Operator scheduling method and device for power service scenario, medium and equipment

Abstract

This invention discloses an operator scheduling method, apparatus, medium, and device for power business scenarios, belonging to the field of scheduling. This application acquires power business data and data security level labels, computing node resource status, and trust level information. Based on a preset operator metadata profile, it determines the input / output specifications and resource requirement fingerprints. After preliminary screening, it determines a candidate node set, silently probes idle nodes to obtain actual inference time, inputs it into a multi-dimensional fit scoring model, determines a security mask value based on a comparison of data security level and trust level, comprehensively calculates the fit score, outputs the target execution node, distributes operators and data to complete local scheduling, and achieves efficient adaptive scheduling of heterogeneous operators on the cloud, edge, and terminal. This application effectively solves the problem that existing technologies cannot accurately and efficiently schedule AI operators.

Description

Technical Field

[0001] This invention relates to the field of scheduling, and in particular to an operator scheduling method, apparatus, medium and equipment for power business scenarios. Background Technology

[0002] With the deep application of artificial intelligence technology in the power industry, AI capabilities have been widely introduced into business scenarios such as power material supply chain, power grid operation and maintenance, and on-site safety management. These capabilities include document OCR recognition, demand forecasting and time-series analysis, transportation route optimization and operational calculation, and on-site violation visual recognition. These AI models are typically developed by different teams based on heterogeneous deep learning frameworks such as TensorFlow, PyTorch, and PaddlePaddle, and deployed as independent APIs on cloud or edge computing nodes.

[0003] However, in existing technical solutions, AI models are highly coupled with business code. When business processes change, the entire system needs to be redeveloped, compiled, and deployed, resulting in long development cycles and poor reusability. Furthermore, existing systems lack adaptive scheduling capabilities for heterogeneous computing environments across cloud, edge, and device, and cannot dynamically select the optimal execution node based on real-time network conditions, node load, and data security requirements. This leads to computational lag when forcibly running large models on edge devices, or transmission delays when processing real-time tasks in the cloud, and poses a security risk of unauthorized transmission of confidential data to the public cloud.

[0004] Furthermore, the data formats of heterogeneous AI models differ significantly, and the tensor layouts and normalization methods of input and output are different. The business layer needs to write a lot of format conversion code, which further increases the system complexity and maintenance costs. These shortcomings make it impossible for existing technologies to accurately and efficiently schedule AI operators. Summary of the Invention

[0005] This invention provides an operator scheduling method, apparatus, medium, and equipment for power business scenarios, in order to solve the problem that existing technologies cannot accurately and efficiently schedule AI operators.

[0006] Firstly, this application provides an operator scheduling method for power business scenarios, including:

[0007] Obtain the power business data to be processed and the data security level label of the power business data, and obtain the resource status data and trust level information of each computing node;

[0008] Based on the preset operator metadata profile, the power business data is parsed to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-constructed based on the containerized encapsulation information and standardized interface descriptions of heterogeneous AI operators;

[0009] Based on the resource demand fingerprint and the resource status data, the computing power nodes are initially screened to determine the candidate node set;

[0010] Silently probe the idle nodes in the candidate node set by sending preset virtual input samples to them to obtain actual inference time data;

[0011] The data security level label, the trust level information, the actual inference time data, and the resource status data are input into a preset multidimensional fit scoring model. The multidimensional fit scoring model determines the security mask value based on the comparison result between the data security level label and the trust level information. After calculating the fit score of each candidate node based on the actual inference time data, the resource status data, and the security mask value, the target execution node is output.

[0012] The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives.

[0013] The target operator and the power business data are distributed to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

[0014] This application achieves standardized encapsulation of heterogeneous AI operators by constructing operator metadata profiles, thereby decoupling AI algorithms from business logic and freeing business personnel from concern themselves with underlying framework differences. Preliminary screening using resource requirement fingerprints and real-time resource status quickly narrows down the candidate node range, improving scheduling efficiency. Silent probing of idle nodes using virtual samples obtains actual inference time data to correct static performance benchmarks, addressing real-time environment deviations. A security mask value is determined by comparing data security level labels with trust level information, blocking the transmission path of confidential data to low-trust-level nodes at the scheduling algorithm level, ensuring data security and compliance. A multi-dimensional adaptability score integrating actual inference time, network transmission cost, model accuracy, historical reputation, and node load achieves a comprehensive optimal decision considering security, performance, accuracy, and stability. Finally, the target operator is distributed to the optimal node and local scheduling is executed based on input-output reduction, thus achieving efficient adaptive scheduling of heterogeneous operators in complex cloud-edge-device environments while ensuring data security. This application effectively solves the problem that existing technologies cannot accurately and efficiently schedule AI operators.

[0015] Furthermore, the acquisition of resource status data and trust level information of each computing node specifically includes:

[0016] By deploying resource-aware agents in the cloud and at the edge, the CPU utilization, memory usage, network bandwidth, and network latency data of each computing node are collected at preset intervals.

[0017] Read the preset node attribute files of each computing node to obtain the trust level information;

[0018] The resource status data is obtained by integrating the CPU utilization, memory usage data, network bandwidth data, network latency data, and trust level information.

[0019] This application constructs a real-time resource view across the entire domain by deploying resource-aware agents in the cloud and at the edge to collect CPU utilization, memory usage, network bandwidth, and network latency data of each computing node at preset intervals. This addresses the disconnect between static resource configuration and dynamic operating status. Trust level information is obtained by reading pre-defined node attribute files, incorporating security attributes into the resource status data system. This allows subsequent scheduling decisions to be based on the matching relationship between data security level and node trust level, blocking the transmission path of classified data to untrusted nodes from the data source. By integrating computing resource status and security attributes into unified resource status data, performance awareness and security awareness are organically combined, providing complete decision input for the multi-dimensional adaptability scoring model. This improves the real-time performance and accuracy of heterogeneous operator scheduling while ensuring data security and compliance.

[0020] Furthermore, the preliminary screening of each computing power node based on the resource demand fingerprint and the resource status data to determine the candidate node set specifically involves:

[0021] The resource requirement fingerprint is parsed to obtain the minimum number of CPUs, minimum graphics processing unit memory, and target architecture type required for runtime.

[0022] The minimum number of CPUs is compared with the number of available CPUs in the resource status data, the minimum graphics processing unit memory is compared with the memory usage data in the resource status data, the target architecture type is compared with the actual architecture type of each computing node, computing nodes that do not meet any of the comparison requirements are eliminated, and a candidate node set is determined based on the remaining computing nodes.

[0023] This application analyzes resource requirement fingerprints to extract hard resource constraints such as the minimum number of CPUs, minimum GPU memory, and target architecture type required for runtime, thereby transforming abstract operator runtime requirements into quantifiable screening conditions. By comparing the minimum number of CPUs with the number of available CPUs, the minimum GPU memory and memory usage, and the target architecture type with the actual architecture type, nodes that do not meet the conditions are eliminated from both computing power capacity and hardware compatibility dimensions. This quickly filters invalid candidates in the early stages of scheduling decisions, reducing the computational overhead of subsequent multi-dimensional suitability scoring. By determining the candidate node set based on the remaining computing power nodes, it ensures that all nodes entering the scoring stage have the basic capability to support the target operator, avoiding scheduling failures or execution interruptions due to insufficient resources, thereby improving the scheduling success rate and system operational stability.

[0024] Furthermore, the step of silently probing idle nodes in the candidate node set by issuing preset virtual input samples to obtain actual inference time data specifically involves:

[0025] Monitor the CPU utilization and GPU utilization of each computing node in the candidate node set;

[0026] Nodes with CPU utilization and GPU utilization below a preset idle threshold are marked as idle nodes.

[0027] A preset virtual input sample is sent to the idle state node, triggering the idle state node to perform a complete inference operation;

[0028] Record the actual time taken for the idle state node to complete the inference operation, and use the actual time taken as the actual inference time data.

[0029] This application monitors the CPU and GPU utilization of each computing node in the candidate node set to perceive the node load status in real time. By marking nodes with utilization below a preset idle threshold as idle nodes, it accurately identifies computing resources with detection conditions, avoiding additional performance interference to high-load nodes. By sending preset virtual input samples to idle nodes and triggering complete inference operations, it simulates real computing scenarios without occupying actual business data. By recording the actual time taken for idle nodes to complete inference operations, it obtains dynamic data reflecting the current real performance of the nodes, thereby correcting the deviation between static performance benchmarks and real-time environments. This solves the problem of inaccurate performance prediction caused by node load fluctuations, temperature changes, or resource contention, providing accurate performance input for the multi-dimensional fit scoring model and improving the reliability and timeliness of scheduling decisions.

[0030] Furthermore, the calculation of the suitability score for each candidate node based on the actual inference time data, the resource status data, and the security mask value specifically involves:

[0031] Based on the network bandwidth data and network latency data in the resource status data, calculate the network transmission time of each candidate node;

[0032] Based on the model version of the target operator running at each candidate node, determine the model accuracy index of each candidate node;

[0033] Calculate the historical reputation correction factor based on the degree of deviation between the historical predicted time and the actual time of each candidate node;

[0034] The basic score is calculated based on the actual inference time data, the network transmission time, the model accuracy index, the historical reputation correction factor, and the node load data in the resource status data.

[0035] The fitness score is obtained by multiplying the base score value by the security mask value.

[0036] This application quantifies the time cost of data migration by calculating network transmission time based on network bandwidth and latency data; it incorporates computational accuracy into the scoring dimension by determining model accuracy indicators based on model version; it identifies and penalizes nodes with unstable performance predictions by calculating historical reputation correction factors based on the deviation between historical predicted and actual transmission times, thus avoiding scheduling black holes; it achieves multi-objective optimization of speed, accuracy, stability, and load balancing by comprehensively calculating a basic score value based on actual inference time, network transmission time, model accuracy indicators, historical reputation correction factors, and node load data; and it forcibly blocks the scheduling path of confidential data to low-trust-level nodes from the algorithm's underlying layer by multiplying the basic score value with a security mask value, directly resetting the fit score to zero when the security mask value is zero, thereby ensuring data security and compliance while achieving accurate adaptive scheduling of heterogeneous operators in complex cloud-edge-device environments, significantly improving the comprehensiveness and reliability of scheduling decisions.

[0037] Furthermore, before parsing the power business data based on the preset operator metadata profile, the process also includes:

[0038] Receive an operator orchestration request initiated by a user through a visual canvas; the operator orchestration request includes the selected operator nodes and the data flow relationship between the nodes;

[0039] Based on the data flow relationship, a directed acyclic graph (DAG) of business logic is constructed; the vertices of the DAG correspond to the operator nodes, and the edges of the DAG correspond to the data flow relationship.

[0040] Traverse the directed acyclic graph to check for logical loops. If a loop exists, block the arrangement and return an error message.

[0041] The output reduction of the upstream node of each edge in the directed acyclic graph is matched and verified with the input reduction of the downstream node. If the reductions do not match, the connection is blocked and an error message is returned.

[0042] Based on the directed acyclic graph after successful verification, the target operator and the upstream operator identifier of the target operator are determined.

[0043] This application transforms the construction of complex AI business processes into intuitive drag-and-drop operations by receiving operator orchestration requests initiated by users through a visual canvas, thus lowering the technical threshold for business personnel. It decouples AI operators from business processes by constructing a directed acyclic graph (DAG) of business logic based on data flow relationships, enabling reusable operators and flexible process adjustments. By traversing the DAG to detect logical loops, it blocks circular dependencies during the orchestration phase, avoiding runtime deadlock risks. By matching and validating the output reduction of upstream nodes with the input reduction of downstream nodes, it identifies data format incompatibility issues before deployment, preventing execution failures due to format errors. Based on the validated DAG, it determines the target operator and upstream operator identifiers, providing a clear topological basis for subsequent metadata parsing, resource selection, and scheduling execution. This ensures agile orchestration and efficient iteration of AI applications while guaranteeing the correctness and executability of the business processes.

[0044] Furthermore, the step of distributing the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol, specifically involves:

[0045] The target operator and the power business data are distributed to the target execution node, and the upstream operator identifier of the target operator is obtained;

[0046] Based on the upstream operator identifier, determine whether the upstream operator and the target operator are assigned to the same physical device;

[0047] If allocated on the same physical device, the operation result data of the upstream operator is obtained through shared memory. The upstream operator writes the operation result data into the shared memory area and passes the memory handle. The target execution node reads the operation result data through memory mapping.

[0048] If they are not assigned to the same physical device, the operation result data of the upstream operator is obtained through network communication, and the upstream operator transmits the operation result data to the target execution node through network protocol;

[0049] After the target execution node obtains the computation result data, it checks whether the tensor layout of the computation result data is consistent with the required layout in the input-output reduction. If they are inconsistent, it calls a preset math library to perform transpose and normalization operations to obtain aligned input data.

[0050] The target operator is operated on based on the aligned input data to complete the degree execution of the target operator.

[0051] This application establishes the execution dependency between upstream and downstream operators by distributing the target operator and power business data to the target execution node and obtaining the upstream operator identifier. It dynamically selects the data transmission method based on whether the upstream and target operators are allocated on the same physical device. If on the same device, memory handles are passed via shared memory to avoid data serialization and kernel / user space copy overhead, reducing transmission latency from milliseconds to microseconds. If on different devices, network communication ensures the reliability of cross-node data transmission. It automatically identifies format differences between heterogeneous frameworks by detecting the consistency of the tensor layout and input / output reduction of the computation result data. Transformation and normalization operations are performed using a preset mathematical library to achieve transparent conversion of heterogeneous tensor data, eliminating the burden of manually writing format conversion code at the business layer. Finally, the target operator operation is executed based on the aligned input data to complete the scheduled execution. This improves the data transmission efficiency between heterogeneous operators while achieving seamless compatibility with heterogeneous AI frameworks, significantly reducing system complexity and maintenance costs.

[0052] Secondly, this application provides an operator scheduling device for power business scenarios. The operator scheduling device for power business scenarios includes:

[0053] The acquisition module is used to acquire the power business data to be processed and the data security level label of the power business data, and to acquire the resource status data and trust level information of each computing node.

[0054] The parsing module is used to parse the power business data based on a preset operator metadata profile to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-built based on the containerized encapsulation information and standardized interface description of heterogeneous AI operators;

[0055] The filtering module is used to perform preliminary filtering on each computing power node based on the resource demand fingerprint and the resource status data to determine the candidate node set;

[0056] The detection module is used to silently detect idle nodes in the candidate node set by sending preset virtual input samples to obtain actual inference time data.

[0057] The output module is used to input the data security level label, the trust level information, the actual inference time data, and the resource status data into a preset multidimensional fit scoring model, so that the multidimensional fit scoring model determines the security mask value based on the comparison result of the data security level label and the trust level information, calculates the fit score of each candidate node based on the actual inference time data, the resource status data, and the security mask value, and outputs the target execution node;

[0058] The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives.

[0059] The scheduling module is used to distribute the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

[0060] This application acquires power business data and its data security level tags through an acquisition module, while simultaneously collecting resource status data and trust level information of each computing node, thus providing dual inputs for scheduling decisions: data security attributes and node security capabilities. A parsing module analyzes the power business data based on pre-defined operator metadata profiles, abstracting heterogeneous AI operators into unified input-output specifications and resource requirement fingerprints, achieving decoupling between the algorithm and the business. A filtering module performs preliminary filtering based on resource requirement fingerprints and resource status data, quickly filtering out nodes that do not meet hard constraints and narrowing the candidate range. A detection module silently probes idle nodes by sending virtual input samples. The system acquires actual inference time data to correct static performance benchmarks; through the output module, multidimensional data is input into a preset multidimensional fit scoring model, which determines the security mask value based on the comparison results of data security level and trust level, and calculates the fit score by combining actual inference time, resource status and security mask value, and outputs the target execution node, thereby achieving multidimensional optimization decision-making while ensuring data security and compliance; through the scheduling module, the target operator and power business data are distributed to the target execution node and local execution is triggered, completing the closed-loop scheduling from perception, parsing, filtering, detection, decision-making to execution, significantly improving the scheduling efficiency and security of heterogeneous operators in the complex cloud-edge-device environment.

[0061] Thirdly, this application provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the operator scheduling method for power business scenarios as described above. Its beneficial effects are the same as those of the operator scheduling method for power business scenarios provided in the first aspect of this application.

[0062] Fourthly, this application provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any of the operator scheduling methods for power business scenarios as described in the first aspect. Attached Figure Description

[0063] Figure 1 : A schematic flowchart of an embodiment of the operator scheduling method for power business scenarios provided in this application;

[0064] Figure 2 This is a schematic diagram of an embodiment of the operator scheduling device for power business scenarios provided in this application. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1

[0067] Please refer to Figure 1 In order to solve the problem that existing technologies cannot accurately and efficiently schedule AI operators, this invention provides an operator scheduling method for power business scenarios, including steps S01-S06.

[0068] To ensure the efficient execution of subsequent business processes, two basic preparations were completed before the system went live:

[0069] Standardized encapsulation of heterogeneous AI operators: AI model files (.pth / .pb), inference engines (ONNXRuntime / TensorRT), and dependent libraries from frameworks such as TensorFlow, PyTorch, and PaddlePaddle are packaged into lightweight Docker images to achieve physical isolation of the underlying runtime environment. Simultaneously, a JSON-formatted metadata description file (Profile) is built for each operator, including I / O specifications (MIME type, tensor shape such as shape:[-1,3,640,640], layout format such as NCHW / NHWC), resource fingerprints (minimum number of CPU cores, GPU VRAM, architecture x86 / ARM), performance benchmarks (benchmark inference time T_base and model accuracy Acc under a T4 graphics card), security tags (offline operation support, trust level of third-party code), and scenario preference weights (speed weight coefficient α, accuracy weight coefficient β, emergency repair scenario configuration α>>β, billing and metering scenario configuration β>>α).

[0070] DAG orchestration engine deployment: Provides a visual canvas, allowing users to drag and drop operators to instantiate as nodes, connect them to define data flow, and construct a DAG. A built-in topology verification engine traverses the DAG structure in real time, automatically detecting logical loops. If a loop exists, orchestration is blocked and an error message is returned. I / O specifications of upstream and downstream nodes are matched and verified. The output specification of the upstream node is extracted and compared with the input specification of the downstream node. If the specifications do not match, the connection is blocked and an error message is returned, ensuring the correctness and executability of the business process. Branching and gateway selection / aggregation waiting are supported between node connections to achieve complex business logic orchestration. The verified DAG structure clearly identifies the target operator and upstream operator identifiers, providing a topology basis for real-time scheduling.

[0071] S01: Obtain the power business data to be processed and the data security level label of the power business data, and obtain the resource status data and trust level information of each computing node.

[0072] In a preferred embodiment of this invention, the steps of acquiring the power business data to be processed and the data security level label of the power business data, and acquiring the resource status data and trust level information of each computing node, specifically include:

[0073] When a business system generates a power business processing requirement, the system receives the power business data to be processed (such as power grid inspection images, equipment time series data, scanned copies of material documents, etc.), and parses its data security level label (public / internal / sensitive / confidential, in accordance with the power industry classification standard, such as StateGrid-L1 top secret level).

[0074] Simultaneously, resource-aware agents deployed in the cloud and at the edge collect state vectors of each computing node at a preset interval of 5 seconds, including CPU utilization, GPU utilization, memory usage, network latency, and network bandwidth. The trust level (private cloud / industry cloud / public cloud) is obtained by reading the node's pre-defined attribute file. This trust level indicates the node's trustworthiness and is used for subsequent security hard constraint verification. The collected CPU utilization, GPU utilization, memory usage, network latency, and network bandwidth are integrated into resource status data, with the trust level information independently correlated, providing a data foundation for subsequent multi-dimensional adaptability scoring.

[0075] In addition, based on the pre-prepared DAG orchestration engine, the verified DAG structure has clearly identified the target operator and upstream operator identifiers, and this relationship is used as the basic data input for this step.

[0076] S02: Based on the preset operator metadata profile, the power business data is parsed to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-constructed based on the containerized encapsulation information and standardized interface description of heterogeneous AI operators.

[0077] In a preferred embodiment of this invention, the step of parsing the power business data based on a preset operator metadata profile to determine the input / output reduction and resource requirement fingerprint of the target operator specifically involves:

[0078] The system matches target operators from the operator warehouse based on the business type identifier of the power business data (such as "document OCR recognition"), and reads its JSON format metadata description file (Profile) for standardized parsing:

[0079] Input / output specifications: Define the MIME type (e.g., image / jpeg, application / json), tensor layout (e.g., NCHW, NHWC), shape description (shape: [-1,3,640,640]), and output data structure;

[0080] Resource requirement fingerprint: Extract minimum number of CPU cores, minimum GPU memory (VRAM), and target architecture type (x86 / ARM).

[0081] Performance benchmarks and scenario weights: The benchmark inference time T_base, model accuracy Acc, and scenario preference weights (speed weight coefficient α and accuracy weight coefficient β) are synchronized with the metadata.

[0082] Through this analysis, heterogeneous AI operators are abstracted into a unified interface entity, ensuring scheduling compatibility.

[0083] S03: Based on the resource demand fingerprint and the resource status data, perform preliminary screening of each computing power node to determine the candidate node set.

[0084] In a preferred embodiment of this invention, the preliminary screening of each computing power node based on the resource demand fingerprint and the resource status data to determine the candidate node set specifically involves:

[0085] The system parses the resource requirement fingerprint to obtain hard constraints, and compares them item by item with the resource status data of each computing node:

[0086] CPU core count comparison: The number of available CPU cores in a node is greater than or equal to the minimum number of CPU cores.

[0087] GPU memory comparison: Node remaining GPU memory (VRAM) ≥ minimum GPU memory requirement;

[0088] Architecture comparison: The actual architecture of the node (x86 / ARM) is consistent with the target architecture;

[0089] Nodes that do not meet any of the criteria are removed, and the remaining nodes form a candidate node set N_candidate, narrowing down the scope of subsequent decisions and avoiding execution interruptions due to insufficient resources.

[0090] S04: Send preset virtual input samples to the idle nodes in the candidate node set for silent detection to obtain actual inference time data.

[0091] In a preferred embodiment of this invention, the step of silently probing idle nodes in the candidate node set by issuing preset virtual input samples to obtain actual inference time data specifically involves:

[0092] The system monitors the CPU utilization and GPU utilization of each node in the candidate node set N_candidate, and marks nodes with less than 20% (preset idle threshold) as idle nodes.

[0093] A pre-set virtual input sample (DummyInput) is sent to idle nodes. This sample conforms to the target operator I / O specification, is only 1% the size of the actual business data, and does not generate any valid business output. After receiving the sample, the node silently performs the complete inference operation in the background, and the system records the actual inference time T_real (i.e., the actual inference time data). This actual inference time data reflects the node's true processing capability under the current load, temperature, and resource contention conditions, and is used to correct the deviation of the static baseline time T_base, providing accurate real-time performance input for subsequent multi-dimensional adaptability scoring.

[0094] S05: Input the data security level label, the trust level information, the actual inference time data, and the resource status data into a preset multidimensional fit scoring model, so that the multidimensional fit scoring model determines the security mask value based on the comparison result of the data security level label and the trust level information, calculates the fit score of each candidate node based on the actual inference time data, the resource status data, and the security mask value, and outputs the target execution node;

[0095] The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives.

[0096] In a preferred embodiment of this example, the data security level label, the trust level information, the actual inference time data, and the resource status data are input into a preset multidimensional fit scoring model. This allows the multidimensional fit scoring model to determine a security mask value based on the comparison result between the data security level label and the trust level information. After calculating the fit score for each candidate node based on the actual inference time data, the resource status data, and the security mask value, the model outputs the target execution node. Specifically:

[0097] The multidimensional fit scoring model employs a four-dimensional constraint algorithm of "safety-performance-accuracy-feedback", and the calculation logic is as follows:

[0098]

[0099] Parameter details:

[0100] (Hard security constraint, value 0 or 1): The system parses the data security level label (e.g., StateGrid-L1 Top Secret) and the node's trust level information of the business data packet. If the data security level is higher than the node's trust level... (If the classified data corresponds to a public cloud node), then Security mask Setting it to 0 forcibly denies the node's scheduling eligibility at the algorithm's underlying level, physically blocking the transmission path of classified data to nodes with lower trust levels; if the data's security level is not higher than the node's trust level, then Security... mask Set to 1.

[0101] (Historical reputation correction factor, value from 0.0 to 1.0):

[0102] .

[0103] This factor reflects the "honesty" and stability of a node. If a node has a history of frequently taking 100ms for a prediction but actually taking 500ms, its reputation value will decrease, thus reducing its probability of being selected and avoiding a "scheduling black hole".

[0104] (Network transmission cost): Based on the currently detected bandwidth and delay Calculate the data transmission time: For local edge nodes, .

[0105] T infer This is the actual inference time data. Load k α represents the node load data (i.e., the weighted value of CPU utilization and GPU utilization), where α is the speed weighting coefficient, β is the accuracy weighting coefficient, and γ is the load penalty coefficient.

[0106] (Model accuracy): The accuracy of the model version running on this node (e.g., cloud version 0.99, edge version 0.85).

[0107] Decision distribution: selection The highest node is used as the target execution unit.

[0108] S06: Distribute the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

[0109] In a preferred embodiment of this invention, the step of distributing the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol, specifically involves:

[0110] The system distributes the Docker image of the target operator and power business data to the target execution node, reads the upstream operator identifier of the target operator, and executes scheduling according to the following logic:

[0111] Data transmission method selection:

[0112] For the same physical device (such as an edge box): HTTP / RPC calls are disabled, and POSIX shared memory technology is adopted. The upstream operator writes the operation result data into the preset shared memory area, and only the memory handle and offset are passed. The target execution node reads through memory mapping (mmap), realizing zero-copy transmission and reducing the latency to the microsecond level.

[0113] Different physical devices: Transmission is carried out via the HTTP / 2 protocol to ensure cross-node reliability.

[0114] Automatic alignment of heterogeneous tensors:

[0115] After the target execution node obtains the upstream data, the built-in tensor adaptation layer compares the tensor layout of the operation result data with the required layout in the input and output specifications (such as upstream NHWC / required NCHW). If they are inconsistent, the corresponding library is called according to the hardware type: CPU nodes call the BLAS library to perform transposition, and GPU nodes call the CUDA Kernel to perform transposition and normalization. No adaptation code is required in the business layer.

[0116] Operator execution and result output:

[0117] The target execution node calls the inference engine (ONNXRuntime / TensorRT) to perform target operator operations based on the aligned input data, and outputs business results (such as OCR recognition results and violation detection reports) for downstream nodes or business systems to call, thus completing the scheduling and execution of the target operator on the local node.

[0118] In summary, this application achieves standardized encapsulation of heterogeneous AI operators by constructing operator metadata profiles, thereby decoupling AI algorithms from business logic and freeing business personnel from concern themselves with underlying framework differences. Preliminary screening using resource requirement fingerprints and real-time resource status quickly narrows down the candidate node range, improving scheduling efficiency. Silent probing of idle nodes using virtual samples obtains actual inference time data to correct static performance benchmarks, addressing real-time environment deviations. A security mask value is determined by comparing data security level labels with trust level information, blocking the transmission path of confidential data to low-trust-level nodes at the scheduling algorithm level, ensuring data security and compliance. A multi-dimensional adaptability score integrating actual inference time, network transmission cost, model accuracy, historical reputation, and node load achieves a comprehensive optimal decision considering security, performance, accuracy, and stability. Finally, the target operator is distributed to the optimal node and local scheduling is executed based on input-output reduction, thus achieving efficient adaptive scheduling of heterogeneous operators in complex cloud-edge-device environments while ensuring data security. This application effectively solves the problem that existing technologies cannot accurately and efficiently schedule AI operators.

[0119] Example 2

[0120] Please refer to Figure 2 This is an operator scheduling device for power business scenarios provided in the embodiments of this application.

[0121] In this embodiment, the operator scheduling device for power business scenarios includes an acquisition module 10, a parsing module 20, a filtering module 30, a detection module 40, an output module 50, and a scheduling module 60.

[0122] The acquisition module 10 is used to acquire the power business data to be processed and the data security level label of the power business data, and to acquire the resource status data and trust level information of each computing node.

[0123] The parsing module 20 is used to parse the power business data based on a preset operator metadata profile to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-constructed based on the containerized encapsulation information and standardized interface descriptions of heterogeneous AI operators;

[0124] The filtering module 30 is used to perform preliminary filtering on each computing power node based on the resource demand fingerprint and the resource status data to determine a set of candidate nodes;

[0125] The detection module 40 is used to silently detect idle nodes in the candidate node set by sending preset virtual input samples to obtain actual inference time data.

[0126] Output module 50 is used to input the data security level label, the trust level information, the actual inference time data and the resource status data into a preset multidimensional fit scoring model, so that the multidimensional fit scoring model determines the security mask value based on the comparison result of the data security level label and the trust level information, calculates the fit score of each candidate node based on the actual inference time data, the resource status data and the security mask value, and outputs the target execution node;

[0127] The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives.

[0128] The scheduling module 60 is used to distribute the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

[0129] For ease of description and brevity, the embodiments of the device of the present invention include all the implementation methods in the above embodiments of the operator scheduling method for power business scenarios, and will not be repeated here.

[0130] Example 3

[0131] This application provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the operator scheduling method for power business scenarios.

[0132] The operator scheduling method for power business scenarios, when implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0133] Example 4

[0134] This embodiment provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the operator scheduling methods for power business scenarios as described in Embodiment 1.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. An operator scheduling method for power business scenarios, characterized in that, include: Obtain the power business data to be processed and the data security level label of the power business data, and obtain the resource status data and trust level information of each computing node; Based on the preset operator metadata profile, the power business data is parsed to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-constructed based on the containerized encapsulation information and standardized interface descriptions of heterogeneous AI operators; Based on the resource demand fingerprint and the resource status data, the computing power nodes are initially screened to determine the candidate node set; Silently probe the idle nodes in the candidate node set by sending preset virtual input samples to them to obtain actual inference time data; The data security level label, the trust level information, the actual inference time data, and the resource status data are input into a preset multidimensional fit scoring model, so that the multidimensional fit scoring model determines a security mask value based on the comparison result of the data security level label and the trust level information; and based on the security mask value, the fit score of each candidate node is calculated according to the actual inference time data and the resource status data, and the target execution node is output. The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives. The target operator and the power business data are distributed to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

2. The operator scheduling method for power business scenarios according to claim 1, characterized in that, The acquisition of resource status data and trust level information for each computing node specifically includes: By deploying resource-aware agents in the cloud and at the edge, the CPU utilization, memory usage, network bandwidth, and network latency data of each computing node are collected at preset intervals. Read the preset node attribute files of each computing node to obtain the trust level information; The resource status data is obtained by integrating the CPU utilization, memory usage data, network bandwidth data, network latency data, and trust level information.

3. The operator scheduling method for power business scenarios according to claim 1, characterized in that, The preliminary screening of each computing power node based on the resource demand fingerprint and the resource status data to determine the candidate node set is specifically as follows: The resource requirement fingerprint is parsed to obtain the minimum number of CPUs, minimum graphics processing unit memory, and target architecture type required for runtime. The minimum number of CPUs is compared with the number of available CPUs in the resource status data, the minimum graphics processing unit memory is compared with the memory usage data in the resource status data, the target architecture type is compared with the actual architecture type of each computing node, computing nodes that do not meet any of the comparison requirements are eliminated, and a candidate node set is determined based on the remaining computing nodes.

4. The operator scheduling method for power business scenarios according to claim 1, characterized in that, The step of silently probing idle nodes in the candidate node set by issuing preset virtual input samples to obtain actual inference time data specifically involves: Monitor the CPU utilization and GPU utilization of each computing node in the candidate node set; Nodes with CPU utilization and GPU utilization below a preset idle threshold are marked as idle nodes. A preset virtual input sample is sent to the idle state node, triggering the idle state node to perform a complete inference operation; Record the actual time taken for the idle state node to complete the inference operation, and use the actual time taken as the actual inference time data.

5. The operator scheduling method for power business scenarios according to claim 1, characterized in that, The process of calculating the suitability score for each candidate node based on the actual inference time data, the resource status data, and the security mask value is as follows: Based on the network bandwidth data and network latency data in the resource status data, calculate the network transmission time of each candidate node; Based on the model version of the target operator running at each candidate node, determine the model accuracy index of each candidate node; Calculate the historical reputation correction factor based on the degree of deviation between the historical predicted time and the actual time of each candidate node; The basic score is calculated based on the actual inference time data, the network transmission time, the model accuracy index, the historical reputation correction factor, and the node load data in the resource status data. The fitness score is obtained by multiplying the base score value by the security mask value.

6. The operator scheduling method for power business scenarios according to claim 1, characterized in that, Before parsing the power business data based on the preset operator metadata profile, the process also includes: Receive an operator orchestration request initiated by a user through a visual canvas; the operator orchestration request includes the selected operator nodes and the data flow relationship between the nodes; Based on the data flow relationship, a directed acyclic graph (DAG) of business logic is constructed; the vertices of the DAG correspond to the operator nodes, and the edges of the DAG correspond to the data flow relationship. Traverse the directed acyclic graph to check for logical loops. If a loop exists, block the arrangement and return an error message. The output reduction of the upstream node of each edge in the directed acyclic graph is matched and verified with the input reduction of the downstream node. If the reductions do not match, the connection is blocked and an error message is returned. Based on the directed acyclic graph after successful verification, the target operator and the upstream operator identifier of the target operator are determined.

7. The operator scheduling method for power business scenarios according to claim 1, characterized in that, The step of distributing the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol, specifically involves: The target operator and the power business data are distributed to the target execution node, and the upstream operator identifier of the target operator is obtained; Based on the upstream operator identifier, determine whether the upstream operator and the target operator are assigned to the same physical device; If allocated on the same physical device, the operation result data of the upstream operator is obtained through shared memory. The upstream operator writes the operation result data into the shared memory area and passes the memory handle. The target execution node reads the operation result data through memory mapping. If they are not assigned to the same physical device, the operation result data of the upstream operator is obtained through network communication, and the upstream operator transmits the operation result data to the target execution node through network protocol; After the target execution node obtains the computation result data, it checks whether the tensor layout of the computation result data is consistent with the required layout in the input-output reduction. If they are inconsistent, it calls a preset math library to perform transpose and normalization operations to obtain aligned input data. The target operator is operated on based on the aligned input data to complete the scheduled execution of the target operator.

8. An operator scheduling device for power business scenarios, characterized in that, include: The acquisition module is used to acquire the power business data to be processed and the data security level label of the power business data, and to acquire the resource status data and trust level information of each computing node. The parsing module is used to parse the power business data based on a preset operator metadata profile to determine the input-output specifications and resource requirement fingerprints of the target operator; wherein, the operator metadata profile is pre-built based on the containerized encapsulation information and standardized interface description of heterogeneous AI operators; The filtering module is used to perform preliminary filtering on each computing power node based on the resource demand fingerprint and the resource status data to determine the candidate node set; The detection module is used to silently detect idle nodes in the candidate node set by sending preset virtual input samples to obtain actual inference time data. The output module is used to input the data security level label, the trust level information, the actual inference time data, and the resource status data into a preset multidimensional fit scoring model, so that the multidimensional fit scoring model determines a security mask value based on the comparison result of the data security level label and the trust level information; and based on the security mask value, calculates the fit score of each candidate node according to the actual inference time data and the resource status data, and outputs the target execution node. The multidimensional fit scoring model is trained based on historical scheduling data and preset multidimensional optimization objectives. The scheduling module is used to distribute the target operator and the power business data to the target execution node, so that the target execution node can schedule the target operator based on the input-output protocol.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the operator scheduling method for power business scenarios as described in any one of claims 1 to 7.

10. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the operator scheduling method for power business scenarios as described in any one of claims 1 to 7.

Images