semantic model, which is used to solve the technical problem in the prior art that the linkage calculation resource allocation is not highly consistent with the actual situation, resulting in a low linkage calculation response speed.
[0029] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0030] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or inherent to these processes, methods, products or devices.
[0031] Embodiment 1
[0032] like Figure 1 As shown, the present application provides a linkage calculation method for a power scenario semantic model, wherein the method includes:
[0033] S100: interacting with the power scenario semantic model, extracting inquiry records in a preset inquiry window, and generating a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times;
[0034] In a possible embodiment, the power scenario semantic model is a semantic model that analyzes and understands text information related to the power system, such as reports, news, forum discussions, and other data, and understands questions raised by users and gives corresponding answers. Optionally, the power scenario semantic model is a model built based on natural language processing technology. Preferably, the power scenario semantic model can understand relevant inquiries about the power system, and the scope of the relevant inquiries is not limited, including inquiries such as service progress, electricity consumption and electricity charges, and power outage maintenance plans.
[0035] The query record of the power scenario semantic model is extracted in the preset query window to obtain the multiple query record data, so as to collect the questions that the user wants to know. Among them, the preset query window is an inquiry time period set by technical personnel in this field, which can be within 3 hours, within one day, within three days, etc. The multiple query record data are used to describe the questions raised by the user, including multiple query dialogue information and multiple query start times. The multiple query dialogue information is used to describe the questions raised by multiple users and the corresponding replies of the power scenario semantic model. The multiple query start times are used to describe the time points when multiple users start to make inquiries.
[0036] By obtaining the multiple inquiry record data, the technical effect of providing basic analysis data for subsequent reliable calculation of calling a dedicated model and providing users with reliable inquiry record results is achieved.
[0037] S200: performing twin comparison on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business orientation identifiers;
[0038] Furthermore, step S200 of the embodiment of the present application further includes:
[0039] Performing keyword vector feature analysis on the multiple query dialogue information to obtain multiple query word vector feature sets;
[0040] Randomly extract a query word vector feature set from the multiple query word vector feature sets as a first benchmark query word vector feature set;
[0041] The multiple query word vector feature sets are traversed to perform twin comparison with the first benchmark query word vector feature set, and multiple query word vector feature sets whose comparison results meet a preset twin similarity threshold are divided into a first query record cluster.
[0042] Furthermore, step S200 of the embodiment of the present application further includes:
[0043] Eliminate the first query record cluster from the multiple query word vector feature sets, and then randomly select a query word vector feature set from the multiple query word vector feature sets again as the second benchmark query word vector feature set;
[0044] Traversing the multiple query word vector feature sets and performing twin comparison with the second benchmark query word vector feature set, and dividing the multiple query word vector feature sets whose comparison results meet the preset twin similarity threshold into the second query record cluster;
[0045] After multiple twin comparisons, multiple query record clusters are obtained;
[0046] The word vector features in the plurality of query record clusters are identified using a business feature identifier to obtain the plurality of business directionality identifiers.
[0047] In one embodiment, the question types corresponding to multiple inquiry dialogue information of multiple users in the preset inquiry window may be repeated. In order to save computing power and avoid repeated calculations, the same type of inquiry dialogue information is classified into one category by performing twin comparison on multiple inquiry dialogue information, thereby obtaining multiple inquiry record clusters. And use business feature recognition to identify multiple inquiry word vector feature sets of multiple inquiry record clusters, determine the question types corresponding to multiple inquiry record clusters, such as power query, maintenance time query, meter replacement progress query, load query and other business types, and determine the multiple business directional identifiers according to the recognition results. By obtaining the multiple business directional identifiers, the technical effect of providing a basis for subsequent calls to the corresponding dedicated models is achieved.
[0048] In one possible embodiment, a suitable keyword extraction algorithm is used to extract representative keywords from multiple query dialogue information to obtain multiple keyword sets, such as the TF-IDF algorithm, the TextRank extraction algorithm, etc.; then, the Word2Vec word embedding technology is used to map the extracted multiple keyword sets to a high-dimensional vector space to obtain multiple keyword vector sets, and the multiple keyword vectors are subjected to mode extraction and mean processing to obtain multiple query word vector feature sets. The goal of extracting and analyzing the keyword vectors in multiple query dialogue information and performing feature analysis is achieved.
[0049] A query word vector feature set is randomly selected from the multiple query word vector feature sets as a first benchmark query word vector feature set, and the first benchmark query word vector feature set is used as a benchmark for binary classification of the multiple query word vector feature sets.
[0050] Traverse the multiple query word vector feature sets and perform twin comparison with the first benchmark query word vector feature set. Optionally, calculate the cosine similarity of the multiple query word vector feature sets and the first benchmark query word vector feature set respectively, use the calculation result as the comparison result, and judge whether the comparison result is greater than or equal to the preset twin similarity threshold. If so, divide the corresponding multiple query word vector feature sets into the first query record cluster. Among them, the first query record cluster is a cluster of query word vector feature sets with high similarity to the first benchmark query word vector feature set. The preset twin similarity threshold is the minimum cosine similarity between the query word vector feature set and the first benchmark query word vector feature set when the two are similar, which is pre-set by a technician in this field.
[0051] Furthermore, the first query record cluster is removed from the multiple query word vector feature sets, and then a query word vector feature set is randomly selected from the multiple query word vector feature sets again as the second benchmark query word vector feature set.
[0052] Based on the same obtaining principle as the first query record cluster, traverse the multiple query word vector feature sets and perform twin comparison with the second benchmark query word vector feature set, calculate the cosine similarity between the multiple query word vector feature sets after eliminating the first query record cluster and the second benchmark query word vector feature set, obtain the comparison result, and then determine whether the comparison result is greater than or equal to the preset twin similarity threshold. If so, divide the corresponding multiple query word vector feature sets into the second query record cluster.
[0053] Based on the same acquisition principle, after multiple twin comparisons, multiple query record clusters are obtained. Then, the word vector features in the multiple query record clusters are identified using a business feature identifier to determine the business type corresponding to each query record cluster, and then the multiple business directionality identifiers are obtained.
[0054] Optionally, by obtaining multiple sample query word vector feature sets and multiple sample business orientation identifiers as training data, the training data is divided into n groups of training data, and the model framework constructed based on the convolutional neural network is trained in turn using the n groups of training data to learn the mapping relationship between the query word vector feature set and the business orientation identifier, and the output accuracy of the previous group of training data is used to update and adjust the network parameters of the model framework until the output converges, that is, the output accuracy meets the requirements, and the trained business feature identifier is obtained. Then, the word vector features with the top k occurrence frequencies in the multiple query record clusters are respectively input into the business feature identifier to obtain multiple business orientation identifiers. The technical effect of quickly, efficiently and accurately identifying the business types of multiple query record clusters is achieved.
[0055] S300: Retrieve a set of multiple dedicated model combination schemes called by the multiple business orientation identifiers in historical time, wherein a dedicated model combination scheme includes a called dedicated model type, a called dedicated model quantity, and a computing performance allocation result;
[0056] In one embodiment, the multiple business orientation identifiers are used as indexes to search in the historical call database to obtain multiple sets of special model combination schemes. The historical call database is a record data repository of special model combinations called by different business orientation identifiers in a historical period. A special model combination scheme includes calling a special model type, calling a special model number and computing performance allocation results. The special model is a machine learning model that calculates specific tasks in the power system.
[0057] Exemplarily, when the business directionality is identified as maintenance progress query, the number of special models called in the corresponding special model combination scheme is 3, and the types of special models called include fault monitoring special models, maintenance progress prediction special models and maintenance plan optimization special models. And the computing performance allocation results in the corresponding special model combination scheme include the computing power resources corresponding to the fault monitoring special model, the maintenance progress prediction special model and the maintenance plan optimization special model respectively. Among them, the fault monitoring special model is used to locate and monitor the fault location in the power system. The maintenance progress prediction special model is used to perform progress prediction based on the maintenance optimization plan output by the maintenance plan optimization special model, combined with the current maintenance project. The maintenance plan optimization special model is used to optimize the corresponding maintenance plan according to the fault location result output by the fault monitoring special model.
[0058] S400: performing fitness screening on the plurality of dedicated model combination scheme sets respectively to obtain a plurality of target dedicated model combination schemes, wherein the plurality of target dedicated model combination schemes include a plurality of linkage calculation time identifiers;
[0059] Further, the plurality of dedicated model combination scheme sets are respectively subjected to fitness screening to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers. Step S400 of the embodiment of the present application further includes:
[0060] Selecting a first dedicated model combination solution set from the plurality of dedicated model combination solution sets;
[0061] Extracting linkage calculation times and response accuracies of a plurality of first special model combination schemes in the first special model combination scheme set to obtain a plurality of first linkage calculation times and a plurality of first response accuracies;
[0062] Calculating the inverse of the ratio of the plurality of first linkage calculation times to the sum of the plurality of first linkage calculation times to obtain a plurality of first sub-fitnesses;
[0063] Calculating a ratio of the plurality of first response accuracies to a sum of the plurality of first response accuracies to obtain a plurality of second sub-fitnesses;
[0064] Performing weighted calculation on the multiple first sub-fitnesses and the multiple second sub-fitnesses according to a preset weight ratio to obtain multiple first fitnesses;
[0065] Using the first dedicated model combination scheme corresponding to the maximum value of the plurality of first fitnesses as the first target dedicated model combination scheme, wherein the first target dedicated model combination scheme includes a first linkage calculation time identifier;
[0066] Fitness calculation and screening are performed on the multiple dedicated model combination scheme sets to obtain multiple target dedicated model combination schemes.
[0067] In one embodiment, after obtaining the plurality of sets of special model combination solutions, since the plurality of sets of special model combination solutions are solutions that have been practiced in the historical time, the linkage calculation time and response accuracy of each special model combination solution in the practice process are extracted, and fitness calculation is performed according to the extraction results to provide a basis for fitness screening, and the plurality of target special model combination solutions are obtained by taking the special model combination solution corresponding to the maximum fitness value in each set of special model combination solutions as the target special model combination solution. The plurality of target special model combination solutions include a plurality of linkage calculation time identifiers.
[0068] Preferably, the linkage calculation time is the total time required for each special model combination scheme to perform linkage calculation. The linkage calculation time is negatively correlated with fitness. The response accuracy is the accuracy of the output result after each special model combination scheme performs linkage calculation. The response accuracy is positively correlated with fitness.
[0069] In one embodiment, a first special model combination scheme set is selected from the multiple special model combination scheme sets, and the first special model combination scheme set is screened for fitness to obtain a first target special model combination scheme. The fitness screening step is as follows: extract the linkage calculation time and response accuracy of multiple first special model combination schemes in the first special model combination scheme set to obtain multiple first linkage calculation times and multiple first response accuracies. Then, the inverse of the ratio of the multiple first linkage calculation times to the sum of the multiple first linkage calculation times is calculated respectively to obtain multiple first sub-fitnesses. Among them, the multiple first sub-fitnesses determine the advantages and disadvantages of multiple first special model combination schemes from the dimension of linkage calculation time.
[0070] Then, the ratio of the multiple first response accuracies to the sum of the multiple first response accuracies is calculated to obtain multiple second sub-fitnesses. The multiple second sub-fitnesses determine the pros and cons of the multiple first special model combination schemes from the dimension of response accuracy. The multiple first sub-fitnesses are obtained by weighted calculation of the multiple first sub-fitnesses and the multiple second sub-fitnesses according to a preset weight ratio, wherein the preset weight ratio is a weight ratio set by a technician in this field according to actual conditions.
[0071] Optionally, a first dedicated model combination scheme corresponding to the maximum value of the plurality of first fitnesses is used as a first target dedicated model combination scheme, wherein the first target dedicated model combination scheme includes a first linkage calculation time identifier. Based on the same acquisition principle as the first target dedicated model combination scheme, the plurality of dedicated model combination scheme sets are subjected to fitness calculation screening to obtain a plurality of target dedicated model combination schemes.
[0072] S500: determining multiple earliest inquiry start times of the multiple inquiry record clusters according to the multiple inquiry start times, and performing linkage calculation timing analysis on the multiple inquiry record clusters according to the multiple earliest inquiry start times to obtain multiple inquiry record cluster processing time series;
[0073] In one embodiment, according to the sequence of the multiple query start times, the time point at which the first query started in each query record cluster occurs is determined, and the multiple first query start times of the multiple query record clusters are obtained. Since different query record clusters may overlap when calling a dedicated model, if the computing performance resources allocated to the dedicated model are insufficient, the query response speed may not meet expectations. Therefore, according to the multiple first query start times, the multiple query record clusters are sorted in order from front to back in time, so as to obtain the processing time series of the multiple query record clusters. Since the query records corresponding to each query record cluster are the same type of business, when the query record corresponding to the first query start time obtains the linkage calculation result, it can be synchronized to the subsequent query records in each query record cluster. Therefore, the multiple first query start times are used as analysis data.
[0074] By obtaining the processing time series of the plurality of query record clusters, data support is provided for subsequent analysis of whether overlap occurs in dedicated model calls, and further optimization of computing performance allocation results is performed according to the analysis results.
[0075] S600: performing overlapping call time analysis of dedicated models of the multiple target dedicated model combination solutions according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time sequences, activating a linkage optimization scheduling module according to the analysis result, optimizing the calculation performance allocation results of the multiple target dedicated model combination solutions by using the linkage optimization scheduling module, and obtaining multiple optimized target dedicated model combination solutions;
[0076] S700: performing linkage calculations on the multiple query record clusters based on the multiple optimization target dedicated model combination solutions to obtain multiple query record results.
[0077] Further, according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, the dedicated model overlapping call time analysis of the multiple target dedicated model combination solutions is performed, and the linkage optimization scheduling module is activated according to the analysis results. The computing performance allocation results of the multiple target dedicated model combination solutions are optimized by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination solutions. The step S600 of the embodiment of the present application also includes:
[0078] Using a dedicated model as an index, the multiple target dedicated model combination schemes are retrieved to obtain K dedicated model call data, wherein the K dedicated model call data include K model call times and K model call scheme sets, the K model call scheme sets are a set consisting of multiple target dedicated model combination schemes that respectively call K dedicated models for linkage calculation, K is the total number of dedicated model types called in the multiple target dedicated model combination schemes, and K is a positive integer greater than or equal to 2;
[0079] Determine whether the number of calls of K models is greater than 1. If so, add the corresponding K model call solution sets into the overlapping model call solution cluster, where the overlapping model call solution cluster has M model call solution sets, where M is the number of special models with a call number greater than 1, and is a positive integer less than or equal to K;
[0080] According to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, the dedicated model overlapping call time analysis is performed on the M model call scheme sets respectively to obtain M overlapping call time interval sets, wherein the M overlapping call time interval sets have M dedicated model type identifiers.
[0081] Furthermore, step S600 in the embodiment of the present application further includes:
[0082] Extracting a first model calling solution set from the M model calling solution sets;
[0083] Determining, according to the multiple linkage calculation time identifiers, multiple start calculation times of multiple target-specific model combination solutions in the first model calling solution set;
[0084] Sort the multiple start calculation times in ascending order to obtain a first start calculation time sequence;
[0085] According to the first start calculation time sequence, in combination with the multiple linkage calculation time identifiers, multiple completion calculation times are determined, and it is judged whether the multiple completion calculation times are greater than the subsequent multiple first start calculation times, and if so, a first overlapping time interval set is counted;
[0086] A dedicated model overlapping call time analysis is performed according to the M model calling scheme sets to obtain M overlapping call time interval sets.
[0087] Furthermore, step S600 in the embodiment of the present application further includes:
[0088] Inputting the M overlapping call time interval sets and the M special model type identifiers into the linkage optimization scheduling module to generate M computing performance allocation scheme sets, wherein each computing performance allocation scheme corresponds to an overlapping call time interval;
[0089] According to the M computing performance allocation scheme sets, the M overlapping call time interval sets and the M special model type identifiers, the computing performance allocation results of the multiple target special model combination schemes are optimized and updated for the computing performance allocation of the special models in the corresponding time intervals to obtain multiple optimized target special model combination schemes.
[0090] In a possible embodiment, the multiple linkage calculation time identifiers are used to describe the length of time required for linkage calculation of multiple target-specific model combination schemes. The multiple query record cluster processing time series is the start time series of special model calls according to the multiple target-specific model combination schemes. The linkage optimization scheduling module is used to optimize the computing performance allocation results of the multiple target-specific model combination schemes, optimize the computing performance of different special models in different overlapping call time intervals, and then obtain the multiple optimized target-specific model combination schemes. The technical effect of optimizing the computing performance of the linkage calculation process according to actual computing needs and improving computing reliability and response rate is achieved.
[0091] Then, according to the plurality of optimization target dedicated model combination schemes, the query dialogue information corresponding to the plurality of query record clusters is respectively calculated in linkage, and according to the calling dedicated model type and the calling dedicated model quantity in the optimization target dedicated model combination scheme, and the optimized computing performance allocation result, the plurality of query word vector feature sets corresponding to the plurality of first query start times in the plurality of query record clusters are respectively identified to obtain the plurality of query record results. Each query record cluster corresponds to one query record result.
[0092] In one embodiment, the multiple target-specific model combination schemes are retrieved with the dedicated model as the index to obtain K dedicated model calling data, wherein the K dedicated model calling data include K model calling times and K model calling scheme sets, the K model calling scheme sets are a set of multiple target-specific model combination schemes that respectively call K dedicated models for linkage calculation, K is the total number of dedicated model types called in the multiple target-specific model combination schemes, and K is a positive integer greater than or equal to 2.
[0093] Then, determine whether the number of calls of the K models is greater than 1. If so, it indicates that the corresponding dedicated model is called multiple times, then add the corresponding K model calling solution sets into the overlapping model calling solution cluster, the overlapping model calling solution cluster has M model calling solution sets, M is the number of dedicated models with a call count greater than 1, and is a positive integer less than or equal to K.
[0094] Then, according to the multiple linkage computing time identifiers and the multiple query record cluster processing time series, the M model call scheme sets are analyzed to determine whether the M multiple call dedicated models corresponding to the overlapping model call scheme clusters have overlapping call times, and if so, obtain M overlapping call time interval sets. The M overlapping call time interval sets are sets of multiple overlapping time intervals when the M multiple call dedicated models participate in different linkage computing schemes.
[0095] According to the multiple linkage calculation time identifiers and the multiple query record cluster processing time sequences, the steps of performing dedicated model overlapping call time analysis on the M model call solution sets are as follows: extracting a first model call solution set from the M model call solution sets, and then determining multiple start calculation times of multiple target dedicated model combination solutions in the first model call solution set according to the multiple linkage calculation time identifiers, sorting the multiple start calculation times in ascending order to obtain a first start calculation time sequence, and determining multiple completion calculation times according to the first start calculation time sequence and combining the multiple linkage calculation time identifiers in sequence. That is, the multiple start calculation times in the first start calculation time sequence are added to the linkage calculation duration described by the corresponding multiple linkage calculation time identifiers to obtain multiple completion calculation times. The multiple completion calculation times reflect the end time of each query dialogue linkage calculation when the dedicated model corresponding to the first model call solution set completes the linkage calculation of multiple query dialogue information.
[0096] And determine whether the multiple completion calculation times are greater than the subsequent multiple first start calculation times. If so, it indicates that when calling the dedicated model to perform linkage calculation on the two query dialogue information, the model call time overlaps. The time period between the multiple completion calculation times and the subsequent multiple first start calculation times is respectively used as the first overlapping time interval to obtain the first overlapping time interval set. Based on the same acquisition principle as the first overlapping time interval set, the dedicated model overlapping call time analysis is performed according to the M model call scheme sets to obtain the M overlapping call time interval sets.
[0097] Preferably, the M overlapping call time interval sets are input into the linkage optimization scheduling module to generate M computing performance allocation scheme sets, wherein each computing performance allocation scheme corresponds to an overlapping call time interval. Each computing performance allocation scheme includes the computing performance resources required for a dedicated model in an overlapping call time interval. Optionally, the computing performance allocation scheme is the sum of multiple computing performance allocation results required for the dedicated model in the overlapping call time interval to participate in different linkage calculations.
[0098] By obtaining multiple sample overlapping call time intervals, multiple sample dedicated model type identifiers and multiple sample computing performance allocation plans as training data, supervised training is performed on the network layer built based on the convolutional neural network, and the mapping relationship between the overlapping call time interval-dedicated model type identifier-computing performance allocation plan is learned until the output converges, thereby obtaining the trained linkage optimization scheduling module.
[0099] According to the M computing performance allocation scheme sets and the M overlapping call time interval sets, the computing performance allocation results of the multiple target dedicated model combination schemes are optimized and updated in the corresponding time intervals. That is to say, the computing performance allocation results of different dedicated models in the multiple target dedicated model combination schemes are updated in the corresponding overlapping call time intervals according to the computing performance allocation schemes corresponding to the overlapping call time intervals, and the resources of the dedicated models are allocated according to the computing performance allocation schemes corresponding to the overlapping call time intervals, thereby obtaining multiple optimized target dedicated model combination schemes. The technical effect of allocating the resources of the dedicated models participating in the linkage calculation in accordance with the actual situation and improving the calculation efficiency is achieved.
[0100] In summary, the embodiments of the present application have at least the following technical effects:
[0101] This application conducts in-depth analysis of multiple query record data in the preset query window of the power scenario semantic model, obtains multiple query record clusters through twin comparison, performs preliminary processing on the query records, and achieves the goal of improving processing efficiency. Then, according to the multiple business orientation identifiers corresponding to the multiple query record clusters, multiple target special model combination schemes that meet the conditions of the multiple query record clusters are determined, and the types and quantities of special models that need to be called and the computing performance required for each special model are determined. Then, the overlapping time intervals of special model calls are analyzed for multiple query record clusters in chronological order, effectively solving the problem of reduced operation response speed due to the participation of special models in multiple linkage calculations at the same time. The linkage optimization scheduling module is used to optimize and update the computing performance of special models with overlapping call times in the overlapping call time interval, and multiple optimization target special model combination schemes are obtained. Then, based on the multiple optimization target special model combination schemes, multiple query record clusters are linked to calculate and obtain multiple query record results. The technical effect of improving the response rate of linkage calculations in complex scenarios and allocating resources according to the actual linkage calculation situation of special models is achieved.
[0102] Embodiment 2
[0103] Based on the same inventive concept as the linkage calculation method using the power scenario semantic model in the aforementioned embodiment, Figure 2 As shown, the present application provides a linkage computing system for a power scenario semantic model, and the system and method embodiments in the present application are based on the same inventive concept. The system includes:
[0104] The inquiry record data generating unit 11 is used to interact with the power scenario semantic model, extract the inquiry records in the preset inquiry window, and generate a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times;
[0105] The inquiry record cluster obtaining unit 12 is used to perform twin comparison on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business orientation identifiers;
[0106] The combination scheme set calling unit 13 is used to call a plurality of special model combination scheme sets called by the plurality of business directionality identifiers in the historical time, wherein a special model combination scheme includes a calling special model type, a calling special model quantity and a computing performance allocation result;
[0107] A target-specific model combination scheme obtaining unit 14 is used to perform fitness screening on the plurality of dedicated model combination scheme sets respectively to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers;
[0108] A processing time series obtaining unit 15 is used to determine a plurality of earliest inquiry start times of the plurality of inquiry record clusters according to the plurality of inquiry start times, and to perform linkage calculation time series analysis on the plurality of inquiry record clusters according to the plurality of earliest inquiry start times to obtain a plurality of inquiry record cluster processing time series;
[0109] The optimization scheme obtaining unit 16 is used to analyze the overlapping calling time of the dedicated models of the multiple target dedicated model combination schemes according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, activate the linkage optimization scheduling module according to the analysis result, and optimize the computing performance allocation results of the multiple target dedicated model combination schemes by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination schemes;
[0110] The query record result cluster obtaining unit 17 is used to perform linkage calculation on the multiple query record clusters based on the multiple optimization target dedicated model combination schemes to obtain multiple query record results.
[0111] Furthermore, the system also includes:
[0112] Performing keyword vector feature analysis on the multiple query dialogue information to obtain multiple query word vector feature sets;
[0113] Randomly extract a query word vector feature set from the multiple query word vector feature sets as a first benchmark query word vector feature set;
[0114] The multiple query word vector feature sets are traversed to perform twin comparison with the first benchmark query word vector feature set, and multiple query word vector feature sets whose comparison results meet a preset twin similarity threshold are divided into a first query record cluster.
[0115] Furthermore, the system also includes:
[0116] Eliminate the first query record cluster from the multiple query word vector feature sets, and then randomly select a query word vector feature set from the multiple query word vector feature sets again as the second benchmark query word vector feature set;
[0117] Traversing the multiple query word vector feature sets and performing twin comparison with the second benchmark query word vector feature set, and dividing the multiple query word vector feature sets whose comparison results meet the preset twin similarity threshold into the second query record cluster;
[0118] After multiple twin comparisons, multiple query record clusters are obtained;
[0119] The word vector features in the plurality of query record clusters are identified using a business feature identifier to obtain the plurality of business directionality identifiers.
[0120] Furthermore, the system also includes:
[0121] Selecting a first dedicated model combination solution set from the plurality of dedicated model combination solution sets;
[0122] Extracting linkage calculation times and response accuracies of a plurality of first special model combination schemes in the first special model combination scheme set to obtain a plurality of first linkage calculation times and a plurality of first response accuracies;
[0123] Calculating the inverse of the ratio of the plurality of first linkage calculation times to the sum of the plurality of first linkage calculation times to obtain a plurality of first sub-fitnesses;
[0124] Calculating a ratio of the plurality of first response accuracies to a sum of the plurality of first response accuracies to obtain a plurality of second sub-fitnesses;
[0125] Performing weighted calculation on the multiple first sub-fitnesses and the multiple second sub-fitnesses according to a preset weight ratio to obtain multiple first fitnesses;
[0126] Using the first dedicated model combination scheme corresponding to the maximum value of the plurality of first fitnesses as the first target dedicated model combination scheme, wherein the first target dedicated model combination scheme includes a first linkage calculation time identifier;
[0127] Fitness calculation and screening are performed on the multiple dedicated model combination scheme sets to obtain multiple target dedicated model combination schemes.
[0128] Furthermore, the system also includes:
[0129] Using a dedicated model as an index, the multiple target dedicated model combination schemes are retrieved to obtain K dedicated model call data, wherein the K dedicated model call data include K model call times and K model call scheme sets, the K model call scheme sets are a set consisting of multiple target dedicated model combination schemes that respectively call K dedicated models for linkage calculation, K is the total number of dedicated model types called in the multiple target dedicated model combination schemes, and K is a positive integer greater than or equal to 2;
[0130] Determine whether the number of calls of K models is greater than 1. If so, add the corresponding K model call solution sets into the overlapping model call solution cluster, where the overlapping model call solution cluster has M model call solution sets, where M is the number of special models with a call number greater than 1, and is a positive integer less than or equal to K;
[0131] According to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, dedicated model overlapping call time analysis is performed on the M model call solution sets respectively to obtain M overlapping call time interval sets.
[0132] Furthermore, the system also includes:
[0133] Extracting a first model calling solution set from the M model calling solution sets;
[0134] Determining, according to the multiple linkage calculation time identifiers, multiple start calculation times of multiple target-specific model combination solutions in the first model calling solution set;
[0135] Sort the multiple start calculation times in ascending order to obtain a first start calculation time sequence;
[0136] According to the first start calculation time sequence, in combination with the multiple linkage calculation time identifiers, multiple completion calculation times are determined, and it is judged whether the multiple completion calculation times are greater than the subsequent multiple first start calculation times, and if so, a first overlapping time interval set is counted;
[0137] A dedicated model overlapping call time analysis is performed according to the M model call scheme sets to obtain M overlapping call time interval sets, wherein the M overlapping call time interval sets have M dedicated model type identifiers.
[0138] Furthermore, the system also includes:
[0139] Inputting the M overlapping call time interval sets and the M special model type identifiers into the linkage optimization scheduling module to generate M computing performance allocation scheme sets, wherein each computing performance allocation scheme corresponds to an overlapping call time interval;
[0140] According to the M computing performance allocation scheme sets, the M overlapping call time interval sets and the M special model type identifiers, the computing performance allocation results of the multiple target special model combination schemes are optimized and updated for the computing performance allocation of the special models in the corresponding time intervals to obtain multiple optimized target special model combination schemes.
[0141] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. Other embodiments are within the scope of the attached claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0142] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.
[0143] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.
Claims
1. A linkage calculation method using a power scenario semantic model, characterized in that: The method comprises: Interacting with the power scenario semantic model, extracting inquiry records in a preset inquiry window, and generating a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times; Performing twin comparisons on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business-oriented identifiers; Retrieving a set of multiple dedicated model combination schemes called by the multiple business orientation identifiers in historical time, wherein a dedicated model combination scheme includes a called dedicated model type, a called dedicated model quantity, and a computing performance allocation result; Perform fitness screening on the plurality of dedicated model combination scheme sets respectively to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers; Determining multiple earliest inquiry start times of the multiple inquiry record clusters according to the multiple inquiry start times, and performing linkage calculation timing analysis on the multiple inquiry record clusters according to the multiple earliest inquiry start times to obtain multiple inquiry record cluster processing time series; Perform overlapping call time analysis of dedicated models of the multiple target dedicated model combination solutions according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, activate the linkage optimization scheduling module according to the analysis results, and optimize the calculation performance allocation results of the multiple target dedicated model combination solutions by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination solutions; Based on the multiple optimization target dedicated model combination schemes, the multiple query record clusters are linked calculated to obtain multiple query record results.
2. The method according to claim 1, characterized in that The method comprises: Performing keyword vector feature analysis on the multiple query dialogue information to obtain multiple query word vector feature sets; Randomly extract a query word vector feature set from the multiple query word vector feature sets as a first benchmark query word vector feature set; The multiple query word vector feature sets are traversed to perform twin comparison with the first benchmark query word vector feature set, and multiple query word vector feature sets whose comparison results meet a preset twin similarity threshold are divided into a first query record cluster.
3. The method according to claim 2, characterized in that The method comprises: Eliminate the first query record cluster from the multiple query word vector feature sets, and then randomly select a query word vector feature set from the multiple query word vector feature sets again as the second benchmark query word vector feature set; Traversing the multiple query word vector feature sets and performing twin comparison with the second benchmark query word vector feature set, and dividing the multiple query word vector feature sets whose comparison results meet the preset twin similarity threshold into the second query record cluster; After multiple twin comparisons, multiple query record clusters are obtained; The word vector features in the plurality of query record clusters are identified using a business feature identifier to obtain the plurality of business directionality identifiers.
4. The method according to claim 1, characterized in that The plurality of dedicated model combination scheme sets are respectively subjected to fitness screening to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers, and the method includes: Selecting a first dedicated model combination solution set from the plurality of dedicated model combination solution sets; Extracting linkage calculation times and response accuracies of a plurality of first special model combination schemes in the first special model combination scheme set to obtain a plurality of first linkage calculation times and a plurality of first response accuracies; Calculating the inverse of the ratio of the plurality of first linkage calculation times to the sum of the plurality of first linkage calculation times to obtain a plurality of first sub-fitnesses; Calculating a ratio of the plurality of first response accuracies to a sum of the plurality of first response accuracies to obtain a plurality of second sub-fitnesses; Performing weighted calculation on the multiple first sub-fitnesses and the multiple second sub-fitnesses according to a preset weight ratio to obtain multiple first fitnesses; Using the first dedicated model combination scheme corresponding to the maximum value of the plurality of first fitnesses as the first target dedicated model combination scheme, wherein the first target dedicated model combination scheme includes a first linkage calculation time identifier; Fitness calculation and screening are performed on the multiple dedicated model combination scheme sets to obtain multiple target dedicated model combination schemes.
5. The method according to claim 1, characterized in that According to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, the dedicated model overlapping call time analysis of the multiple target dedicated model combination solutions is performed, and the linkage optimization scheduling module is activated according to the analysis result. The calculation performance allocation results of the multiple target dedicated model combination solutions are optimized by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination solutions. The method includes: Using a dedicated model as an index, the multiple target dedicated model combination schemes are retrieved to obtain K dedicated model call data, wherein the K dedicated model call data include K model call times and K model call scheme sets, the K model call scheme sets are a set consisting of multiple target dedicated model combination schemes that respectively call K dedicated models for linkage calculation, K is the total number of dedicated model types called in the multiple target dedicated model combination schemes, and K is a positive integer greater than or equal to 2; Determine whether the number of calls of K models is greater than 1. If so, add the corresponding K model call solution sets into the overlapping model call solution cluster, where the overlapping model call solution cluster has M model call solution sets, where M is the number of special models with a call number greater than 1, and is a positive integer less than or equal to K; According to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, dedicated model overlapping call time analysis is performed on the M model call solution sets respectively to obtain M overlapping call time interval sets.
6. The method according to claim 5, characterized in that The method comprises: Extracting a first model calling solution set from the M model calling solution sets; Determining, according to the multiple linkage calculation time identifiers, multiple start calculation times of multiple target-specific model combination solutions in the first model call solution set; Sort the multiple start calculation times in ascending order to obtain a first start calculation time sequence; According to the first start calculation time sequence, in combination with the multiple linkage calculation time identifiers, multiple completion calculation times are determined, and it is judged whether the multiple completion calculation times are greater than the subsequent multiple first start calculation times, and if so, a first overlapping time interval set is counted; A dedicated model overlapping call time analysis is performed according to the M model call scheme sets to obtain M overlapping call time interval sets, wherein the M overlapping call time interval sets have M dedicated model type identifiers.
7. The method according to claim 6, characterized in that The method comprises: Inputting the M overlapping call time interval sets and the M dedicated model type identifiers into the linkage optimization scheduling module to generate M computing performance allocation scheme sets, wherein each computing performance allocation scheme corresponds to an overlapping call time interval; According to the M computing performance allocation scheme sets, the M overlapping call time interval sets and the M special model type identifiers, the computing performance allocation results of the multiple target special model combination schemes are optimized and updated in the corresponding time intervals to obtain multiple optimized target special model combination schemes.
8. A linkage computing system using a power scenario semantic model, characterized in that the system comprises: An inquiry record data generating unit, used to interact with the power scenario semantic model, extract inquiry records in a preset inquiry window, and generate a plurality of inquiry record data, wherein the plurality of inquiry record data includes a plurality of inquiry dialogue information and a plurality of inquiry start times; An inquiry record cluster obtaining unit, configured to perform twin comparison on the plurality of inquiry dialogue information respectively to obtain a plurality of inquiry record clusters, wherein the plurality of inquiry record clusters have a plurality of business orientation identifiers; A combination scheme set calling unit, used to call a plurality of special model combination scheme sets called by the plurality of business orientation identifiers in historical time, wherein a special model combination scheme includes a calling special model type, a calling special model quantity and a computing performance allocation result; A target-specific model combination scheme obtaining unit is used to perform fitness screening on the plurality of special model combination scheme sets respectively to obtain a plurality of target-specific model combination schemes, wherein the plurality of target-specific model combination schemes include a plurality of linkage calculation time identifiers; a processing time series obtaining unit, configured to determine a plurality of earliest inquiry start times of the plurality of inquiry record clusters according to the plurality of inquiry start times, and to perform linkage calculation time series analysis on the plurality of inquiry record clusters according to the plurality of earliest inquiry start times to obtain a plurality of inquiry record cluster processing time series; An optimization scheme obtaining unit is used to perform overlapping call time analysis of the dedicated models of the multiple target dedicated model combination schemes according to the multiple linkage calculation time identifiers and the multiple query record cluster processing time series, activate the linkage optimization scheduling module according to the analysis result, and optimize the computing performance allocation results of the multiple target dedicated model combination schemes by using the linkage optimization scheduling module to obtain multiple optimized target dedicated model combination schemes; The query record result cluster obtaining unit is used to perform linkage calculation on the multiple query record clusters based on the multiple optimization target dedicated model combination schemes to obtain multiple query record results.