Complex task scheduling decision method and system for power scenarios based on semantic big model

Abstract

The present invention discloses a method and system for scheduling complex tasks in power scenarios based on a semantic big model, which belongs to the field of power task scheduling, wherein the method includes: reading task processing records of diversified power scenarios, mining task scenario features and performing feature granularity refinement separation and clustering processing to determine key task features; traversing key task features to build a semantic big model; receiving pre-processed power tasks; performing task granularity refinement and decision-making layer allocation processing to determine task processing strategies; corresponding task processing strategies to determine a strategy distribution network; transmitting to a power management system to perform task local response management. This application solves the technical problem in the prior art that the decision-making accuracy of complex task scheduling in power scenarios is insufficient, resulting in low quality and low efficiency of task processing, and achieves the technical effect of improving the accuracy of decision-making of complex task scheduling in power scenarios and improving the quality and efficiency of task processing through a semantic big model.

Description

Technical Field

[0001] The present invention relates to the field of power task scheduling, and in particular to a method and system for scheduling complex tasks in power scenarios based on a semantic big model. Background Art

[0002] As the scale of the power grid continues to expand and the supply and demand situation becomes increasingly complex, higher requirements are placed on the intelligence and refinement of power dispatching. Among them, efficient scheduling decisions for complex tasks in power scenarios are the key to ensuring the safe and stable operation of the power system. In the existing technology, the scheduling decisions for complex tasks in power scenarios mainly adopt rule-based or heuristic algorithm methods. Such methods generally have the problem of rough task granularity division, and it is difficult to accurately characterize the characteristics of complex tasks, resulting in insufficient accuracy of scheduling decisions for complex tasks in power scenarios, which leads to poor task processing quality and low efficiency. Summary of the invention

[0003] This application aims to solve the technical problem in the prior art that the decision-making accuracy of complex task scheduling in power scenarios is insufficient, resulting in low quality and efficiency of task processing, by providing a method and system for scheduling complex tasks in power scenarios based on a semantic big model.

[0004] In view of the above problems, the present application provides a method and system for scheduling and deciding complex tasks in power scenarios based on a semantic big model.

[0005] The first aspect disclosed in the present application provides a complex task scheduling decision method for power scenarios based on a semantic big model, the method comprising: reading task processing records of diversified power scenarios, mining task scenario features and performing feature granularity refinement, separation and clustering processing to determine key task features; traversing key task features to construct a semantic big model, wherein the semantic big model comprises a front-end coordination and scheduling layer and a plurality of rear-end parallel task decision layers; receiving a pre-processed power task, the pre-processed power task comprising at least one parallel task; combining the semantic big model, refining the task granularity and performing decision layer allocation processing on the pre-processed power task to determine the task processing strategy; identifying the task source distribution of the pre-processed power task, corresponding to the task processing strategy, and determining the strategy distribution network; transmitting the strategy distribution network to the power management system to perform task local response management.

[0006] Another aspect disclosed in the present application provides a complex task scheduling decision system for power scenarios based on a semantic big model, the system comprising: a task feature determination module, used to read task processing records of diversified power scenarios, mine task scenario features and perform feature granularity refinement, separation and clustering processing to determine key task features; a model construction module, used to traverse key task features and construct a semantic big model, wherein the semantic big model includes a front-end coordination and scheduling layer and a plurality of rear-end parallel task decision layers; a task receiving module, used to receive pre-processed power tasks, the pre-processed power tasks include at least one parallel task; a strategy determination module, used to perform task granularity refinement and decision layer allocation processing on the pre-processed power tasks in combination with the semantic big model, and determine the task processing strategy; a network determination module, used to identify the task source distribution of the pre-processed power tasks, correspond to the task processing strategy, and determine the strategy distribution network; a local management module, used to transmit the strategy distribution network to the power management system for task local response management.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] By reading the task processing records of diversified power scenarios, mining the characteristics of task scenarios and performing feature granularity refinement, separation and clustering processing, key task characteristics are determined, historical data are fully utilized, and the multi-dimensional characteristics of complex power tasks are accurately characterized, laying the foundation for task modeling; traversing the key task characteristics, building a semantic big model, where the semantic big model includes a front-end coordination layer and multiple rear-end parallel task decision layers, using hierarchical modeling, the coordination layer is responsible for macro-scheduling, and the task decision layer is responsible for the refined processing of specific tasks, collaborative optimization, and improving the overall decision-making effect; receiving pre-processing power tasks, the pre-processing power tasks include at least one parallel task, introducing parallel tasks, making full use of computing resources, and improving task processing efficiency; combining the semantic big model with task granularity for the pre-processing power tasks The task processing strategy is determined by refining the task granularity and allocating it to the decision-making layer, and the task processing strategy is determined. The task granularity is further refined, and the strategy is formulated in combination with the task characteristics to achieve precise scheduling and improve the execution quality of a single task. The task source distribution of pre-processed power tasks is identified, the task processing strategy is matched, the strategy distribution network is determined, and the strategy distribution network is transmitted to the power management system for local task response management. The task response plan is quickly triggered at the terminal to achieve real-time dynamic adjustment of complex tasks, and the flexibility and efficiency of power scheduling are improved. This technical solution solves the technical problem that the accuracy of complex task scheduling decisions in power scenarios in the existing technology is insufficient, resulting in low quality and low efficiency of task processing, and achieves the technical effect of improving the accuracy of complex task scheduling decisions in power scenarios and improving the quality and efficiency of task processing through a large semantic model.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of a complex task scheduling decision method for power scenarios based on a semantic big model is provided for the embodiment of the present application;

[0011] Figure 2 A flowchart of constructing a semantic big model in a complex task scheduling decision method for power scenarios based on a semantic big model is provided for an embodiment of the present application;

[0012] Figure 3 A structural schematic diagram of a complex task scheduling decision system for power scenarios based on a semantic big model is provided for an embodiment of the present application.

[0013] Explanation of the accompanying drawings: task feature determination module 11, model construction module 12, task receiving module 13, strategy determination module 14, network determination module 15, local management module 16. DETAILED DESCRIPTION

[0014] The overall idea of ​​the technical solution provided by this application is as follows:

[0015] The embodiment of the present application provides a complex task scheduling decision method and system for power scenarios based on a semantic big model. First, by mining historical task data, extracting key task features, and characterizing task diversity, rich corpus is provided for semantic modeling. Then, using the semantic big model, a multi-level, collaborative task scheduling decision framework is built to seek a balance between overall planning and fine-grained processing, taking into account both macro and micro, achieving global optimization, and improving the accuracy of complex task scheduling decisions in power scenarios. Next, a parallel mechanism is introduced to optimize resource allocation and improve system throughput. At the same time, the task granularity is refined, and precise scheduling strategies are formulated to ensure the execution quality of a single task. Furthermore, according to the distributed source characteristics of the task, a strategy distribution network is dynamically generated, and the decision-making tentacles are extended to the edge end, so that the response is closer to the demand source. Finally, the strategy distribution network is transmitted to the power management system to perform local task response management, give play to real-time regulation and control efficiency, and improve task processing quality and efficiency.

[0016] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically described below in conjunction with the drawings in the specification.

[0017] Embodiment 1

[0018] like Figure 1As shown, the embodiment of the present application provides a complex task scheduling decision method for power scenarios based on a semantic big model, the method comprising:

[0019] S1: Read the task processing records of diversified power scenarios, mine the task scenario features, perform feature granularity refinement, separation and clustering, and determine the key task features;

[0020] Furthermore, this step includes:

[0021] Traversing the task scene features, identifying feature complexity, performing feature granularity refinement and separation on the scene task features that meet the complexity threshold, and determining multiple groups of task features;

[0022] Based on the task processing requirements, hierarchical clustering is performed on the multiple groups of task features to determine feature clustering clusters;

[0023] The feature clusters are identified, and the key task features are determined, wherein the key task features correspond one-to-one to the feature clusters.

[0024] In an embodiment of the present application, in order to construct a semantic large model for complex task scheduling decisions in power scenarios, the task processing records in the diversified scenarios of the power system are first read, which contain the execution status data of various tasks, such as the input parameters of the task, the output results, the called service interface, the execution timestamp, etc. By analyzing these task processing records, the commonalities and individualities of task execution in different scenarios are found, thereby obtaining the task scenario characteristics. Since the actual task scenarios are ever-changing, in order to make the features more refined and targeted, the task scenario features initially mined are further fine-grained. Dissociation. Among them, the dissociation process can adopt a variety of strategies, such as splitting features based on semantic levels, segmenting features based on data attributes, etc., to decompose complex task scenario features into several fine-grained sub-features, which are convenient for subsequent feature representation and processing. After obtaining the task scenario features after fine-grained dissociation, we further explore the intrinsic connections between the features. By using clustering algorithms such as hierarchical clustering and K-means clustering, we classify the semantically related and similar execution mode features together and divide the features into different clusters. Each cluster represents a type of task scenario, thereby determining the key task features.

[0025] In a feasible implementation, for the mined task scene features, it is necessary to further judge their complexity to determine whether it is necessary to perform feature granularity refinement and separation processing. First, a feature complexity identification algorithm is used to traverse and analyze each task scene feature to identify the feature complexity of the task scene feature. Among them, the feature complexity identification algorithm evaluates the complexity of the feature from multiple dimensions, such as the number of attributes contained in the feature, the discrete degree of the attribute value, the correlation between the feature and other features, etc. Next, the identified feature complexity value is compared with a preset complexity threshold. Among them, the complexity threshold is set according to the experiment and represents the critical point of the feature complexity. When the feature complexity of a task scene feature exceeds the threshold, it is considered that the feature is too complex and needs to be further fine-grained. For task scene features whose feature complexity does not reach the threshold, the original granularity remains unchanged. For task scene features that meet the complexity threshold, a feature granularity refinement and separation operation will be performed on them to divide the complex scene features into multiple semantically independent and finer granularity sub-features, thereby reducing the complexity of a single feature and improving the interpretability and processability of the feature. Among them, the dissociation method is determined according to the semantic structure and data type of the feature, for example, segmentation according to attribute dimensions, division according to numerical intervals, disassembly based on association rules, etc. After feature granularity refinement and dissociation, the original complex task scene feature will be split into multiple groups of fine-grained task features. Each group of task features is independent of each other, but together they depict different aspects of the original task scene features, resulting in multiple groups of task features.

[0026] According to the task processing requirements of the power scenario, the objectives and evaluation indicators of cluster analysis are determined, including the real-time requirements of the task, the priority level of the task, the resource consumption of the task, etc. After obtaining multiple groups of task features after refinement and separation, the similarity and correlation between these features are further analyzed, and the features with similar semantics and similar execution modes are classified into one category. Among them, the hierarchical clustering method is used to cluster multiple groups of task features. Each group of task features is first regarded as an independent class, and then the most similar classes are gradually merged until all task features are classified into one class or the preset number of clusters is reached. In the merging process, the distance or similarity between each class is calculated by Euclidean distance, cosine similarity, Pearson correlation coefficient, etc. Through hierarchical clustering processing, multiple groups of task features are gradually classified to form a tree-like clustering structure. In the tree-like clustering structure, the leaf node represents the original task feature, the intermediate node represents the aggregated feature class, and the root node represents the final large class. Subsequently, the tree-like clustering structure is segmented to obtain several feature clusters, each of which contains a group of task features with similar semantics and processing modes, representing a type of task scenario.

[0027] Subsequently, for each feature cluster, a key feature extraction algorithm is used for analysis. Among them, the key feature extraction algorithm comprehensively considers the feature distribution, feature weight, feature frequency and other factors within the cluster, and automatically identifies the key features that are representative and distinguishing for the cluster; the key feature extraction methods include TF-IDF weight method, chi-square test method, mutual information method, etc. By calculating the importance score of each feature in the feature cluster, one or several features with the highest score are selected as the key task features of the cluster, thereby determining the key task features. When determining the key task features, it is necessary to ensure that each feature cluster has a corresponding key task feature to form a one-to-one mapping relationship. This mapping relationship is represented by a dictionary or hash table to achieve fast retrieval and matching.

[0028] S2: Traverse the key task features and build a semantic big model, wherein the semantic big model includes a front-end coordination and arrangement layer and a plurality of rear-end parallel task decision layers;

[0029] Further, such as Figure 2 As shown, this step includes:

[0030] Identify the key mission characteristics, configure the mission decision algorithm and computing power scale;

[0031] Preprocessing the task processing records to determine training samples, the training samples include an extraction sample set of each power scenario task-key task feature and an analysis sample set of the key task feature-task processing strategy;

[0032] Mapping the task decision algorithm, the computing power scale and the training samples, and supervising the training of the semantic big model that meets the single convergence condition and the overall convergence condition.

[0033] Furthermore, generating a semantic large model also includes:

[0034] Based on the analysis sample set, sample division is performed based on the attribution of the key task features, and supervised training of the task decision layer is performed with the mapped task decision algorithm and model scale as constraints, wherein the task decision layer corresponds to the key task features one by one;

[0035] Based on the extracted sample set, supervise the training of the overall planning layer;

[0036] The task decision layer is configured in parallel and is post-connected to the overall planning layer to generate the semantic big model.

[0037] In an embodiment of the present application, after obtaining the key task features, a semantic big model is constructed based on the key task features to arrange and make decisions on complex tasks in power scenarios. The semantic big model adopts a layered architecture that combines front-end coordinated arrangement with rear-end parallel decision-making, which can take into account both the global optimization and local refinement of task processing. Specifically, all key task features obtained are traversed, and a corresponding task decision layer is created for each key task feature. Among them, the task decision layer is the basic processing unit of the semantic big model, which is responsible for special arrangement and decision-making for a certain type of task scenario, and through a special task decision layer, refined processing of different types of tasks is achieved. Then, at the front end of all task decision layers, a coordinated arrangement layer is set. Among them, the coordinated arrangement layer is responsible for preliminary analysis and disassembly of the complex tasks input, and distributes the tasks to different task decision layers for processing. Afterwards, the overall orchestration layer is connected in series with all task decision layers to form a complete semantic model. In this model, the overall orchestration layer is at the front end and is responsible for receiving complex tasks from external input; multiple task decision layers are deployed in parallel at the back end to handle different categories of task sub-scenarios respectively. The overall orchestration layer and the task decision layer interact through data channels and control protocols to achieve coordination and optimization of the task processing process.

[0038] In a feasible implementation, in order to build an efficient and accurate semantic big model, for each key task feature, a task decision algorithm adapted to it is designed, and a reasonable computing power scale is configured. First, identify the key task features obtained, analyze the task scenario category represented by each key feature, and evaluate the complexity, real-time, concurrency and other characteristics of the task in this category. Then, different task decision algorithms are adopted according to the characteristics of the task scenario category. For example, for task categories with clear rules and fixed processes, a rule-based decision engine is used to quickly generate task processing strategies through a predefined rule base and reasoning mechanism; for task categories with complex logic and variable strategies, machine learning algorithms such as decision trees and random forests are used to automatically learn and optimize decision strategies in a data-driven manner; for task categories that require deep understanding and reasoning, deep neural networks, knowledge graphs and other artificial intelligence technologies are used to establish semantic representations and associated reasoning models of tasks. After selecting the task decision algorithm, appropriate computing resources are configured according to factors such as the data scale and computational complexity of the task scenario category. For example, for task categories with large data volumes and intensive computing, more computing nodes and memory resources are equipped to support large-scale parallel computing and real-time response; for task categories with high real-time requirements and delay sensitivity, low-latency, high-throughput computing architectures are adopted, such as FPGA, GPU and other hardware accelerators, to meet the timeliness requirements of task processing.

[0039] Then, the massive task processing records of the power system are preprocessed. The task processing records contain detailed information such as the input, output, service call, and execution process of various types of tasks; preprocessing includes data cleaning, format conversion, field extraction and other operations to convert the original task processing records into structured and standardized data representations. Then, based on the preprocessed task records, two types of training sample sets are constructed, namely the extraction sample set and the analysis sample set. Among them, the extracted sample set is used to train the overall planning layer in the semantic big model. For each pre-processed task processing record, the key features of the task are extracted from it, and the power scenario category to which the task belongs is marked, and the paired samples of power scenario task-key task features are obtained, which describe the feature representation of tasks under different power scenarios and are used to train the task scenario recognition and task feature extraction capabilities of the overall planning layer; the analysis sample set is used to train each task decision layer in the semantic big model. For each pre-processed task processing record, its key task features are paired with the task processing strategy actually adopted to form an analysis sample of key task features-task processing strategy, which characterizes the optimal decision behavior under specific task features and is used to train the strategy optimization and resource scheduling capabilities in the task decision layer, so that each task decision layer can master the processing rules and optimization strategies of the task category. After completing the configuration of the task decision algorithm and computing power scale, and building the training sample set, they are mapped and supervised training of the semantic big model is carried out. For each power scenario task-key task feature sample in the extracted sample set, according to its power scenario category, it is mapped to the corresponding task decision layer and its configured decision algorithm and computing power scale, so as to realize the sample division based on the attribution of key task features. Similarly, for each key task feature in the analysis sample set - the task processing strategy sample, it is mapped to the corresponding task decision layer. Through the mapping relationship, the corresponding relationship between the training sample and the components of the semantic large model is established, preparing for the supervised training of the model.

[0040] Then, based on the mapped task decision algorithm, computing power scale and training samples, supervised training is performed on the semantic large model. Specifically, for each key task feature, the corresponding task decision layer is selected for special supervised training. During the training process, the sample subset corresponding to the key task feature is used as training data, and the mapped task decision algorithm and model scale are used as constraints to perform parameter learning and performance optimization on the policy optimization and resource scheduling model within the task decision layer. Through repeated training on exclusive samples, each task decision layer can handle such tasks more accurately and efficiently, forming a special task processing capability. In the supervised training of the task decision layer, the training objectives and convergence conditions are set, and the loss function is selected according to the characteristics of each task decision layer, such as task completion quality, resource utilization efficiency, etc., to measure the training effect of the decision layer.

[0041] In addition to targeted training for each task decision layer, the overall orchestration layer of the semantic big model also needs to be specially supervised and trained. The supervised training of the overall orchestration layer is based on the constructed extracted sample set, which contains paired samples of power scenario tasks and key task features, and characterizes the feature representation of tasks in different power scenarios. During the training process, the power scenario task of each extracted sample is used as input, and the key task features are used as the expected output to supervise the scene recognition and feature extraction model within the overall orchestration layer. By calculating the error between the model output and the expected output, and using the optimization algorithm to adjust the model parameters, the overall orchestration layer can identify the scene to which the task belongs as accurately as possible and extract key features from it. Through the special training of the overall orchestration layer, the semantic big model can handle various complex power tasks intelligently and adaptively.

[0042] After completing the supervised training of each task decision layer and the overall orchestration layer, the trained overall orchestration layer is integrated with each task decision layer to form a complete semantic large model. Since each task decision layer is specially optimized for a specific type of key task features, they are independently and parallelly deployed. By configuring each task decision layer on the same level, they can simultaneously receive task distribution from the overall orchestration layer and independently and synchronously process tasks, giving full play to the processing capabilities of each task decision layer and improving the concurrency and throughput of task processing. Then, the parallel configured task decision layer is post-connected to the overall orchestration layer, that is, the output of the overall orchestration layer will be directly used as the input of each task decision layer. Among them, the overall orchestration layer distributes the task to one or more corresponding task decision layers for processing based on the scene recognition and key feature extraction results of the task. A complete semantic model is formed by coordinating the parallel configuration and hierarchical connection of the orchestration layer and the task decision layer. The unified access, scene recognition, feature extraction and intelligent distribution of tasks are achieved through the orchestration layer. The specialized processing and optimized decision-making of tasks are achieved through the parallel task decision layer. The global coordination and local refinement of task processing are taken into account, which significantly improves the processing efficiency and decision-making quality of complex tasks.

[0043] During the training process of the semantic big model, the convergence termination conditions of the model are set, including single convergence conditions and overall convergence conditions. Among them, the single convergence condition is for each component of the semantic big model, such as the overall arrangement layer, each task decision layer, etc., by setting the loss function, evaluation indicators, etc., to determine whether the training of each part has reached convergence; the overall convergence condition is for the overall performance of the semantic big model, by setting the generalization error of the model, task completion quality and other indicators, to determine whether the model has reached the expected performance requirements. Only when the two conditions of single convergence and overall convergence are met, the semantic big model training is considered completed.

[0044] Through supervised training of the semantic big model, it can form intelligent task scheduling and decision-making capabilities, fully understand the characteristics of power scenarios, adaptively handle various complex tasks, generate high-quality and efficient task decision plans, and provide support for the intelligent operation of the power system.

[0045] S3: receiving a pre-processing power task, wherein the pre-processing power task includes at least one parallel task;

[0046] In the embodiment of the present application, after the construction and training of the semantic big model is completed, the model can be used to process actual power scenario tasks. During the operation of the power system, a large number of tasks will be generated, such as fault diagnosis, load forecasting, equipment maintenance, scheduling optimization, etc.

[0047] Receive task requests from various application scenarios of the power system. Each task request contains basic information of the task, such as task type, priority, deadline, etc., as well as detailed description of the task and input data. After receiving the task request, preprocess the task request and convert the original task request into a standardized task format that can be directly processed by the semantic big model as a preprocessed power task. Among them, the preprocessed power task is at least a parallel task. By means of parallel tasks, the parallel processing capability of the semantic big model can be fully utilized to improve the efficiency of task processing and achieve the optimal task processing performance.

[0048] S4: combining the semantic big model, refining the task granularity and decision-making layer allocation processing of the pre-processing power task, and determining the task processing strategy;

[0049] Furthermore, this step includes:

[0050] Identify the pre-processing power task, perform dissociation analysis and decision processing sorting of key task characteristics based on the overall planning layer, and determine the task processing plan;

[0051] Based on the task processing plan, multiple task decision layers are selectively activated, computing power is redistributed and task serialization is performed, and the task processing strategy is obtained.

[0052] Furthermore, based on the overall planning layer, the dissociation analysis and decision processing of key task features are carried out, including:

[0053] By conducting task dissociation analysis, unrelated subtasks and interrelated subtasks are identified;

[0054] Interacting the processing scale of each task decision layer to determine the idle computing power configuration, where the idle computing power configuration is a time zone variable;

[0055] Based on the idle computing power configuration, the non-related subtasks and the interrelated subtasks are comprehensively sorted in the order of executing task processing to determine the task processing plan, wherein the comprehensive sorting standard includes the intersection of parallel sorting and sequential sorting.

[0056] In an embodiment of the present application, after receiving the pre-processed power task, the constructed semantic big model is used to further refine the task granularity and allocate the decision-making layer to determine the optimal task processing strategy.

[0057] First, the semantic big model extracts the pre-processing power task and inputs it into the overall planning layer. The overall planning layer conducts in-depth semantic analysis and structural analysis on the pre-processing power task. Through the task dissociation analysis technology, the task is split into multiple relatively independent subtasks according to the semantic logic and data dependency. On the basis of task dissociation, the overall planning layer further identifies the correlation between subtasks and divides the subtasks into two categories: non-correlated subtasks and inter-correlated subtasks. Among them, non-correlated subtasks are subtasks without direct data dependency or execution sequence relationship, which can be processed independently and have the potential for parallel execution. For example, in the comprehensive energy optimization task, subtasks such as renewable energy generation forecasting and load demand forecasting in different regions can be carried out in parallel without explicit correlation between each other; inter-correlated subtasks are subtasks with certain data dependency or execution sequence relationship, which need to be processed serially in a certain logical order, or necessary synchronization and coordination are carried out on the basis of parallel processing. For example, in the power grid fault diagnosis task, subtasks such as fault data collection, fault type judgment, and fault location analysis need to be carried out in sequence in a certain logical order, and there is an obvious correlation.

[0058] After the overall planning layer completes the identification of the subtasks of the pre-processing power task, the overall planning layer interacts with each task decision layer to obtain the current processing scale information of each task decision layer. Among them, the processing scale information includes the total computing resources of the task decision layer, the resource occupancy of the assigned tasks, the length of the task queue, etc. By analyzing the processing scale information of each task decision layer, the overall planning layer can accurately evaluate the real-time load status and remaining computing power of each task decision layer. After obtaining the processing scale information of each task decision layer, the overall planning layer calculates the idle computing power configuration of each decision layer by considering factors such as the total computing power, occupied computing power, and task saturation of the task decision layer. Among them, the idle computing power configuration represents the amount of idle computing resources that each task decision layer can use for new task processing at the current moment. The idle computing power configuration is a time zone variable, that is, it changes dynamically over time. In different time periods, the task load and resource status of the power system may change, causing the idle computing power of the task decision layer to fluctuate accordingly. In order to adapt to this dynamic change, the overall coordination layer will interact with the task decision layer in real time to update the idle computing power configuration, ensure that the overall coordination layer always grasps the latest processing capabilities of the task decision layer, and make timely and accurate task scheduling decisions.

[0059] After obtaining the idle computing power configuration and subtask correlation identification results, the overall planning layer comprehensively sorts the processing order of tasks based on this information and generates the optimal task processing plan. For non-correlated subtasks, the overall planning layer sorts them from the perspective of parallel processing. According to factors such as the task volume, computational complexity, and idle computing power configuration of each non-correlated subtask, each non-correlated subtask is assigned to different task decision layers for parallel processing in accordance with certain priorities and load balancing principles, so as to maximize the parallel processing capabilities of the semantic large model and significantly improve the efficiency of task processing. For interrelated subtasks, the overall planning layer sorts them from the perspective of task dependencies. By analyzing the data dependencies and execution order logic between interrelated subtasks, the interrelated subtasks are arranged in a certain topological sequence to ensure the logical accuracy of task execution. At the same time, the overall planning layer considers the idle computing power configuration of the task decision layer and assigns the interrelated subtasks to the task decision layer with lighter load as much as possible to achieve load balancing and optimal resource utilization. In addition, in the actual process of overall arrangement, non-related subtasks and interrelated subtasks are carried out in an interleaved manner, and the overall arrangement layer needs to jointly arrange non-related subtasks and interrelated subtasks. For example, for an interrelated subtask, there may be non-related subtasks that can be processed in parallel. The overall arrangement layer will sort and allocate the non-related subtasks in parallel while ensuring the execution order of the interrelated subtasks, so as to maximize the parallel processing potential of the task.

[0060] After generating the optimized task processing plan, since the task processing plan has clarified the decision layer allocation of each subtask, the overall orchestration layer can activate specific task decision layers in a targeted manner; and for the task decision layers that have not been assigned tasks temporarily, the overall orchestration layer chooses to temporarily shut down or set them to low power consumption mode to save computing resources. On the basis of selective activation of the decision layer, the overall orchestration layer further reallocates computing resources. According to the task allocation of each task decision layer in the task processing plan and the idle computing power configuration of the task decision layer, the overall orchestration layer dynamically adjusts the resource quota of each task decision layer. For decision layers with heavy task loads, the overall orchestration layer will appropriately increase their computing power resource quota to meet the needs of task processing; for task decision layers with light task loads, the overall orchestration layer will correspondingly reduce their computing power resource quota and transfer excess computing power resources to other task decision layers. Through the dynamic computing power reallocation mechanism, the optimization of resource utilization can be achieved to ensure the overall processing efficiency of the semantic large model. After completing the decision layer activation and computing power reallocation, the overall orchestration layer begins to serialize the tasks to obtain the task processing strategy. According to the task execution order determined by the task processing plan, the overall planning layer will send the tasks to the corresponding task decision layer for processing. For non-related subtasks, the overall planning layer will distribute them to multiple task decision layers for parallel processing; for interrelated subtasks, the overall planning layer will strictly follow the logical dependency order of the interrelated subtasks and distribute them to the task decision layer for serial processing.

[0061] Identify the distribution of task sources of the pre-processed power task, correspond to the task processing strategy, and determine the strategy distribution network;

[0062] In an embodiment of the present application, after obtaining the task processing strategy, the source distribution of the pre-processed power task is identified, and the task processing strategy is matched with the task source to form a strategy distribution network.

[0063] First, the distribution of task sources of pre-processed power tasks is identified to reflect the attribute information of power tasks such as geographical origin, organizational affiliation, and business type. For example, by analyzing the metadata and context of task requests, the source distribution of each pre-processed power task can be accurately identified, such as identifying that a task comes from a specific substation, a task belongs to a specific distribution network area, and a task belongs to a specific business department. Then, based on the identification of task source distribution, the task processing strategy is matched with the task source, and each subtask in the task processing strategy is associated with its corresponding task source, so that each subtask establishes a mapping relationship with one or more task sources to form a ternary combination of task-strategy-task source. After completing the matching of task processing strategy and task source, a strategy distribution network is further generated to organize the scattered task sources and task processing strategies into a unified framework according to the association logic and scheduling mechanism. In the strategy distribution network, each node represents a task source and a task processing strategy, and the connection relationship between nodes represents the interaction and collaboration between them.

[0064] By identifying the source distribution of power tasks, accurately matching task processing strategies with task sources, and generating a strategy distribution network, the semantic big model is deeply integrated into the actual operating environment of the power system, realizing efficient distribution and execution of task processing strategies in distributed scenarios, and improving the real-time collaborative control and optimized scheduling level of the power system.

[0065] The strategy distribution network is transmitted to the power management system for task local response management.

[0066] In an embodiment of the present application, after obtaining the strategy distribution network, the network is transmitted to the power management system, and the task processing strategy is executed in the local environment of the power system to achieve real-time response and closed-loop management of the task.

[0067] First, the generated strategy distribution network is transmitted to the power management system through a standardized interface protocol. The power management system is the core control and dispatching platform of the power system, responsible for real-time monitoring of the power grid, decision optimization, and control command issuance. When the strategy distribution network is successfully transmitted to the power management system, the power management system will accurately issue the task processing strategy to each local node of the power system, such as substations, distribution rooms, and user terminals, based on the node attributes and connection relationships in the strategy distribution network. Each local node will perform corresponding control actions or optimization measures based on the received task processing strategy and local real-time status data to achieve local response to power tasks.

[0068] In the process of task local response management, the power management system will provide a series of management functions, including strategy status monitoring, effect evaluation, dynamic adjustment, etc. By collecting the task execution status and performance indicators of each local node in real time, the power management system can evaluate the effect of strategy execution and find possible problems or optimization space. According to the evaluation results, the power management system dynamically adjusts the strategy distribution network, optimizes the allocation and execution of task processing strategies at local nodes, and realizes closed-loop optimization of strategy execution.

[0069] Among them, task local response management is a continuous iterative process. The operating environment of the power system is dynamically changing, and new task requirements and optimization goals will continue to emerge. Therefore, the power management system maintains real-time two-way communication with the semantic big model, and regularly transmits feedback data and optimization requirements of task execution to the semantic big model, triggering its re-optimization of the strategy distribution network. At the same time, the semantic big model will continue to generate and transmit the optimized strategy distribution network to guide the power management system to dynamically update and adjust the strategy. Through continuous iterative optimization, the task processing strategy is always kept in the best state and adapted to the real-time changes of the power system, thereby improving the accuracy of complex task scheduling decisions in power scenarios and improving the quality and efficiency of task processing.

[0070] Furthermore, after performing task local response management, the embodiment of the present application further includes:

[0071] Simultaneously conduct task-local response monitoring to mine strategy execution variables that meet generalization criteria;

[0072] Reading the model processing records of a predetermined period to determine the model processing constraints;

[0073] The policy execution variables and the model processing constraints are transmitted back, and incremental learning is performed on the semantic large model to determine an incremental semantic model.

[0074] In a preferred embodiment, during the task local response management process, the power management system monitors the execution effect of the task processing strategy in real time, mines key strategy execution variables, and provides a basis for the continuous optimization of the semantic big model. Among them, the strategy execution variables are factors that affect the strategy execution effect, such as grid topology changes, equipment failure rates, load forecast errors, etc., reflecting various uncertainties and interference factors in the actual operation of the power system.

[0075] Then, read the operation log and performance report of the semantic big model regularly, analyze the key indicators such as resource usage, task saturation, processing delay, etc. of the model, and evaluate the overall operation status of the semantic big model by analyzing the processing records within a predetermined period (such as one week, one month, etc.), to find out whether there are problems such as uneven distribution of computing resources, task congestion, and insufficient model generalization. Based on the analysis results of the model processing records, determine the model processing constraints to guide the incremental learning and dynamic optimization of the semantic big model. For example, constrain the scale ratio of different task processing layers, appropriately expand the processing layers with greater computing power requirements and higher processing frequency, and appropriately reduce the processing layers with lower utilization, so as to achieve overall computing power balance and efficiency improvement; constrain the maximum complexity of the model and the number of training rounds to prevent model overfitting and performance degradation.

[0076] After obtaining the policy execution variables and model processing constraints, the policy execution variables and model processing constraints are fed back to the semantic big model. After receiving the feedback information, the semantic big model will immediately start an incremental learning. On the basis of the original model, the newly obtained policy execution variables and model processing constraints are used to modify and expand the model in a targeted manner to improve the adaptability of the model. The new model obtained by incremental learning, namely the incremental semantic model, inherits the existing capabilities of the original semantic big model while improving some strategies and constraints in a targeted manner to better adapt to the current operation status and regulation requirements of the power grid. After the incremental semantic model is generated, it will be deployed to the power management system again to undergo a new round of testing and optimization. In this way, the semantic big model will continuously evolve itself, resonate with the power system, and continuously improve the accuracy and efficiency of intelligent control of the power grid.

[0077] By synchronously monitoring the effect of strategy execution, extracting strategy execution variables, analyzing model processing constraints, and combining mechanisms such as incremental learning, real-time interaction and closed-loop optimization between the semantic big model and the power system are achieved, so that the semantic big model can continuously learn the operating laws of the power system and use the optimized strategy to guide the power system to achieve more intelligent, efficient and safe operation and scheduling.

[0078] In summary, the complex task scheduling decision method for power scenarios based on a semantic big model provided by the embodiment of the present application has the following technical effects:

[0079] Read the task processing records of diversified power scenarios, mine the characteristics of task scenarios, and perform feature granularity refinement, separation and clustering processing to determine the key task characteristics, providing data foundation and feature support for subsequent task modeling and strategy formulation. Traverse the key task characteristics and build a semantic big model, where the semantic big model includes a front-end coordination and scheduling layer and multiple back-end parallel task decision layers. The coordination and scheduling layer is responsible for global macro scheduling, and the task decision layer is responsible for the micro processing of specific tasks. The two have clear division of labor and coordinated optimization in order to achieve global optimization, thereby improving the accuracy of complex task scheduling decisions in power scenarios. Receive pre-processing power tasks, which include at least one parallel task. Introduce a parallel processing mechanism, and make full use of computing resources by scheduling parallel tasks to improve the system's concurrent processing capability and task throughput. Combined with the semantic big model, perform task granularity refinement and decision layer allocation processing on the pre-processing power tasks, determine the task processing strategy, match the optimal decision for each subtask, and ensure the execution quality of a single task. Identify the distribution of task sources for pre-processing power tasks, match task processing strategies, determine the strategy distribution network, and sink the task response plan to the demand terminal, so that the decision is closer to reality and the response is more timely and sensitive. Transmit the strategy distribution network to the power management system, perform local task response management, perform real-time scheduling and dynamic optimization at the local location of the task source, so that the execution plan can be quickly implemented and effective, give full play to the advantages of edge computing, and improve the technical effect of task processing quality and efficiency.

[0080] Embodiment 2

[0081] Based on the same inventive concept as the power scenario complex task scheduling decision method based on the semantic big model in the aforementioned embodiment, Figure 3 As shown, the embodiment of the present application provides a complex task scheduling decision system for power scenarios based on a semantic big model, the system comprising:

[0082] The task feature determination module 11 is used to read the task processing records of diversified power scenarios, mine the task scenario features, perform feature granularity refinement, separation and clustering processing, and determine the key task features;

[0083] A model building module 12 is used to traverse the key task features and build a semantic big model, wherein the semantic big model includes a front-end coordination and arrangement layer and a plurality of rear-end parallel task decision layers;

[0084] A task receiving module 13, used for receiving a pre-processing power task, wherein the pre-processing power task includes at least one parallel task;

[0085] A strategy determination module 14 is used to combine the semantic big model, perform task granularity refinement and decision layer allocation processing on the pre-processing power task, and determine the task processing strategy;

[0086] A network determination module 15, for identifying the distribution of task sources of the pre-processed power task, corresponding to the task processing strategy, and determining a strategy distribution network;

[0087] The local management module 16 is used to transmit the strategy distribution network to the power management system to perform task local response management.

[0088] Furthermore, the task feature determination module 11 includes the following execution steps:

[0089] Traversing the task scene features, identifying feature complexity, performing feature granularity refinement and separation on the scene task features that meet the complexity threshold, and determining multiple groups of task features;

[0090] Based on the task processing requirements, hierarchical clustering is performed on the multiple groups of task features to determine feature clustering clusters;

[0091] The feature clusters are identified, and the key task features are determined, wherein the key task features correspond one-to-one to the feature clusters.

[0092] Furthermore, the model building module 12 includes the following execution steps:

[0093] Identify the key mission characteristics, configure the mission decision algorithm and computing power scale;

[0094] Preprocessing the task processing records to determine training samples, the training samples include an extraction sample set of each power scenario task-key task feature and an analysis sample set of the key task feature-task processing strategy;

[0095] Mapping the task decision algorithm, the computing power scale and the training samples, and supervising the training of the semantic big model that meets the single convergence condition and the overall convergence condition.

[0096] Furthermore, the model building module 12 also includes the following execution steps:

[0097] Based on the analysis sample set, sample division is performed based on the attribution of the key task features, and supervised training of the task decision layer is performed with the mapped task decision algorithm and model scale as constraints, wherein the task decision layer corresponds to the key task features one by one;

[0098] Based on the extracted sample set, supervise the training of the overall planning layer;

[0099] The task decision layer is configured in parallel and is post-connected to the overall planning layer to generate the semantic big model.

[0100] Furthermore, the strategy determination module 14 includes the following execution steps:

[0101] Identify the pre-processing power task, perform dissociation analysis and decision processing sorting of key task characteristics based on the overall planning layer, and determine the task processing plan;

[0102] Based on the task processing plan, multiple task decision layers are selectively activated, computing power is redistributed and task serialization is performed, and the task processing strategy is obtained.

[0103] Furthermore, the strategy determination module 14 further includes the following execution steps:

[0104] By conducting task dissociation analysis, unrelated subtasks and interrelated subtasks are identified;

[0105] Interacting the processing scale of each task decision layer to determine the idle computing power configuration, where the idle computing power configuration is a time zone variable;

[0106] Based on the idle computing power configuration, the non-related subtasks and the interrelated subtasks are comprehensively sorted in the order of executing task processing to determine the task processing plan, wherein the comprehensive sorting standard includes the intersection of parallel sorting and sequential sorting.

[0107] Furthermore, the embodiment of the present application also includes an incremental learning module, which includes the following execution steps:

[0108] Simultaneously conduct task-local response monitoring to mine strategy execution variables that meet generalization criteria;

[0109] Reading the model processing records of a predetermined period to determine the model processing constraints;

[0110] The policy execution variables and the model processing constraints are transmitted back, and incremental learning is performed on the semantic large model to determine an incremental semantic model.

[0111] Any step of the method described above can be stored as a computer instruction or program in an unlimited computer memory, and can be called and recognized by an unlimited computer processor to implement any method in the embodiments of the present application, without any unnecessary restrictions.

[0112] Obviously, those skilled in the art can make various changes and modifications 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 equivalent technologies, the present application is intended to include these modifications and variations.

Claims

1. A complex task scheduling decision method for power scenarios based on a semantic big model, characterized by: The method comprises: Read the task processing records of diversified power scenarios, mine the characteristics of task scenarios, perform feature granularity refinement, separation and clustering, and determine key task characteristics; Traversing the key task features, constructing a semantic big model, wherein the semantic big model includes a front-end coordination layer and multiple back-end parallel task decision layers, the coordination layer is responsible for preliminary analysis and decomposition of the input complex tasks, and distributes the tasks to different task decision layers for processing; receiving a pre-processed power task, the pre-processed power task comprising at least one parallel task; Combined with the semantic big model, the pre-processing power task is refined in task granularity and allocated to the decision layer to determine the task processing strategy; Identify the distribution of task sources of the pre-processed power task, correspond the task processing strategy, and determine the strategy distribution network, wherein: identify the distribution of task sources of the pre-processed power task to reflect the attribute information of the geographical origin, organizational affiliation, and business type of the power task; on the basis of identifying the distribution of task sources, match the task processing strategy with the task source, associate each subtask in the task processing strategy with its corresponding task source, establish a mapping relationship between each subtask and one or more task sources, and form a ternary combination of task-strategy-task source; transmitting the strategy distribution network to the power management system for task local response management; The mining of task scene features and performing feature granularity refinement, separation and clustering processing include: Traversing the task scene features, identifying feature complexity, performing feature granularity refinement and separation on the task scene features that meet the complexity threshold, and determining multiple groups of task features, wherein: the dissociation method is determined according to the semantic structure and data type of the feature, specifically segmenting according to attribute dimensions, dividing according to numerical intervals, or disassembling based on association rules; Based on the task processing requirements, the multiple groups of task features are hierarchically clustered to determine the feature clustering clusters, wherein: according to the task processing requirements of the power scenario, the objectives and evaluation indicators of the clustering analysis are determined, including the real-time requirements of the task, the priority level of the task, and the resource consumption of the task; in the merging process, the distance or similarity between the classes is calculated by means of Euclidean distance, cosine similarity, and Pearson correlation coefficient; through the hierarchical clustering process, the multiple groups of task features are gradually classified to form a tree-like clustering structure; Identify the feature clusters and determine the key task features, wherein the key task features correspond one-to-one to the feature clusters; Traversing the key task features, the step of constructing a semantic big model includes: Identify the key mission characteristics, configure the mission decision algorithm and computing power scale; Preprocessing the task processing records to determine training samples, the training samples include an extraction sample set of each power scenario task-key task feature and an analysis sample set of the key task feature-task processing strategy; Mapping the task decision algorithm, the computing power scale and the training samples, and supervising the training of the semantic big model that meets the single convergence condition and the overall convergence condition; The method further comprises: Based on the analysis sample set, sample division is performed according to the attribution of the key task features, and supervised training of the task decision layer is performed with the mapped task decision algorithm and model scale as constraints, and the task decision layer corresponds to the key task features one by one; Based on the extracted sample set, supervise the training of the overall planning layer; The task decision layer is configured in parallel and is subsequently connected to the overall planning layer to generate the semantic big model; The pre-processing power task is subjected to task granularity refinement and decision-making layer allocation processing, including: Identify the pre-processing power task, perform dissociation analysis and decision processing sorting of key task characteristics based on the overall planning layer, and determine the task processing plan; Based on the task processing plan, multiple task decision layers are selectively activated, computing power is redistributed and task serialization is performed, and the task processing strategy is obtained; Based on the overall planning layer, the key task characteristics are analyzed and the decision-making process is sorted, including: By conducting task dissociation analysis, unrelated subtasks and interrelated subtasks are identified; Interacting the processing scale of each task decision layer to determine the idle computing power configuration, where the idle computing power configuration is a time zone variable, wherein the idle computing power configuration represents the amount of idle computing resources available for new task processing at each task decision layer at the current moment, and the idle computing power configuration is a time zone variable that changes dynamically over time; Based on the idle computing power configuration, the non-related subtasks and the interrelated subtasks are comprehensively sorted in the order of executing task processing to determine the task processing plan, wherein the comprehensive sorting standard includes the intersection of parallel sorting and sequential sorting.

2. The method according to claim 1, characterized in that After performing task local response management, the method further includes: Simultaneously carry out task local response monitoring and mine strategy execution variables that meet the generalization standard, where the strategy execution variables are factors that affect the strategy execution effect, including grid topology changes, equipment failure rates, and load forecast errors, reflecting various uncertainties and interference factors in the actual operation of the power system; Reading the model processing records of a predetermined period to determine the model processing constraints, wherein: the model processing constraints are determined based on the analysis results of the model processing records to guide the incremental learning and dynamic optimization of the semantic large model; The policy execution variables and the model processing constraints are transmitted back, and incremental learning is performed on the semantic large model to determine an incremental semantic model.

3. A complex task scheduling decision system for power scenarios based on a semantic big model, characterized by: The system for implementing the complex task scheduling decision method for power scenarios based on a semantic big model as described in any one of claims 1-2 comprises: A task feature determination module, which is used to read task processing records of diversified power scenarios, mine task scenario features, perform feature granularity refinement, separation and clustering processing, and determine key task features; A model building module, the model building module is used to traverse the key task features and build a semantic big model, wherein the semantic big model includes a front-end overall planning layer and a plurality of rear-end parallel task decision layers; A task receiving module, the task receiving module is used to receive a pre-processing power task, the pre-processing power task includes at least one parallel task; A strategy determination module, the strategy determination module is used to combine the semantic big model, perform task granularity refinement and decision layer allocation processing on the pre-processing power task, and determine the task processing strategy; A network determination module, the network determination module is used to identify the task source distribution of the pre-processed power task, correspond to the task processing strategy, and determine the strategy distribution network; A local management module is used to transmit the strategy distribution network to the power management system to perform task local response management.

Images