Methods, systems, devices, and media for inter-core task scheduling optimization based on multi-core controllers

Abstract

This application belongs to the field of wind turbine generator control technology, and particularly relates to a method, system, device, and medium for optimizing inter-core task scheduling based on a multi-core controller. The method includes: dividing the wind turbine generator into several core control tasks according to its operational requirements, determining corresponding task scheduling units, and constructing a data dependency graph; monitoring the real-time operating status parameters of the wind turbine generator and the current load status of each processing core of the multi-core controller, and determining whether the performance of the processing core is abnormal based on the current load status and fault indication signals to trigger fault recovery scheduling, or acquiring dynamic sensing information; inputting the data dependency graph and dynamic sensing information into a scheduling decision model of a state-aware graph neural network to calculate and generate an initial task allocation scheme; executing the initial task allocation scheme to obtain performance feedback, and adjusting the scheduling weight parameters based on the performance feedback to update the initial task allocation scheme. This application improves the efficiency of wind turbine generator task scheduling.

Description

Technical Field

[0001] This application belongs to the field of wind turbine generator control technology, and in particular relates to a method, system, equipment and medium for inter-core task scheduling optimization based on a multi-core controller. Background Technology

[0002] As wind turbines become larger and more intelligent, main control systems are increasingly adopting multi-core controllers to enhance parallel processing capabilities. However, the currently commonly used fixed-core scheduling scheme (which statically binds different control tasks to specific CPU cores) has significant limitations in terms of flexibility and real-time performance. For example, in traditional schemes, each control task is typically pre-assigned to run on a specific core, without considering differences in core performance or runtime load variations during scheduling. This static scheduling lacks awareness and adaptability to heterogeneous multi-core environments, resulting in underutilization of computing resources.

[0003] Existing multi-core task scheduling algorithms have failed to address this issue. For example, some studies have used offline analysis to select the "best" kernel for a task, but their effectiveness is limited because they cannot account for dynamic changes during runtime. Traditional real-time operating systems also tend to reduce task migration to avoid cache misses, but this can lead to some kernels being overloaded while others remain idle. Furthermore, if tasks frequently migrate between multiple kernels, cache hit rates will decrease due to cache data reloading, which in turn negatively impacts system performance.

[0004] On the other hand, inter-core communication latency is also a concern in multi-core scheduling. In wind turbine main control systems, different tasks often run on different cores and require frequent data exchange. If the scheduling algorithm does not optimize inter-core communication, the end-to-end latency of short message transmission may be high, which can severely reduce real-time control performance. Reports indicate that some existing heterogeneous multi-core scheduling algorithms (such as the HEFT algorithm) do not consider the cost of inter-task communication. Neglected communication overhead can lead to data desynchronization or increased latency, hindering real-time monitoring and control of wind turbines. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method, system, device, and medium for optimizing inter-core task scheduling based on a multi-core controller. By combining dynamic perception with intelligent decision-making, it achieves efficient scheduling of inter-core tasks for the multi-core controller, which can significantly improve the real-time performance and resource utilization of the wind turbine main control system.

[0006] Firstly, this application provides a method for optimizing inter-core task scheduling based on a multi-core controller, the method comprising: Based on the operational requirements of the wind turbine, several core control tasks are divided, and several corresponding task scheduling units are determined based on these core control tasks. A data dependency graph between these task scheduling units is then constructed. Monitor the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller, and determine whether the performance of the processing core is abnormal based on the current load status and fault indication signals, so as to trigger fault recovery scheduling according to the judgment result, or obtain dynamic sensing information. The data dependency graph and the dynamic perception information are input into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme. The initial task allocation scheme is executed to obtain performance feedback. Based on the performance feedback, the scheduling weight parameters corresponding to the task scheduling unit are adjusted to update the initial task allocation scheme.

[0007] Furthermore, Constructing a data dependency graph among several task scheduling units, specifically including: The task scheduling unit is used as the initial task node to form a node set; Based on the data flow relationship between tasks, define the directed dependency edges between the initial task nodes to form a set of directed edges; Based on the set of nodes and the set of directed edges, a task-dependent directed graph and its corresponding adjacency matrix are generated. For each of the initial task nodes, an initial node feature vector is constructed that includes at least one of the following: task computation requirements, task cycle, deadline, and scheduling priority. An initial node feature matrix is ​​then formed based on the initial node feature vector. The data dependency graph is obtained based on the task dependency directed graph, the corresponding adjacency matrix, and the initial node feature matrix.

[0008] Furthermore, Determine whether the core performance is abnormal based on the current load status, specifically including: Detect the current task execution rate and current health status value of any processing core, and determine the current load status based on the current task execution rate and current health status value; The current load state is compared with a preset load state threshold, and the presence of fault indication signals is monitored. Based on the comparison results and the fault indication signal monitoring results, it is determined whether the core processing performance is abnormal.

[0009] Furthermore, Based on the judgment result, fault recovery scheduling is triggered, or dynamic perception information is obtained, specifically including: When the current load state corresponding to any processing core is less than or equal to the preset load state threshold, or when a fault indication signal exists, fault recovery scheduling is triggered to transfer the task scheduling unit on the abnormal core and ensure that the status data is consistent. When the current load status is greater than the preset load status threshold and there is no fault indication signal, dynamic sensing information is obtained based on the real-time operating status parameters and the current load status.

[0010] Furthermore, The task scheduling unit on the exception core is transferred, and state data consistency is ensured. Specifically, this includes: Identify the abnormal core and the normal processing core, and determine the set of target task scheduling units on the abnormal core; Analyze the current load status of the normal processing core, and select the target migration processing core based on the current load status; Through a communication synchronization mechanism, the status data of the task scheduling unit set on the anomaly core is copied to the target migration processing core in real time.

[0011] Furthermore, The calculation generates an initial task allocation scheme, specifically including: The data dependency graph and the dynamic perception information are input into the state perception graph neural network; The initial node feature matrix is ​​linearly transformed by a trainable weight matrix to obtain an initial feature representation. The initial feature representation and the task-dependent directed graph are aggregated by a graph convolutional layer, and the initial relationship weights between the initial task nodes are calculated using an attention mechanism. Based on the initial relation weights, the features of the neighbor nodes of any initial task node are weighted and summed to obtain the neighbor information; The neighbor information is fused with its corresponding initial feature representation, and then processed by a nonlinear activation function to output the updated node embedding feature vector. The node embedding feature vector of each initial task node is input into the fully connected layer and mapped to an output vector of the same dimension as the number of processing cores; Apply the softmax function to any of the output vectors to calculate the probability distribution of the corresponding task across all processing cores. Based on the probability distribution of all tasks on each processing core, a probability sampling method is used to determine the initial core to which any task scheduling unit belongs, forming an initial mapping relationship between tasks and cores, i.e., the initial task allocation scheme.

[0012] Furthermore, Adjusting the scheduling weight parameters corresponding to the task scheduling unit based on performance feedback to update the initial task allocation scheme specifically includes: The initial task allocation scheme is embedded and pooled, and then concatenated with the real-time utilization rate, real-time wind conditions and real-time core power of each processing core to construct the environmental state vector at the current moment. The environmental state vector is input into the policy network to obtain the score of each candidate scheduling action. After filtering illegal actions by combining the mask formed by task dependency and core state, the action probability distribution is calculated and sampled to generate the mapping relationship between tasks and core and execute scheduling. After the scheduling is executed, the system's performance indicators are collected. The performance indicators are normalized and compared based on preset indicator thresholds, the weights of each indicator are adaptively adjusted, and the instant reward is calculated in a multi-objective weighted manner. The advantage function is calculated using the temporal difference and generalized advantage estimation algorithm, and the parameters of the policy network and value network are updated based on the policy gradient method, combined with the immediate reward and the advantage function. The advantages and contributions of each task in historical scheduling decisions are statistically analyzed. The scheduling weight parameters corresponding to each task are updated using the exponential moving average method, and these parameters are used as priority biases to introduce into subsequent strategy decisions. The optimization is considered to have converged when the cumulative reward improvement is lower than the threshold or the key performance indicators continue to meet the preset conditions.

[0013] Secondly, based on the same inventive concept, this application provides an inter-core task scheduling optimization system based on a multi-core controller, the system comprising: A construction module is used to divide several core control tasks according to the operation requirements of the wind turbine, determine several corresponding task scheduling units based on the several core control tasks, and construct a data dependency graph between the several task scheduling units. The judgment module is used to monitor the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller. Based on the current load status and fault indication signals, it judges whether the performance of the processing core is abnormal, and triggers fault recovery scheduling according to the judgment result, or obtains dynamic sensing information. The generation module is used to input the data dependency graph and the dynamic perception information into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme. An update module is used to execute the initial task allocation scheme to obtain performance feedback, and adjust the scheduling weight parameters corresponding to the task scheduling unit based on the performance feedback to update the initial task allocation scheme.

[0014] Thirdly, this application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of any of the multi-core controller-based inter-core task scheduling optimization methods described above.

[0015] Fourthly, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the inter-core task scheduling optimization methods based on a multi-core controller as described above.

[0016] Compared with the prior art, this application has the following advantages: 1. This application constructs a data dependency graph between tasks and utilizes a state-aware graph neural network for deep feature extraction, enabling it to accurately capture the complex data flow and logical dependencies between core control tasks of wind turbine units. Simultaneously, by combining dynamic sensing information, the scheduling model can make precise decisions based on the current system environment.

[0017] 2. This application collects performance indicators after executing the initial scheme and performs normalized comparison and adaptive weighting based on indicator thresholds, which can transform real-time operational feedback into quantifiable immediate rewards. By combining temporal difference and generalized advantage estimation algorithms to update the policy network, the scheduling model can continuously iterate and update the scheduling weight parameters of tasks according to the performance degradation or operating condition drift of wind turbines during long-term operation until it converges to the optimal policy, thus avoiding the performance degradation of traditional static scheduling strategies caused by environmental changes.

[0018] 3. This application can quickly and accurately identify abnormal cores by real-time monitoring of the task execution rate and health status value of each processing core and comparing them with preset thresholds and fault indication signals. When fault recovery scheduling is triggered, it can not only transfer the task units on the abnormal core, but also ensure the strict consistency of status data during the migration process through a communication synchronization mechanism, effectively preventing control task interruption or data loss caused by core failure.

[0019] 4. When generating the initial allocation scheme, this application utilizes an attention mechanism to dynamically calculate the relationship weights between task nodes, accurately capturing the coupling degree between heterogeneous tasks. During the scheme update phase, a multi-objective weighted approach is used to comprehensively consider key performance indicators such as real-time utilization, core power, and wind conditions, and a priority bias based on historical advantage contributions is introduced into the strategy decision-making process. This design breaks through the limitations of single-objective optimization, enabling a balanced optimization of system resource utilization and energy consumption while ensuring task real-time performance and dependencies.

[0020] 5. After the policy network outputs scores for each candidate action, this application combines task dependencies with a mask constructed from the current core state to pre-filter out all illegal actions that violate physical constraints or system rules, ensuring the generated scheduling scheme is executable on a real multi-core controller. Based on this, a probabilistic sampling method is used for task allocation, leveraging the model's experience while retaining the ability to explore potentially better solutions, effectively preventing the algorithm from getting trapped in local optima.

[0021] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating an inter-core task scheduling optimization method based on a multi-core controller according to an embodiment of this application is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Figure 1 A flowchart illustrating an inter-core task scheduling optimization method based on a multi-core controller according to an embodiment of this application is shown, as follows: Figure 1 As shown in the figure, the inter-core task scheduling optimization method based on a multi-core controller according to an embodiment of this application includes, S1. Divide the wind turbine into several core control tasks according to the operation requirements of the wind turbine, determine several corresponding task scheduling units based on the several core control tasks, and construct a data dependency graph between the several task scheduling units. In the embodiments of this application, several core control tasks include at least pitch regulation, converter control, unit status monitoring, wind speed prediction, and power smoothing.

[0026] In this embodiment of the application, step S1 specifically includes: S11, the task scheduling unit is used as the initial task node to form a node set; S12, Based on the data flow relationship between tasks, define the directed dependency edges between the initial task nodes to form a set of directed edges; S13, Based on the set of nodes and the set of directed edges, generate a task-dependent directed graph and the corresponding adjacency matrix; S14, and construct an initial node feature vector for any of the initial task nodes, which includes at least one of the following: task computation requirements, task cycle, deadline and scheduling priority, and form an initial node feature matrix based on the initial node feature vector; S15, the data dependency graph is obtained based on the task dependency directed graph, the corresponding adjacency matrix, and the initial node feature matrix.

[0027] In this embodiment of the application, forming a node set specifically includes: Define the set of scheduling units for core control tasks: ,in, For the complete set of task scheduling units, For pitch adjustment, For converter control, For unit condition monitoring, For wind speed prediction, For power smoothing.

[0028] In this embodiment of the application, the set of directed edges is specifically comprised of: Based on the data flow relationships between tasks, i.e., providing a set of dependency pairs indicating "who provides data to whom," edges are formed accordingly. ,in, Let be a set of dependent pairs (directed), representing the output of the source task as the input of the destination task. Indicates wind speed prediction output Used for pitch adjustment; Indicates the given pitch value Enter converter control; Indicates active / reactive output Provides power smoothing.

[0029] In this embodiment of the application, the construction of the task dependency graph and adjacency matrix, and the provision of the minimum initial task node features, include: ; ; ; In the formula, , is a task-dependent directed graph, where The set of nodes in the graph is equal to , Let be a set of directed edges, equal to ; For elements of the adjacency matrix, if ,but ,otherwise ,in, This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. This represents the i-th task scheduling unit (i.e., the core control task), such as pitch regulation, converter control, etc. This represents the k-th task scheduling unit (i.e., the core control task), such as pitch regulation, converter control, etc.

[0030] Number of tasks (in this embodiment) ); The initial node feature matrix (a minimum of 4 dimensions is sufficient to drive subsequent scheduling); For the first The minimum eigenvector of each task, where... For the first The computational requirements of each task (e.g., the number of instructions / milliseconds within the estimated cycle). For the first The task cycle (e.g., milliseconds) of each task. Deadline (unit: ) Consistent) For the first The scheduling priority of each task (the larger the value, the more critical it is).

[0031] S2 monitors the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller. Based on the current load status and fault indication signals, it determines whether the performance of the processing core is abnormal, and triggers fault recovery scheduling according to the judgment result, or obtains dynamic sensing information. In this embodiment of the application, the real-time operating status parameters include at least real-time wind speed, real-time power generation, and real-time blade angle.

[0032] In this embodiment of the application, step S2 specifically includes: S21, detect the current task execution rate and current health status value of any processing core, and determine the current load status based on the current task execution rate and current health status value; S22, compare the current load state with the preset load state threshold, and monitor whether there is a fault indication signal. Based on the comparison result and the fault indication signal monitoring result, determine whether the core processing performance is abnormal.

[0033] In this embodiment of the application, when any abnormality is detected in any processing core, the fault triggering judgment condition is met: in, This represents the performance metrics (such as task execution rate or health status value) of the nth processing core. The preset load state threshold, It is the fault indication function of the nth processing core (when the core... When malfunction (The value is 0 during normal operation). Once the above condition is true (i.e., a core's performance is below the threshold or a failure occurs), the fault recovery scheduling mechanism is immediately triggered.

[0034] In this embodiment of the application, step S2 further includes: S23, when the current load state corresponding to any processing core is less than or equal to the preset load state threshold, or when there is a fault indication signal, trigger fault recovery scheduling, transfer the task scheduling unit on the abnormal core, and ensure that the status data is consistent. S24, when the current load state is greater than the preset load state threshold and there is no fault indication signal, dynamic sensing information is obtained based on the real-time operating status parameters and the current load state.

[0035] In this embodiment of the application, in the multi-core control system of the wind turbine, operating parameters such as wind speed, power generation, and blade angle are acquired in real time, as well as the current load status of each processing core of the multi-core controller, to construct the dynamic state vector of the system. Assume the system has... There are 1 processing core (core number n=1,2,...M), then in time... The state vector can be represented as: in Let be the real-time wind speed at the current time t. This represents the real-time power generation at the current time t. L represents the real-time blade angle at the current time t. n (t) represents the current load (such as CPU utilization or task queue length) of the nth processing core at time t. This dynamically sensed information vector It updates in real time and serves as input for task scheduling decisions. The scheduling system based on... Dynamically assess the operating status of each core load and wind turbine to provide a basis for subsequent task allocation.

[0036] In this embodiment of the application, step S23 specifically includes: S231, Identify the abnormal core and the normal processing core, and determine the set of target task scheduling units on the abnormal core; S232, Analyze the current load status of the normal processing core, and select the target migration processing core based on the current load status; S233, through a communication synchronization mechanism, the status data of the task scheduling unit set on the anomaly core is copied to the target migration processing core in real time.

[0037] In this embodiment of the application, the processing core with the optimal current load state is selected as the target migration processing core.

[0038] In this embodiment of the application, the set of tasks on the abnormal core a is denoted as New scheduling mapping after fault recovery Map these tasks to the core health set A certain health core : This means that for each task k originally executed by the faulty core a, under fault recovery scheduling, it will be migrated to a normally functioning target migration processing core j for execution. To ensure the consistency of task execution state after migration, the system uses a communication synchronization mechanism to transmit the state data of task k from the old core before migration to the target migration processing core. For example, using... If the state data of task k on the abnormal core a is represented, then the target is migrated to the processing core through synchronous replication. Task status At the moment the fault occurred equal: Where t is the fault trigger time. and These represent the instants before and after the migration. The aforementioned communication synchronization ensures seamless transitions between task contexts, but it also incurs some communication overhead. (For example, the amount of state data transmitted) will be taken into account during scheduling optimization.

[0039] S3, input the data dependency graph and the dynamic perception information into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme; In this embodiment of the application, step S3 specifically includes: S31, input the data dependency graph and the dynamic perception information into the state perception graph neural network; S32, the initial node feature matrix is ​​linearly transformed by a trainable weight matrix to obtain an initial feature representation. The initial feature representation and the task-dependent directed graph are aggregated by a graph convolutional layer, and the initial relationship weights between the initial task nodes are calculated using an attention mechanism. S33, based on the initial relation weights, perform a weighted summation of the features of the neighbor nodes of any initial task node to aggregate and obtain the neighbor information; S34, the neighbor information is fused with its corresponding initial feature representation, processed by a nonlinear activation function, and the updated node embedding feature vector is output. S35, the node embedding feature vector of each initial task node is input into the fully connected layer and mapped to an output vector with the same dimension as the number of processing cores; S36. Apply the softmax function to any of the output vectors to calculate the probability distribution of the corresponding task on all processing cores. S37. Based on the probability distribution of all tasks on each processing core, the initial core to which any task scheduling unit belongs is determined by a probability sampling method, forming an initial mapping relationship between tasks and cores, i.e., the initial task allocation scheme.

[0040] In this embodiment of the application, the construction of the input features of the state-aware graph neural network model includes the following steps: Step 1: Represent the multi-core task scheduling state of the wind power main control system as a graph structure, determine the set of task nodes and the edges between nodes, with each task corresponding to one node, and the edges between nodes established according to the task dependency relationship; Step 2: Construct the adjacency matrix Represents the topological connections between the nodes, where A y,z =1 when task node y and task node z are connected by an edge; otherwise, A y,z =0; Step 3: Obtain the attribute parameters of each task node and construct the initial node feature matrix. The size of the matrix is Each node y corresponds to a feature vector h. y The feature vector includes at least the task's computational requirements, priority, and deadline information.

[0041] In this embodiment of the application, the graph convolution operation of the state-aware graph neural network model includes the following steps: Step 1: Convert the node feature matrix Multiply by the weight matrix , recorded as To obtain the transformed node feature representation; Step 2: For node y and its neighbor node z, calculate the unnormalized attention score. ,in This is a trainable attention parameter vector. This indicates vector concatenation; Step 3: Check all neighbors of node y The attention scores are normalized using the softmax function to obtain the attention weights. , where N(y) is the set of neighboring nodes of node y.

[0042] In this embodiment of the application, the embedding and aggregation step of the state-aware graph neural network model includes the following steps: Step 1: Perform weighted summation and aggregation of the feature vectors of neighboring nodes: For each node y, calculate the transformed feature vector (h) of each of its neighboring nodes z. z W) and the corresponding attention weight a yz Multiply and sum to obtain the aggregate vector of node y. ; Step 2: Aggregate vector m y Input nonlinear activation function (e.g., ReLU) to obtain the updated embedding feature vector of node y. .

[0043] In this embodiment, the scheduling output of the state-aware graph neural network model is given in the form of a task-core allocation probability distribution, including the following steps: Step 1: For each task node, map its final embedded feature vector to a vector of length [length missing]. The output vector ,in The number of cores in a multi-core processor. Indicates task Assigned to the Core scoring values; Step 2: Process the output vector Applying the softmax function, we obtain the task. exist The probability vector assigned to each core ,in And satisfy .

[0044] In this embodiment, the reward function of the reinforcement learning of the scheduling decision model of the state-aware graph neural network The definition is as follows: Step 1: Determine the target items for scheduling optimization, including at least task completion time. and core load balancing ,in This represents the total time or average completion time for all tasks. This indicates the degree of load imbalance among the cores of the multi-core processor; Step 2: Define the instant reward value based on the target item. ,in and This is a weighting parameter used to adjust the contribution weight of each objective item to the reward; Step 3: The reinforcement learning process aims to maximize the cumulative reward, thereby minimizing the weighted objective value mentioned above.

[0045] S4, execute the initial task allocation scheme to obtain performance feedback, and adjust the scheduling weight parameters corresponding to the task scheduling unit based on the performance feedback to update the initial task allocation scheme.

[0046] In this embodiment of the application, step S4 specifically includes: S41, the initial task allocation scheme is embedded and pooled, and then concatenated with the real-time utilization rate, real-time wind conditions and real-time core power of each processing core to construct the environmental state vector at the current moment. S42, input the environment state vector into the policy network to obtain the score of each candidate scheduling action, filter illegal actions by combining the mask formed by task dependency and core state, calculate the action probability distribution and sample it, generate the mapping relationship between tasks and core and execute scheduling. S43, after executing the scheduling, collect the system's performance indicators; S44, based on preset indicator thresholds, normalize and compare the performance indicators, adaptively adjust the weights of each indicator, and calculate the immediate reward in a multi-objective weighted manner. S45, the advantage function is calculated using the temporal difference and generalized advantage estimation algorithm, and the parameters of the policy network and value network are updated based on the policy gradient method, combined with the instantaneous reward and the advantage function. S46. Calculate the contribution of each task to the historical scheduling decision, update the scheduling weight parameters of each task using the exponential moving average method, and introduce them as priority bias into subsequent strategy decisions. When the cumulative reward increase is lower than the threshold or the key performance indicators continue to meet the preset conditions, it is determined to be optimization convergence.

[0047] In this application embodiment, the embodiment of S4 specifically includes: Step 1 (State Construction): Construct the environment state vector s at time t. t , represented as ;in, The pooling vector for task embedding output by the state-aware graphical neural network model, denoted as... , It is the node embedding feature vector of task i after being updated by the graph neural network, u t To process the core utilization vector, Where uj,t represents the load (such as CPU utilization or task queue length) of core j at time t; v t This is a wind condition vector, containing real-time external environmental information such as wind speed, wind direction, and turbulence intensity. ;p t For the core power vector, .

[0048] Step 2 (Strategy Output and Action Selection): ... t Input policy network F θ Obtain the scoring vector z t And calculate the action probability distribution. ; where m t For the action mask, elements corresponding to illegal scheduling actions are set to 0, and ⊕ indicates element-wise addition; action a is obtained by probability sampling. t Based on this, the task-core mapping matrix M is generated. t To satisfy the condition that for each task i there is , of which M ij,t This indicates whether task i is assigned to core j at time t (usually 0 or 1, and the sum of each row is 1, meaning each task can only be assigned to one core).

[0049] Step 3 (Execution and Metric Collection): Execute Mapping M t The set of performance metrics at time t is obtained later [d] t, c t, e t, b t Among them, the delay index Communication metrics Energy consumption indicators ,in Load imbalance index ,in .

[0050] Step 4 (Reward Calculation and Goal Balancing): Calculate immediate rewards using a multi-goal weighted approach. And based on a reference threshold, the weights are adaptively balanced, specifically by normalizing the index ratio vector. The weight vector is updated component by component. ,

[0051] Step 5 (Dominance Estimation): Calculate the time series difference and dominance function, specifically as follows: ; ; where γ is the discount factor and τ is the generalized advantage estimation coefficient.

[0052] Step 6 (Policy Objective and Entropy Regularization): Apply policy objective and entropy regularization in the form of truncated probability ratios, specifically: probability ratios ; intercept the target Value loss G t For discounted returns; entropy term Overall goal .

[0053] Step 7 (Parameter Update): Update the parameters according to the gradient ascent and descent rules, specifically: update the strategy parameters according to the gradient ascent. Update value parameters using gradient descent. ; the weight parameter vector And perform range cropping to ensure that each element is located in (Refined update of scheduling weight parameters): Update the task-level scheduling weight vector. Perform advantage-driven adaptive updates, specifically by: statistically analyzing the advantage contribution of task i at time t. — Index updates via sliding. And cut to Decision cycle, As a priority bias added to the strategy score, denoted as .

[0054] Step 8 (Refining and updating scheduling weight parameters): Update the task-level scheduling weight vector Perform advantage-driven adaptive updates, specifically by: statistically analyzing the advantage contribution of task i at time t. Index updates via sliding. And cut to Decision cycle, As a priority bias added to the strategy score, denoted as .

[0055] Step 9 (Rounding and Convergence): Repeat steps 1 through 8 until any convergence condition is met, specifically: K consecutive rounds. Improvement less than the threshold ; or — the index vector [d,e] simultaneously satisfies and Continuous T hold One cycle.

[0056] Step 10 (Online Deployment): During online operation, the strategy is executed according to steps 1 to 4, and parameters are fine-tuned at a lower frequency according to steps 5 to 8 to ensure real-time adaptive scheduling and performance-energy balance under different operating conditions such as strong gusts, low wind speed startup, and grid disturbances; correspondingly, the online execution and update frequency meets the requirements. .

[0057] Based on the above method, this application also provides an inter-core task scheduling optimization system based on a multi-core controller, corresponding to the above method, the system comprising: A construction module is used to divide several core control tasks according to the operation requirements of the wind turbine, determine several corresponding task scheduling units based on the several core control tasks, and construct a data dependency graph between the several task scheduling units. The judgment module is used to monitor the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller. Based on the current load status and fault indication signals, it judges whether the performance of the processing core is abnormal, and triggers fault recovery scheduling according to the judgment result, or obtains dynamic sensing information. The generation module is used to input the data dependency graph and the dynamic perception information into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme. An update module is used to execute the initial task allocation scheme to obtain performance feedback, and adjust the scheduling weight parameters corresponding to the task scheduling unit based on the performance feedback to update the initial task allocation scheme.

[0058] Based on the same inventive concept disclosed above, this application also provides an electronic device. The electronic device of this application includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0059] It should be noted that the electrical connections between the above-mentioned units do not necessarily represent the connections between lines. Indirect connections are applicable to the embodiments of this application as long as they achieve the purpose of this application.

[0060] Based on the same inventive concept, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above method.

[0061] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for optimizing inter-core task scheduling based on a multi-core controller, characterized in that, The method includes, Based on the operational requirements of the wind turbine, several core control tasks are divided, and several corresponding task scheduling units are determined based on these core control tasks. A data dependency graph between these task scheduling units is then constructed. Monitor the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller, and determine whether the performance of the processing core is abnormal based on the current load status and fault indication signals, so as to trigger fault recovery scheduling according to the judgment result, or obtain dynamic sensing information. The data dependency graph and the dynamic perception information are input into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme. The initial task allocation scheme is executed to obtain performance feedback. Based on the performance feedback, the scheduling weight parameters corresponding to the task scheduling unit are adjusted to update the initial task allocation scheme.

2. The method according to claim 1, characterized in that, Constructing a data dependency graph among several task scheduling units, specifically including: The task scheduling unit is used as the initial task node to form a node set; Based on the data flow relationship between tasks, define the directed dependency edges between the initial task nodes to form a set of directed edges; Based on the set of nodes and the set of directed edges, a task-dependent directed graph and its corresponding adjacency matrix are generated. For each of the initial task nodes, an initial node feature vector is constructed that includes at least one of the following: task computation requirements, task cycle, deadline, and scheduling priority. An initial node feature matrix is ​​then formed based on the initial node feature vector. The data dependency graph is obtained based on the task dependency directed graph, the corresponding adjacency matrix, and the initial node feature matrix.

3. The method according to claim 1, characterized in that, Determine whether the core performance is abnormal based on the current load status, specifically including: Detect the current task execution rate and current health status value of any processing core, and determine the current load status based on the current task execution rate and current health status value; The current load state is compared with a preset load state threshold, and the presence of fault indication signals is monitored. Based on the comparison results and the fault indication signal monitoring results, it is determined whether the core processing performance is abnormal.

4. The method according to claim 3, characterized in that, Based on the judgment result, fault recovery scheduling is triggered, or dynamic perception information is obtained, specifically including: When the current load state corresponding to any processing core is less than or equal to the preset load state threshold, or when a fault indication signal exists, fault recovery scheduling is triggered to transfer the task scheduling unit on the abnormal core and ensure that the status data is consistent. When the current load status is greater than the preset load status threshold and there is no fault indication signal, dynamic sensing information is obtained based on the real-time operating status parameters and the current load status.

5. The method according to claim 4, characterized in that, The task scheduling unit on the exception core is transferred, and state data consistency is ensured. Specifically, this includes: Identify the abnormal core and the normal processing core, and determine the set of target task scheduling units on the abnormal core; Analyze the current load status of the normal processing core, and select the target migration processing core based on the current load status; Through a communication synchronization mechanism, the status data of the task scheduling unit set on the anomaly core is copied to the target migration processing core in real time.

6. The method according to claim 1, characterized in that, The calculation generates an initial task allocation scheme, specifically including: The data dependency graph and the dynamic perception information are input into the state perception graph neural network; The initial node feature matrix is ​​linearly transformed by a trainable weight matrix to obtain an initial feature representation. The initial feature representation and the task-dependent directed graph are aggregated by a graph convolutional layer, and the initial relationship weights between the initial task nodes are calculated using an attention mechanism. Based on the initial relation weights, the features of the neighbor nodes of any initial task node are weighted and summed to obtain the neighbor information; The neighbor information is fused with its corresponding initial feature representation, and then processed by a nonlinear activation function to output the updated node embedding feature vector. The node embedding feature vector of each initial task node is input into the fully connected layer and mapped to an output vector of the same dimension as the number of processing cores; Apply the softmax function to any of the output vectors to calculate the probability distribution of the corresponding task across all processing cores. Based on the probability distribution of all tasks on each processing core, a probability sampling method is used to determine the initial core to which any task scheduling unit belongs, forming an initial mapping relationship between tasks and cores, i.e., the initial task allocation scheme.

7. The method according to claim 1, characterized in that, Adjusting the scheduling weight parameters corresponding to the task scheduling unit based on performance feedback to update the initial task allocation scheme specifically includes: The initial task allocation scheme is embedded and pooled, and then concatenated with the real-time utilization rate, real-time wind conditions and real-time core power of each processing core to construct the environmental state vector at the current moment. The environmental state vector is input into the policy network to obtain the score of each candidate scheduling action. After filtering illegal actions by combining the mask formed by task dependency and core state, the action probability distribution is calculated and sampled to generate the mapping relationship between tasks and core and execute scheduling. After the scheduling is executed, the system's performance indicators are collected. The performance indicators are normalized and compared based on preset indicator thresholds, the weights of each indicator are adaptively adjusted, and the instant reward is calculated in a multi-objective weighted manner. The advantage function is calculated using the temporal difference and generalized advantage estimation algorithm, and the parameters of the policy network and value network are updated based on the policy gradient method, combined with the immediate reward and the advantage function. The advantages and contributions of each task in historical scheduling decisions are statistically analyzed. The scheduling weight parameters corresponding to each task are updated using the exponential moving average method, and these parameters are used as priority biases to introduce into subsequent strategy decisions. The optimization is considered to have converged when the cumulative reward improvement is lower than the threshold or the key performance indicators continue to meet the preset conditions.

8. A multi-core controller-based inter-core task scheduling optimization system, characterized in that, The system includes: A construction module is used to divide several core control tasks according to the operation requirements of the wind turbine, determine several corresponding task scheduling units based on the several core control tasks, and construct a data dependency graph between the several task scheduling units. The judgment module is used to monitor the real-time operating status parameters of the wind turbine and the current load status of each processing core of the multi-core controller. Based on the current load status and fault indication signals, it judges whether the performance of the processing core is abnormal, and triggers fault recovery scheduling according to the judgment result, or obtains dynamic sensing information. The generation module is used to input the data dependency graph and the dynamic perception information into the scheduling decision model of the state-aware graph neural network to calculate and generate an initial task allocation scheme. An update module is used to execute the initial task allocation scheme to obtain performance feedback, and adjust the scheduling weight parameters corresponding to the task scheduling unit based on the performance feedback to update the initial task allocation scheme.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the inter-core task scheduling optimization method based on a multi-core controller as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the inter-core task scheduling optimization method based on a multi-core controller as described in any one of claims 1-7.

Images