Task value and delay sensitivity multi-dimensional classification-based algorithm and power coordination scheduling method and system

Abstract

The application relates to the technical field of algorithm and network cooperative scheduling, and discloses an algorithm and power cooperative scheduling method and system based on multi-dimensional classification of task value and time delay sensitivity, which comprises the following steps: acquiring multi-source data streams of computing power tasks, power markets and network topologies; performing multi-dimensional classification processing based on a time delay sensitivity index and a comprehensive value score to generate a task type code; extracting power market environment features to dynamically generate a multi-target optimization weight vector, and screening network topology nodes to generate a candidate data center set; combining the code and the weight vector to calculate the comprehensive cooperative utility of each node, generate a to-be-verified scheduling decision, input the decision into a consortium chain network to perform a quoted price tolerance check and consensus determination, and generate a scheduling instruction after the check and determination; collecting actual operation parameters according to the instruction to construct an experience four-tuple, and inputting the experience four-tuple into a meta-learning network for parameter fine-tuning. Through multi-dimensional feature mapping and distributed checking, the application realizes closed-loop cooperative scheduling of heterogeneous computing power and dynamic power.

Description

Technical Field

[0001] This invention relates to the field of computer-network collaborative scheduling technology, specifically to a computer-network collaborative scheduling method and system based on multi-dimensional classification of task value and latency sensitivity. Background Technology

[0002] With the development of cross-regional data centers, collaborative scheduling of computing and power has become a technical means to allocate computing resources and control power costs. However, in existing scheduling methods, multi-source heterogeneous computing tasks often lack a unified quantitative standard before being scheduled. Existing technologies fail to fully extract multi-dimensional characteristics such as data scale and business type of computing tasks, and also fail to combine and structure features in conjunction with latency requirements and security levels. This results in the system being unable to map different types of tasks to a unified feature representation, making it difficult for subsequent data center nodes to directly perform resource matching based on quantified values.

[0003] Meanwhile, when dealing with dynamic and fluctuating electricity market environments, existing scheduling models used to generate multi-objective optimization weights typically rely on random initialization of model parameters each time the market environment changes. This mechanism fails to effectively calculate and utilize pre-trained parameters from similar historical environmental scenarios as the starting point for model updates, increasing the number of forward inference iterations and resulting in longer convergence times for the model's parameters in response to dynamic electricity market fluctuations.

[0004] Furthermore, existing scheduling schemes lack distributed verification of price data and closed-loop correction mechanisms for operational status after decision generation. The system typically fails to perform tolerance-based comparisons between quoted electricity prices and actual external electricity prices, making it prone to introducing scheduling deviations due to abnormal price data. Moreover, after dispatch instructions are issued, the system lacks feedback on actual indicator parameters, making it unable to construct samples based on actual execution status to supervise and update the underlying network model, resulting in difficulties in effectively controlling execution errors during the actual scheduling process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a computing-electricity collaborative scheduling method and system based on multi-dimensional classification of task value and latency sensitivity. This solves the problem that the lack of mapping between heterogeneous computing power task characteristics and the fluctuating state of the electricity market leads to limited parameter convergence speed, uneven allocation of network transmission resources, and lack of distributed reliable verification of scheduling instructions when cross-regional scheduling decisions meet latency and cost constraints.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a computing-powered collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity, comprising the following steps: The system acquires multi-source input data, including computing task data streams, electricity market data streams, and network topology data streams. It performs quantitative analysis on the computing task data streams, calculating latency sensitivity indices and comprehensive value scores. These indices and scores are then used to perform multi-dimensional classification mapping of computing tasks, generating task type codes. Simultaneously, market environment features are extracted from the electricity market data streams, and multi-objective optimization weight vectors are generated through model inference to respond to real-time market environment changes. The network topology data streams are analyzed, and data center nodes are selected based on preset constraints to generate a candidate set. During the decision-making phase, the task type codes and multi-objective optimization weight vectors are used as input variables to calculate the comprehensive collaborative utility of each node within the candidate set, outputting a scheduling decision to be verified. An effective scheduling instruction is generated through consensus verification. Finally, the execution process is monitored, and actual indicator parameters are collected, performing closed-loop parameter fine-tuning.

[0008] Preferably, the multi-source input data includes security and value assessment data streams. The system performs dimensionless mapping on the continuous values ​​in the multi-source input data, transforming heterogeneous physical quantities into a unified numerical range, and splicing them to generate a preprocessed input dataset.

[0009] Preferably, the task classification process determines the data security level of the computing power task, discretizes the latency sensitivity index, comprehensive value score and data security level, performs multi-dimensional feature dimensionality reduction according to feature cross-combination logic, and outputs a structured task dataset and task type code.

[0010] Preferably, when evaluating the latency sensitivity index, the task type feature function, business type feature function, and data scale feature function are multiplied by their corresponding static weight coefficients, and the product results are summed; when evaluating the comprehensive value score, the business value score, user impact score, and data value score are multiplied by their corresponding global weight coefficients and summed; when evaluating the data security level, the data sensitivity identifier value, data scale, and task type are constructed as joint judgment input variables, and the output result is calculated using the security level evaluation function.

[0011] Preferably, the calculation process of the data scale characteristic function performs logarithmic scaling. The system obtains the minimum and maximum values ​​of the data scale, as well as a preset constant, calculates the ratio of the logarithmic difference between the current data scale and the minimum value to the logarithmic difference between the maximum and the minimum value, and subtracts this ratio using the preset constant. This process reduces the influence of the maximum value on the shift of the characteristic function through logarithmic domain transformation.

[0012] Preferably, the generation process of the multi-objective optimization weight vector involves parameter matching by retrieving historical environmental feature databases. The system constructs a feature vector from the current market environment features, calculates the ratio of its inner product and the product of its magnitude with the feature vectors of historical scenarios, and obtains a similarity score. Based on the maximum similarity, the starting point of the pre-trained meta-model parameters is extracted. Combining the current electricity market sample data, local learning rate step size, and loss function, the parameter gradient update amount is obtained through partial derivative gradient calculation, and the model parameters are fine-tuned. This process utilizes historical experience distribution to reduce the number of training iterations of the model in the current environment, improving the parameter convergence speed.

[0013] Preferably, the network transmission cost calculation process extracts the normalized physical distance parameter, normalized real-time bandwidth utilization rate, normalized network latency fluctuation parameter, and bandwidth instability parameter of the candidate data center node, and performs multiplication and accumulation operations with four preset sets of independently assigned weight coefficients respectively, transforming spatial topology attributes and real-time communication status into quantified transmission cost indicators.

[0014] Preferably, the calculation of comprehensive synergistic utility includes multi-dimensional reward quantification. The system calculates the electricity price economic reward, green electricity consumption reward, latency satisfaction reward, task value reward, and data security reward for candidate data center nodes in relation to computing power tasks. Each reward is weighted and summed with the corresponding coefficients in the multi-objective optimization weight vector, and a Pareto front is constructed using a multi-objective optimization algorithm. Nodes that meet the non-dominated solution conditions are selected to form a non-dominated solution set, from which the node corresponding to the maximum comprehensive synergistic utility is selected as the scheduling target.

[0015] Preferably, the automated verification and closed-loop fine-tuning process is executed in a consortium blockchain network environment. The transaction center nodes in the consortium blockchain network calculate the absolute difference between the quoted electricity price and the real-time external electricity price in the scheduling decision to be verified. If the absolute difference is less than the minimum price tolerance value, the verification is deemed successful, and the consensus nodes are triggered to execute a consensus decision to generate an effective scheduling instruction. In the fine-tuning phase, multi-dimensional comprehensive deviation feedback signals and actual comprehensive business utility feedback signals are extracted. These are combined with environmental feature vectors and weight vectors to construct empirical quadruples. The multi-dimensional comprehensive deviation feedback signals are used as supervisory signals input into the meta-learning network to update the model parameters.

[0016] A second aspect of this invention provides a computing-electricity collaborative scheduling system based on multi-dimensional classification of task value and latency sensitivity, comprising: a data processing module for accessing and preprocessing multi-source input data; a task classification module for performing quantitative evaluation and encoding of task feature dimensions; a meta-learning module for calculating multi-objective optimization weight vectors; a matching module for filtering candidate data center sets that meet constraints; an optimization module for outputting scheduling decisions through multi-dimensional reward weighting and non-dominated solution filtering; an auditing module for issuing scheduling instructions through a consortium blockchain consensus mechanism; and a feedback module for constructing empirical quadruples and performing model parameter iteration. The modules exchange data and transmit control signaling through an internal communication bus.

[0017] This invention provides a computing-powered collaborative scheduling method and system based on multi-dimensional classification of task value and latency sensitivity. It offers the following advantages: 1. This invention extracts multi-dimensional features such as data scale and business type of computing power tasks, calculates latency sensitivity index and comprehensive value score, and combines data security level to perform discretization and feature cross-combination to generate task type code, mapping multi-source heterogeneous computing power task data into a unified structured feature representation, solving the problem of lack of unified quantitative standards for different types of tasks before scheduling, so that subsequent data center nodes can directly perform resource matching based on quantitative values.

[0018] 2. This invention calculates the similarity score between the feature vector of the current market environment and the feature vector of the historical environment scene, extracts the pre-trained meta-model parameters corresponding to the historical scene with the highest similarity as the starting point, and uses the partial derivative gradient calculated by the current environment sample data to perform parameter fine-tuning. This avoids the random initialization of model parameters every time the market environment changes, reduces the number of forward inference iterations in the process of generating weight vectors for multi-objective optimization, and shortens the convergence time of the model for dynamic power market fluctuations.

[0019] 3. This invention inputs the scheduling decision to be verified into the consortium blockchain network. The transaction center node calculates the absolute difference between the quoted electricity price data and the real-time external electricity price data and performs tolerance comparison. After the consensus judgment is passed, the instruction is issued. At the same time, actual indicator parameters are collected to construct an empirical quadruple to perform supervised updates on the meta-learning network. This mechanism uses distributed node comparison to eliminate scheduling deviations caused by abnormal electricity price data, and continuously corrects model parameters through closed-loop feedback of real execution data, thereby controlling the execution error in the actual scheduling process. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the computing-electricity collaborative scheduling system architecture based on multi-dimensional classification of task value and latency sensitivity according to the present invention. Figure 2This is a schematic diagram of the computing and power-coordinated scheduling method based on multi-dimensional classification of task value and latency sensitivity according to the present invention. Figure 3 This is a schematic diagram of the method for multi-objective Pareto front search and optimal node determination in computing power tasks according to the present invention. Figure 4 This is a schematic diagram of the dual feedback closed-loop control of the meta-learning mechanism of the present invention; Figure 5 This is a schematic diagram illustrating the multi-objective Pareto front search and optimal node determination for computing power tasks in this invention. Figure 6 This is a schematic diagram of the convergence curve of the dual-feedback closed-loop meta-learning mechanism of the present invention, wherein... Figure 6 (a) is a curve showing the convergence of the perception layer supervised optimization prediction and the actual physical error in this invention; Figure 6 (b) is a graph showing the upward trend of the actual physical execution reward for the decision-making layer strengthening closed loop of the present invention. Detailed Implementation

[0021] The technical solutions in 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.

[0022] See attached document Figure 1 In this embodiment, the present invention provides a computing-electricity collaborative Pareto scheduling system based on multi-dimensional classification of task value and latency sensitivity, which may include: a data processing module, a task classification module, a meta-learning module, a matching module, an optimization module, an auditing module, and a feedback module.

[0023] The data processing module synchronously collects input data from multi-source data interfaces and performs data cleaning and normalization to generate a preprocessed input dataset. The multi-source data interfaces connect to computing power applications, power trading centers, network monitoring systems, and enterprise security management systems. The data processing module has a built-in pre-defined alignment and verification mechanism that performs synchronous truncation and alignment operations on heterogeneous information such as computing networks and security posture through time benchmark and spatial feature matching. The multi-source data interfaces specifically collect latency and bandwidth parameters from the network topology.

[0024] The task classification module receives the preprocessed input dataset, performs latency sensitivity analysis, calculates task value scores, and determines data security levels, generating task type codes and constructing a structured task dataset. The task classification module sets preset numerical threshold ranges, discretizes the continuously calculated multidimensional scores, and classifies computing tasks into preset refined task type codes based on cross-combination logic of multidimensional features.

[0025] The meta-learning module receives electricity market environment data, extracts environmental features and matches them with historical features, and generates a dynamic multi-objective optimized weight vector adapted to the current electricity market environment through meta-parameter updates. The meta-learning module constructs a meta-training task set by extracting multiple historical electricity market fluctuation scenarios and updates the meta-model parameters using a meta-gradient descent algorithm. When faced with new operating conditions and entering the online execution phase, it uses pre-trained meta-parameters combined with a small amount of sample data from the current environment to perform rapid model parameter fine-tuning, dynamically generating a dynamic multi-objective optimized weight vector adapted to the current operating conditions.

[0026] The matching module performs hard rule filtering based on the structured task dataset and network topology data to generate a candidate data center set. It then calculates network transmission cost and matching weights to generate a preliminary scheduling candidate subset and a preliminary scheduling scheme. The matching module pre-computes physical isolation qualifications and extreme latency constraints to perform hard elimination of data center nodes, generating the candidate data center set. It also integrates multiple network feature parameters to calculate the comprehensive network transmission cost. Finally, the matching module calculates the comprehensive matching weight of each candidate node in the candidate data center set using multi-dimensional matching evaluation indicators, and selects the preliminary scheduling candidate subset as the preliminary scheduling scheme based on descending order of matching weights.

[0027] The optimization module receives a dynamic multi-objective optimization weight vector and a preliminary subset of scheduling candidates, calculates the multi-dimensional utility reward for each candidate node, and generates a scheduling decision to be verified. Combining the previously input dynamic multi-objective optimization weight vector, the optimization module performs a linear weighted summation of various utility evaluation indicators to construct a global multi-objective utility space. Within this multi-dimensional utility space, it searches for the Pareto front and identifies the data center node with the highest weighted total utility score in the resulting set of non-dominated solutions. This generates a scheduling decision to be verified, containing a complete data structure including the computing task ID, target data center ID, expected execution period, expected green electricity usage ratio, and digital signature.

[0028] The audit module broadcasts the scheduling decisions to be verified to the consortium blockchain network for multi-node verification and consensus mechanism execution, generating effective scheduling instructions that are then uploaded to the blockchain. Relying on a consortium blockchain architecture comprising nodes from data centers, power trading centers, and computing power service providers, the audit module issues smart contracts to drive oracles to compare off-chain data consistency. Specifically, trading center nodes verify electricity prices and green electricity data, computing power service provider nodes verify resource idle status, and security nodes verify secure signature matching. During the execution of the Practical Byzantine Fault Tolerance (PBFT) consensus algorithm, the system extracts verification results based on feedback from all nodes in the network and calculates the credibility of the scheduling decisions. When the credibility of the decision exceeds a preset threshold and the number of valid verification votes returned by the entire network meets the preset statutory threshold limit conditions of the consensus algorithm, consensus is determined to be reached, and an immutable block hash value is automatically generated.

[0029] The feedback module monitors the scheduling execution status, collects actual execution data, calculates prediction deviations and actual business utility, and stores them in the experience replay pool, triggering model parameter updates in the meta-learning module. The feedback module compares the deviations of the actual completion delay and the actual settled electricity cost from the initial prediction value, converting them into a multi-dimensional comprehensive deviation feedback signal. Simultaneously, it calculates an actual comprehensive business utility feedback signal based on actual physical indicators. These dual feedback signals, along with the associated environmental state features and execution actions, are encapsulated into Markov decision process experience quadruples and fed into the experience replay pool. Batch samples are extracted at set intervals to drive the meta-learning module to perform supervised learning and reinforcement learning fine-tuning compensation on the weights of the underlying neural network.

[0030] See attached document Figure 2 The present invention also provides a computing-powered collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity. This method is implemented through the system disclosed in the embodiments, and the specific workflow includes the following steps: S1. The system receives real-time task requests and synchronously collects key input data from multi-source data interfaces. The data processing module acquires computing power task data streams, electricity market data streams, network topology data streams, and security and value assessment data streams. It then performs dimensionless mapping on the raw, continuously collected values ​​using preset benchmark rules to generate a preprocessed input dataset.

[0031] S2. The task classification module calculates the latency sensitivity index and comprehensive value score of the task based on the preprocessed input dataset to determine the data security level. The continuous scores are discretized according to the preset threshold, and the task is mapped to the corresponding structured task dataset and task type code based on the multi-dimensional feature cross-combination logic.

[0032] S3, the meta-learning module receives the current electricity market data stream, extracts environmental features to match the historical environmental feature library, and calls the pre-trained meta-model parameters in the meta-parameter library to update and match them with the current market environment samples, and outputs an adaptive scheduling strategy. The adaptive scheduling strategy is specifically manifested as a dynamic multi-objective optimized weight vector for the current market environment dimension.

[0033] S4. The matching module combines the spatiotemporal network topology characteristics to eliminate data center nodes that do not meet the physical isolation and latency limit constraints, generating a candidate data center set. For the nodes in the set, the network transmission cost is calculated, and the node matching weight is calculated in combination with multi-dimensional matching evaluation indicators. Based on the matching weight, the nodes are sorted in descending order and a preliminary scheduling candidate subset is generated.

[0034] S5. The optimization module calculates multi-dimensional independent utility rewards for each node in the initial scheduling candidate subset, and calculates the weighted total utility score by combining the dynamic multi-objective optimization weight vector. The node with the largest weighted total utility score that is at the Pareto front is selected to generate a scheduling decision to be verified.

[0035] S6. The audit module sends the scheduling decision to be verified to each verification node in the consortium blockchain network for automated verification. The system extracts the verification results and calculates the credibility of the scheduling decision. When the credibility of the decision exceeds a preset threshold and the number of valid verification votes in the entire network reaches the preset legal threshold limit of the consensus algorithm, consensus is determined to be reached. The audit module packages the decision data into blocks and stores them on the blockchain, forming an immutable and effective scheduling instruction.

[0036] S7. The feedback module monitors the task execution process of the target data center according to the effective scheduling instructions, collects actual indicator parameters, calculates multi-dimensional comprehensive deviation feedback signals and actual comprehensive business utility feedback signals, and stores the experience quadruple containing environmental characteristics, actions and dual feedback signals into the experience playback pool, driving the meta-learning module to extract data for model parameter fine-tuning.

[0037] See attached document Figure 1 and attached Figure 2 In step S1, the data processing module synchronously collects key input data from the multi-source data interface, performs the acquisition and standardization processing of multi-source heterogeneous data, and establishes a real-time communication link with the external physical environment through the preset multi-source data interface.

[0038] In this embodiment, the computing power task data stream acquired by the data processing module originates from computing power application terminals such as financial transaction systems, 5G edge computing nodes, blockchain nodes, and video rendering platforms. The computing power task data stream contains two types of data tags: basic task information and additional attributes. The basic task information carries the task identity identifier, task type, data scale, and business type, used to characterize the expected consumption of underlying physical resources. The additional attributes carry the user identity identifier who initiated the task, data sensitivity identifier value, and business urgency index, used to construct a baseline for value and security evaluation in subsequent dimensions.

[0039] The data processing module acquires electricity market data streams pulled from external electricity trading center nodes via an application programming interface (API). The collected data covers real-time electricity prices, green electricity output forecasts, grid load status parameters, and regional electricity supply-demand balance indicators for the locations of regional data centers. Simultaneously, the data processing module connects to an external weather forecasting system to obtain wind speed and solar radiation intensity forecasts to capture their impact on renewable energy generation output. The communication protocol configuration for the API (such as standard Hypertext Transfer Protocol or Message Queuing Telemetry Protocol) is well-known in the field and will not be elaborated upon here.

[0040] The network topology data stream acquired by the data processing module originates from the underlying network monitoring system. The network monitoring system utilizes network probes deployed in routers at various regional nodes to transmit data back to the data processing module, including the physical distance between nodes, real-time bandwidth utilization, network latency fluctuations, and network fault warnings. The deployment of network probes and the underlying capture processes, such as network traffic mirroring technology, are well-known technologies in this field and will not be elaborated upon here.

[0041] As a preferred approach, the security and value assessment data streams acquired by the data processing module originate from the enterprise security management system and business support system. The former periodically outputs data security level assessment results for the application end; the latter delivers fine-grained task value assessment parameters, including business importance scores, user impact scope base, and data value coefficients.

[0042] After acquiring all the data, the data processing module performs a data cleaning process. Using built-in fixed data type validation rules and extreme value thresholds, it automatically removes missing and outlier values ​​that do not conform to the specified format. In this embodiment, the extreme value threshold is set based on the device's physical limits and statistical extreme values ​​from the past three months. For example, if the physical transmission delay is set to an extreme value of two thousand milliseconds, data exceeding this threshold will be marked as invalid. The implementation methods for data cleaning, such as regular expression matching, are well-known technologies in the field and will not be elaborated upon here.

[0043] For the retained valid data samples, the data processing module performs dimensionless processing on all continuous numerical input variables. The module employs a min-max normalization algorithm to proportionally map the original continuous numerical input variables to an absolute standard range of zero to one. The normalization calculation formula is defined as follows: ;in, This represents the actual measured value of the original continuous numerical input variable currently being processed by the data processing module. and These represent the lower and upper limits of the corresponding category input variables in the system's historical statistical records, respectively. To prevent the denominator from being zero and to ensure computational completeness, a first minimum positive offset is introduced. (The range of values ​​is) to Normalized absolute value Its value range is strictly limited to a closed interval between zero and one.

[0044] After processing, the data processing module extracts the global timestamp and task identifier of each task request as a baseline, and performs time-dimensional correlation and feature fusion on the cleaned and normalized multi-source heterogeneous data streams. For high-frequency data streams such as computing power tasks and network status, synchronous truncation and alignment operations are performed to remove isolated data packets that exceed the dynamic adaptive tolerance window. The initial baseline time span of this high-frequency tolerance window is set to one hundred milliseconds. The data processing module dynamically updates the real-time time span limit by combining the statistical distribution of high-frequency data packet arrival delays (such as moving average delay and standard deviation) and the current network status within historical time periods to cope with network jitter and avoid the erroneous removal of valid data.

[0045] For low-frequency data streams such as power status and weather forecasts, the data processing module employs a forward-fill algorithm based on a zero-order hold. This means that before the next low-frequency data cycle arrives, the state of the most recently received valid low-frequency data is preserved and mapped to the timestamp nodes of all current high-frequency tasks. Through time base alignment and spatial feature matching, the data processing module integrates computing power attributes, power status, network status, and security assessment information to generate a highly structured preprocessed input dataset. This dataset is then written into a shared memory area as a multi-dimensional matrix data structure for the task classification module to use for judgment.

[0046] See attached document Figure 1 and attached Figure 2 In step S2, the task classification module calls the preprocessed input dataset from the shared memory region to perform a three-dimensional structured decomposition and quantitative analysis of the latency, value, and security of computing power tasks.

[0047] For latency sensitivity quantitative analysis, the task classification module extracts the task type, business type, and data size from the preprocessed input dataset. The task classification module then executes a latency sensitivity index. The calculation formula is defined as follows: ; in, This represents the latency sensitivity index. The system uses a pre-defined type mapping table to map interactive query and other technical types to specific values ​​between zero and one, representing the characteristic function of the task type. The system represents the business type characteristic function, which maps specific business attributes such as financial transactions to standard values ​​between zero and one. This represents a characteristic function of data size. , as well as These represent the corresponding static weight coefficients, and the system stipulates that the sum of the three must equal one. The construction logic of the above mapping table and the initial assignment and setting of each static weight coefficient can be achieved by those skilled in the art using an expert scoring matrix (such as the Analytic Hierarchy Process, AHP) combined with historical scheduling log regression analysis. The specific mapping value calibration and weight calculation process are well-known technologies in this field and will not be elaborated here.

[0048] As a preferred approach, since the marginal impact of data throughput on latency exhibits a non-linear decreasing trend, the data size characteristic function is calculated using logarithmic scaling logic, and its formula is defined as: ; in, This indicates the actual data scale of the current computing power task. and These represent the minimum and maximum values ​​of the input data collected during the system's historical operating cycle, respectively. Specifically, these values ​​are obtained by dynamically statistically analyzing and extracting extreme values ​​from the computing task logs within a fixed-length historical time window. The value 1 within the logarithmic term is the first preset constant used for logarithmic smoothing to prevent the input from being zero. The value 1 at the beginning is the second preset constant used for inverse normalization of the eigenvalues. In practical applications, the first and second preset constants are not limited to the value 1; they can be set to other positive real numbers based on the specific data distribution characteristics through pre-defined grid searches or empirical data fitting. To ensure the absolute completeness of the division operation logic and prevent computational overflow caused by the denominator approaching zero when extreme values ​​coincide, a second minimum positive number offset is introduced. Its value is The specific underlying calculation process of the logarithmic function and the extreme value statistics algorithm under the sliding window can be directly implemented by those skilled in the art using existing basic mathematical libraries and data processing frameworks. These are well-known technologies in the field and will not be elaborated upon here.

[0049] For calculating the comprehensive value score of a multi-dimensional task, the task classification module performs feature fusion across independent dimensions. The formula for calculating the comprehensive task value score is defined as follows: ; in, This represents the overall task value score obtained through calculation. The business value score is obtained by multiplying multiple business attribute feature scalars (such as business type feature scalars and industry standard evaluation scalars) preset by the system by their corresponding sub-weight coefficients and then performing a linear weighted summation. The user impact score is calculated by multiplying multiple user characteristic scalars (such as the number of end users and user importance indicators) by their corresponding sub-weight coefficients. The data value score is generated by multiplying multiple data feature scalars (as a preferred approach, including data sensitivity, overall data size, and data timeliness) by their respective sub-weight coefficients. , as well as These represent the global weight coefficients assigned to the three evaluation sub-dimensions, each with a value within the open interval of zero to one, and their sum is strictly equal to one. The global weight coefficients and sub-weight coefficients can be automatically calculated and generated by those skilled in the art using an objective weighting method based on the variance distribution of historical multidimensional feature data (such as the entropy weighting method). The principle and implementation process of this objective weighting algorithm are well-known technologies in this field and will not be elaborated upon here.

[0050] For data security level assessment and determination, the task classification module extracts data sensitivity identifiers, data size, and task type as joint input variables based on the enterprise's physical isolation requirements for security compliance. The task classification module calls a pre-defined security level assessment function to perform the determination. The security level assessment function is defined as follows: ;in, This represents the data security level score in the form of continuous numerical output. This indicates the data sensitivity indicator value. Indicates the data size. This represents a scalar indicating the task type. This represents a non-linear security mapping function, whose physical purpose is to map the true hard requirements of a task for the physical and logical isolation qualifications of the underlying data center through the cross-combination of three factors. For non-linear security mapping functions... For the specific construction and fitting, those skilled in the art can use multilayer perceptron neural networks or support vector machines based on radial basis function kernels combined with historical compliance cases for supervised training. The output layer needs to be configured with normalized activation functions such as Sigmoid to ensure that the safety level score of the continuous numerical form of the output strictly falls within the standard numerical space of zero to one, thereby ensuring the mathematical validity and feasibility of the subsequent double absolute threshold truncation rule. The model training and function fitting process is a well-known technology in this field and will not be described in detail here.

[0051] After completing the quantization calculation of the three-dimensional continuous numerical values, the task classification module performs discretization dimensionality reduction and refined task type encoding generation operations. In this embodiment, the task classification module constructs a latency sensitivity classification system for the latency sensitivity dimension. To assign an absolute physical boundary to the previously calculated dimensionless exponent, the task classification module incorporates an inverse latency mapping function to convert the calculated continuous latency sensitivity exponent into a linear time-delay sensitivity index. Mapping to expected response time The specific construction of the aforementioned inverse time delay mapping function can be achieved by those skilled in the art through inverse proportional function fitting, linear interpolation, or a preset numerical correspondence table. The mathematical transformation process is a well-known technique in this field and will not be elaborated here.

[0052] Subsequently, the system extracts the expected response time generated by the mapping. Implement hard physical threshold truncation. Specifically, categorize the latency sensitivity of computing tasks into three types: when latency requirements are met... When the real-time response boundary is met, it is determined and mapped to a latency-sensitive type; when the latency requirement is satisfied... When the latency requirement is met, it is determined and mapped to a latency-tolerant type; when the latency requirement is met... When the interaction tolerance boundary is reached, it is determined and mapped to a time-delay flexible type.

[0053] The real-time response boundary (100ms) and the interaction tolerance boundary (500ms) are benchmarks established based on the standards of the computing network convergence engineering project and the physical characteristics of actual business. The physical basis for establishing the real-time response boundary comes from the general physical tolerance limit of highly real-time businesses such as high-frequency financial transactions and vehicle-to-everything (V2X) edge computing. These businesses require control signaling to complete the closed loop in a very short time to avoid logical blockage. The establishment of the interaction tolerance boundary corresponds to the upper limit of user experience tolerance for regular Internet interactions (such as streaming media buffering and complex database queries). For tasks exceeding this boundary (such as offline data mining and batch video rendering), the sensitivity of their business performance to time decreases non-linearly. The system can safely classify them into a flexible task queue that can be scheduled during off-peak hours of the power grid at night based on this physical characteristic.

[0054] For either the overall task value score or the data security level score, the task classification module uniformly sets a first preset classification threshold. and the second preset classification threshold In this embodiment, The value is set to , The value is set to .for and The logic for setting this up can be understood by those skilled in the art by performing normal distribution statistics on the corresponding score data of massive tasks within the system's historical operating cycle, and extracting specific mathematical quantiles (such as the quantile boundary between the top 30% and the bottom 30%) as preset benchmark values. The process of offline threshold calibration and solidification based on statistical quantiles is a well-known technology in the field and will not be elaborated here. For any calculated value or security score, when it is greater than or equal to 0.7, the task classification module marks its physical state as high-level; when it is greater than or equal to 0.3 and less than 0.7, it marks it as medium-level; and when the score is less than 0.3, it marks it as low-level.

[0055] The system maps the high, medium, and low of the task comprehensive value score dimension to high value, medium value, and low value, respectively; and maps the high, medium, and low of the data security level dimension to high security, medium security, and low security, respectively. High security corresponds to physical isolation requirements, medium security corresponds to logical isolation requirements, and low security corresponds to no special security requirements.

[0056] Based on the aforementioned threshold truncation rules, the task classification module maps the continuous multidimensional scores of the three dimensions to three discrete levels. Using orthogonal combinatorial logic in the mathematical space of multidimensional features, the three levels of each of the three dimensions intersect, ultimately forming twenty-seven defined, refined physical combination states within the system. The task classification module precisely maps tasks to corresponding structured task datasets and task type codes; for example, financial transaction tasks are coded as latency-sensitive, high-value, and high-security types. The task classification results and the original task data are linked by foreign keys in the underlying database and stored in a structured manner, forming a structured task dataset, which is then directly pushed into the next level of scheduling control flow as standardized task input. For the foreign key associations and structured storage in the underlying database, those skilled in the art use standard relational databases (such as MySQL or PostgreSQL) and SQL statements to establish master-slave table constraints. The database construction and data persistence processes are well-known technologies in this field and will not be elaborated upon here.

[0057] See attached document Figure 1 and attached Figure 2 In step S3, the meta-learning module receives the electricity market data stream and the structured task dataset, and executes a rapid adaptation mechanism based on meta-learning.

[0058] The system extracts electricity price fluctuation trends, changes in green power output, regional supply and demand balance, and grid load changes from electricity market data. These extracted environmental feature components are then sequentially combined to construct a multi-dimensional comprehensive environmental feature vector.

[0059] In this embodiment, the meta-learning module has a pre-built historical environment feature library, which stores feature vectors of typical power market fluctuation scenarios within historical operating cycles. The system iterates through and matches the current comprehensive environment feature vector with the historical environment feature library, and calculates the similarity score using cosine similarity logic. ;in, This represents the similarity score between the current comprehensive environmental feature vector and the historical environmental scene feature vector. This represents the feature vector of the current comprehensive environment. This represents the feature vector of a historical environment scene in the historical environment feature library. This represents the magnitude of the current integrated environmental feature vector. This represents the magnitude of the feature vector of the historical environment scene. This represents the offset of the third smallest positive number. To ensure the absolute completeness of the algorithm's calculation logic, its value is [value missing]. . It is determined based on the lower limit of floating-point arithmetic precision of the underlying deployed hardware, to ensure that while preventing program crashes caused by a denominator of zero, its minimum value attribute will not cause observable error interference to the mathematical precision of the original cosine similarity score.

[0060] This embodiment employs a two-layer optimization architecture based on model-independent meta-learning. As a preferred implementation, the base model for meta-learning can be a deep multilayer perceptron network architecture. The number of nodes in the input layer of this deep multilayer perceptron network strictly corresponds to the total dimension of the integrated environmental feature vector. The hidden layer contains three fully connected layers and uses the ReLU activation function to capture the nonlinear cross-coupling relationships between multi-source environmental features. The output layer contains five nodes and uses the Softmax activation function to ensure that the sum of all components of the output dynamic multi-objective optimization weight vector is strictly one. The basic architecture of the deep multilayer perceptron network and the standard mathematical operations of the ReLU and Softmax activation functions can be easily implemented by those skilled in the art using existing deep learning frameworks; these are well-known techniques in the field and will not be elaborated upon here.

[0061] During the offline phase of meta-training, the meta-learning module samples from multiple different historical electricity market environments to construct a training task set. Internally, the system pre-builds a meta-parameter library and maintains a set of meta-model parameters. , representing the initial generalization starting point for the deep multilayer perceptron network base model. The global objective function of the meta-learning module is defined as: ;in, This represents the target operator for minimization. This indicates the meta-model parameters that need to be optimized. This represents a mathematical summation operator. This represents the total number of historical sampling tasks. Indicates the first To ensure strict mathematical consistency with the subsequent reinforcement learning-based continuous fine-tuning mechanism, the loss function of each historical task is used. The system adopts a policy optimization error function (such as negative policy gradient error or value function error) based on the idea of ​​maximizing expected utility to quantify the physical deviation between the model output action and the cumulative utility of environmental feedback. Indicates the first The parameters of a specific model after local adaptation in a historical task, which has undergone single-step or few-step local adaptation.

[0062] In the meta-training of the inner loop, for each... For each historical task, the system samples from its training set and performs local parameter updates to obtain the parameters after specific model adaptation. The calculation formula is defined as follows: ;in, Indicates the first The parameters of a specific model after local adaptation on a historical task. Indicates the parameters of the meta-model Initial model parameters for instantiation. This represents the local learning rate step size of the inner loop, and its value range is set to... to . Indicates the initial model parameters Mathematical gradient operator for finding partial derivatives. Indicates the first The loss function for each historical task. To balance local convergence speed and model accuracy, the step size of this local learning rate is set during system initialization. The specific values ​​are determined by offline cross-validation combined with dynamic optimization of the network convergence curve. For the calculation of the objective function based on the gradient of the reinforcement learning policy, and the specific mathematical derivation of the gradient based on partial derivatives and backpropagation to update network parameters, those skilled in the art can refer to conventional deep reinforcement learning theories (such as REINFORCE or PPO algorithms), which are all well-known techniques in this field and will not be elaborated upon here.

[0063] In the meta-update of the outer loop, the validation loss performance of all historical tasks after the inner update is summarized, the meta-gradient is calculated, and the meta-model parameters are updated macroscopically. The calculation formula is as follows: ; in, This represents the parameters of the meta-model. This indicates the assignment update operator. This indicates the meta-learning rate step size of the outer loop. Indicates the parameters of the meta-model Mathematical gradient operator for finding partial derivatives. Indicates the first The loss function for each historical task. Indicates the first The parameters of a specific model after local adaptation on a historical task. In engineering implementation, the meta-learning rate step size. The range of values ​​is to Its order of magnitude is smaller than the inner layer's local learning rate step size to ensure the stability of the macroscopic update of global parameters. Its precise setting depends on the hyperparameter grid search mechanism locking it during the pre-training phase. Through this two-layer gradient update, the meta-model parameters... It is trained to a spatial state with high environmental universality and synchronously and persistently stored in the system's meta-parameter library.

[0064] During the online execution phase, when new environmental data of the current electricity market is received, the system extracts the most similar historical samples (these historical samples are the system's underlying experience records, fully containing historical environmental feature vectors, historical output weight actions, and comprehensive deviation penalty reward signals evaluated by the system afterward) as a small sample of the current electricity market environment. The system uses the meta-parameters in the meta-parameter library as a starting point and performs rapid model parameter fine-tuning based on the current electricity market environment sample. The parameter update formula is: ; in, This represents the new model parameters after minor adjustments to address the current unforeseen circumstances in the electricity market. This indicates the starting point of the pre-trained meta-model parameters loaded from the meta-parameter library. This indicates the step size of the local learning rate. Indicates the starting point of the parameters of the pre-trained meta-model. Mathematical gradient operator for finding partial derivatives. This represents the system loss function calculated based on a small amount of sample data from the current electricity market environment.

[0065] As a preferred approach, a rapidly fine-tuned deep multilayer perceptron network base model is used for forward propagation inference of the current electricity market state. The five nodes of the output layer are combined with the Softmax activation function to generate a continuous dynamic multi-objective optimized weight vector through end-to-end mapping. The dynamic multi-objective optimization weight vector Specifically, it includes weighted components in five dimensions: This indicates the economic weight of electricity prices, reflecting the system's sensitivity to reducing electricity costs; It represents the green electricity consumption weight, which is used to automatically increase the priority of new energy consumption when the grid has a surplus of green electricity; Indicates the weight of delay satisfaction; Indicates the task value weight; This represents the data security weights. The meta-learning module will dynamically optimize this multi-objective weight vector. It is pushed into the subsequent scheduling and control bus as the core coefficient basis for the final Pareto optimization.

[0066] See attached document Figure 1 and attached Figure 2 In step S4, the matching module receives the structured task dataset and combines it with the spatiotemporal network topology features to execute a program for candidate screening of spatiotemporal network topologies and calculation of network transmission costs.

[0067] The task type code and data security level are extracted from the structured task dataset. For computing tasks with a data security level marked as high security (i.e., requiring physical isolation), the matching module retrieves the static attribute database of global data center nodes. The matching module identifies the physical security attribute labels of each candidate data center. If a candidate data center does not have independent physical hardware isolation qualifications, the corresponding candidate data center is directly removed from the current task's selectable list. Based on the physical distance parameter in the network topology data stream, the matching module extracts the physical distance between the task initiator and each candidate data center. .

[0068] As a preferred approach, for latency-sensitive tasks, the system presets a maximum maximum distance. The furthest limit distance. The specific setting logic is as follows: the system extracts the expected response time of the current computing task. After deducting the estimated data processing latency performed within the candidate data center, the remaining available communication latency is obtained. This remaining available communication latency is then multiplied by the reference fiber optic propagation speed coefficient to calculate the limiting physical distance parameter. If the physical distance... Greater than or equal to the farthest limit distance If a candidate data center fails to meet the physical constraints, a hard rejection operation is performed to ensure that latency-sensitive tasks are ultimately retained in the candidate data center set. All of these are edge candidate data centers that meet the extreme latency requirements. After filtering using hard rules, the system generates a set of candidate data centers that satisfy the basic constraints. If the candidate data center set If the set is empty, the system will trigger a degraded scheduling strategy or return a resource exhaustion exception to the task initiator.

[0069] For candidate data center sets The system utilizes multiple network characteristic parameters to perform a multi-dimensional dynamic evaluation of network transmission costs for the remaining candidate data centers. Any candidate data center in The matching module calculates its normalized network transmission cost. Its calculation formula is defined as: ; in, Indicates candidate data centers The normalized network transmission cost, the larger the value, the worse the network transmission performance or the higher the physical communication cost. Indicates candidate data centers The physical distance parameter after normalization. Indicates candidate data centers Real-time bandwidth utilization after normalization. Indicates candidate data centers Network latency fluctuation parameters after normalization. Indicates candidate data centers The bandwidth instability parameter, of which Indicates candidate data centers The bandwidth stability metrics after normalization. These four normalized physical quantities are all mapped to a standard numerical space between zero and one, and the mapping process is performed based on the maximum and minimum range of the system's historical statistical values. , , as well as These represent the weighting coefficients for the normalized physical distance parameter, real-time bandwidth utilization, network latency fluctuation parameter, and bandwidth instability parameter, respectively. The system limits the values ​​of these four weighting coefficients to an open interval between zero and one, and the sum of all four is strictly equal to one. In actual engineering deployments, these four weighting coefficients can be obtained through static matrix calibration using the Analytic Hierarchy Process (AHP) combined with the prior experience of network operation and maintenance experts, or they can be automatically fitted and optimized using machine learning algorithms by inputting a historical network transmission fault sample dataset.

[0070] After completing the network transmission cost assessment, the matching module calculates the comprehensive matching weight by combining task characteristics with the resource status of candidate data centers. (Comprehensive Matching Weight) The calculation formula is defined as follows: ; in, Represents the calculated candidate data centers The overall matching weight. Indicates task type and candidate data center The degree of matching of abilities. Indicates candidate data centers The balance between electricity prices and network costs. Indicates candidate data centers The priority level for green electricity consumption is directly equivalent to that of candidate data centers. Real-time percentage of green electricity supplied. , as well as These represent the static weight coefficients assigned to the three matching factors: the matching degree between task type and candidate data center capabilities, the balance between electricity price and network cost, and the priority level of green electricity consumption. The sum of these three factors is strictly equal to one. These three static weight coefficients are obtained by: assigning initial baseline values ​​using the Delphi method based on the system's current macro-strategic orientation towards economic efficiency (pursuing low prices), environmental protection (pursuing renewable energy consumption), or task performance at different operational stages; and then fine-tuning them through training on historical successful scheduling datasets.

[0071] Based on task type and candidate data centers Compatibility of abilities The system calculates the similarity based on the cosine similarity of supply and demand characteristics, and its mathematical formula is: ;in, This represents the preset computing power task requirement feature vector (and each dimension has been pre-mapped to a positive number space from zero to one). Indicates candidate data centers The provided computing power feature vector (after equal normalization processing). Represents the dot product of two vectors. and Representing the two eigenvectors respectively Norm (i.e., vector length). This is the fourth smallest positive offset, used to prevent computational overflow caused by the eigenvector approaching zero.

[0072] For candidate data centers The balance between electricity prices and network costs Its calculation formula is defined as: ;in, Indicates candidate data centers The real-time settlement price of electricity in the region. Indicates candidate data centers Normalized network transmission costs. The fifth smallest positive offset, the above and All of these are determined based on the lower limit of floating-point arithmetic precision of the underlying deployed hardware (the conventional value is set to...). ). Represents the current set of candidate data centers The theoretical minimum overall cost of all nodes in the process is given by the value of ,in A mathematical operator that represents the minimum value; This represents the local index variable used to traverse the set of candidate data centers. The formula, while maintaining the inverse logic of economic optimization (i.e., the lower the electricity and network costs, the higher the matching degree), will... The output range is normalized to the interval (0, 1).

[0073] The matching module performs descending sorting and space reduction operations on the candidate set based on the calculated comprehensive matching weights of each candidate data center, forming a high-quality preliminary scheduling candidate subset. The system then selects the candidate data centers remaining after filtering by hard rules. In the middle, based on the comprehensive matching weight The values ​​are sorted in descending order, and the top-ranked values ​​are extracted. The nodes in this initial scheduling candidate subset are defined by the following dynamic truncation rule: ;in, This represents the reduced initial scheduling candidate subset. Indicates the set of candidate data centers Extract the top-ranked items based on comprehensive matching weight. Mathematical operators for bit nodes. Indicates the task is in the candidate data center The comprehensive matching weight. In engineering implementation, this parameter... The initial scheduling candidate subset, which can be dynamically set by the system to an integer between 5 and 10 based on real-time computing load, includes nodes with high matching degree. After generation, This will be used as the boundary of the simplified initial solution space and accurately pushed to the next level of the five-dimensional reward optimization control flow based on the Pareto front. This provides basic data support for retaining high-quality solutions in subsequent multi-objective Pareto optimization based on electricity price economy, green power consumption, time delay satisfaction, task value and data security.

[0074] See attached document Figure 3 In step S5, the optimization module receives the candidate data center set and obtains the dynamic multi-objective optimization weight vector. A multi-dimensional independent utility evaluation system is constructed for the remaining candidate data center nodes, and a Pareto front-based multi-objective optimization decision-making procedure is executed.

[0075] The system utilizes the established five-dimensional reward framework to independently calculate the normalized reward in five dimensions for each candidate data center node in the candidate data center set.

[0076] The first dimension is the electricity price economic incentive. For any candidate data center, its associated estimated electricity cost and network transmission cost are extracted. The formula for calculating the electricity price economic incentive is as follows: ;in, This represents the calculated economic incentive for electricity pricing. The closer the calculated result is to a constant of one, the more significant the economic benefit. This indicates a lower limit truncation operator, used to forcibly lower the economic reward to zero when the current estimated total cost exceeds the historical extreme value. This represents the projected electricity cost of the target candidate data center, which is derived by multiplying the real-time node unit price by the estimated energy consumption. This represents the quantified network transmission cost, which is derived by multiplying the previously estimated amount of network transmission data by the real-time network bandwidth unit price. This represents the maximum single-task cost benchmark maintained by the system based on historical data. Its specific value is dynamically updated by periodically analyzing the actual cost of a single task in the historical operation logs and taking the maximum value. To ensure the robustness of the computing architecture and prevent division overflow from occurring in the very early stages of system operation when the historical maximum cost benchmark has not been effectively established, a sixth minimum positive offset is introduced. The standard configuration is .

[0077] The second dimension is the green electricity consumption reward. The formula for calculating the green electricity consumption reward is defined as follows: ;in, This indicates an incentive for green electricity consumption. This indicates the actual amount of green electricity that the target candidate data center can provide to the corresponding computing power task at the current scheduling moment. This represents the estimated total power consumption during the execution of the computing task. Based on the same overflow prevention mechanism, a seventh minimal positive offset is introduced. Standard configuration and Keep it consistent, set to .

[0078] The third dimension is the latency satisfaction reward. The system extracts the latency requirements of the computing task and the estimated end-to-end total execution latency, where the end-to-end total execution latency strictly includes the sum of network transmission latency and node internal data processing latency. As a preferred method, the latency satisfaction reward is calculated based on continuous exponential decay logic, and its mathematical formula is defined as a piecewise function: When the estimated total end-to-end execution latency is less than or equal to the latency requirement of the computing power task, the requirement is met. The system assigns a delay satisfaction reward based on the given conditions. The value is equal to a constant. ; When the estimated total end-to-end execution latency is significantly greater than the latency requirement of the computing task, i.e., when... Under the physical operating conditions, the formula for calculating the delay satisfaction reward is: ;in, This represents the reward for meeting the time delay satisfaction requirement. This represents the estimated total end-to-end execution latency. The threshold value representing the explicit latency requirement of the computing task is extracted directly from the request metadata of the computing task during the data parsing phase. This represents the delay sensitivity attenuation coefficient. The numerical value is directly related to the tolerance of the computing power task itself, and its range is set within... to The value is determined by the system through static matching of a preset tolerance and attenuation coefficient mapping table; a larger value indicates a steeper penalty mechanism. To ensure the attenuation coefficient is effective within a reasonable range, all delay values ​​are uniformly converted to standard units in seconds before being substituted into the formula.

[0079] The fourth dimension is the task value reward. The formula for calculating the task value reward is defined as follows: ;in, This indicates the reward for the task. This represents the comprehensive value score of the computing power task calculated during the classification and evaluation phase. This represents the maximum value limit defined by the system. Its specific value is pre-determined by summing the maximum weighted combination of all predefined business value, user impact, and data value characteristics. This represents the eighth smallest positive offset, as stated above. All of these are determined based on the lower limit of floating-point arithmetic precision of the underlying deployed hardware (the default is usually...). ).

[0080] The fifth dimension is data security rewards. The system determines compliance based on strict isolation qualification comparison logic, and its calculation formula is defined as follows: ;in, This indicates a data security reward. Represents a mathematical characteristic function. Indicates the actual physical or logical security level of the target candidate data center (converted to the corresponding integer value by calling the same quantization mapping dictionary). This indicates the security level required for the computing power task. The system pre-sets a security level quantization mapping dictionary, extracts the discrete labels generated in the previous task classification steps, and maps high security, medium security, and low security to integer values ​​3, 2, and 1 respectively, assigning them to [the appropriate values]. Perform logical operations. When the security level value of a node after mapping is greater than or equal to the security level value required by the computing power task, the mathematical characteristic function outputs a constant one; otherwise, it outputs zero.

[0081] After obtaining independent utility rewards across five dimensions, the system constructs a global multi-objective utility evaluation benchmark by combining dynamically input weight vectors. The system then comprehensively calculates the weighted total utility score for each candidate data center. ; in, This represents the weighted total utility score of the target candidate data center obtained through calculation. This indicates an economic incentive in electricity pricing. This indicates an incentive for green electricity consumption. This represents the reward for meeting the time delay satisfaction requirement. This indicates the reward for the task. This indicates a data security reward. to These correspond to the weights for electricity price economics, green electricity consumption, time delay satisfaction, task value, and data security, respectively. These five consecutive weight coefficients are directly derived dynamically from the meta-learning mechanism, and their values ​​all fall within a closed interval of zero to one. Furthermore, the sum of the five weight coefficients is strictly equal to a constant of one.

[0082] The system's underlying layer first performs a coarse screening and boundary delineation based on Pareto multi-objective optimization. Within the candidate data center set, the system iterates pairwise, comparing the independent utility reward performance of each candidate data center across dimensions such as electricity price economy, green energy consumption, latency satisfaction, task value, and data security. Based on the core mathematical principle of non-dominated sorting, if and only if the node... The node is greater than or equal to all five evaluation dimensions. Furthermore, there exists at least one evaluation metric that is strictly greater than the node. At that time, the system determines the node in terms of physical state. Fully Dominated Node The system undergoes a full cross-validation process, automatically removing all dominated and inferior solution nodes from the solution space, and defining the remaining non-dominated node group as the Pareto front non-dominated solution set. For the iterative comparison and hierarchical partitioning operations of the underlying array pointer cursors in this non-dominated sorting algorithm, those skilled in the art can implement it using a fast non-dominated sorting criterion with an elitist retention strategy. The determination process of its high-dimensional vector dominance relationships is a well-known technique in the field and will not be elaborated upon here.

[0083] After constructing a high-quality Pareto front non-dominated solution set through non-dominated ranking in the physical dimension, the system then reduces the dimensionality of the multidimensional utility space to a scalar space to establish a unique target suitable for the current environment among the high-quality solutions that cannot dominate each other. The system then retrieves the weighted total utility score corresponding to each front node in a targeted manner. The system uses a sorting algorithm to select the unique node with the highest weighted total utility score and formally establishes it as the optimal execution target for the current computing power task.

[0084] Based on the established optimal execution target, the system encapsulates the underlying scheduling metadata to generate a scheduling decision to be verified, containing a strict data structure. The underlying data structure of the scheduling decision to be verified must embed a globally unique identifier for the computing power task, the physical address information of the target data center, the expected start and end timestamp intervals, the expected green electricity quota ratio parameters, and a tamper-proof digital signature generated by encryption using the system's private key. For the underlying cryptographic operations used to generate the tamper-proof digital signature using the system's private key, those skilled in the art can employ standard RSA or elliptic curve digital signature algorithms. The underlying key calculation and signature encapsulation processes are well-known technologies in the field and will not be elaborated upon here. The system then pushes the encapsulated scheduling decision to be verified into the subsequent audit network to trigger the multi-party verification and consensus confirmation process of the consortium blockchain nodes.

[0085] See attached document Figure 1 and attached Figure 2 In step S6, the audit module receives the scheduling decision to be verified. This triggers a trustworthy scheduling decision and consensus-based rights confirmation mechanism based on the consortium blockchain.

[0086] Based on underlying physical considerations to prevent single-point cheating and malicious tampering of critical data, the system will verify the scheduling decision. The broadcast is sent to each verification node in the consortium blockchain network. In this embodiment, each verification node executes isolated automated rule verification based on the underlying distributed smart contract. Specifically, the transaction center node participating in the verification extracts the quoted electricity price data and the expected proportion of green electricity used in the scheduling decision to be verified, and retrieves real-time external real electricity price data and green electricity output prediction data synchronized by the on-chain oracle, thereby executing a multi-dimensional factor consistency verification function. For the underlying code writing of the consortium blockchain distributed smart contract and the data synchronization of the on-chain oracle, those skilled in the art can rely on the standard application programming interfaces provided by mature open-source blockchain frameworks (such as Hyperledger Fabric), and the cross-chain data acquisition and contract deployment process is a well-known technology in the field, and will not be described in detail here.

[0087] As a preferred approach, the underlying logic of the electricity price consistency verification function is defined as follows: Meanwhile, the logic of the green electricity consistency verification function is defined as follows: ; in, and These represent the electricity price consistency verification function and the green electricity consistency verification function, respectively. When the Boolean value of the logical expression in parentheses is true, a verification pass signal is output. This indicates the quoted electricity price data referenced in the dispatch decision to be verified. Real-time external electricity price data synchronized with on-chain oracles The absolute difference between them. This represents the projected proportion of green electricity used in the scheduling decisions to be verified. Real-time external green electricity forecast data synchronized with on-chain oracles The absolute difference between them. and These represent the pre-defined minimum price tolerance and the minimum tolerance for the proportion of green electricity, respectively. Considering the minimum precision unit of the electricity market spot trading system and the reasonable microsecond-level delay in network data synchronization, this minimum price tolerance value... The range of values ​​is usually constrained to . to Within the closed interval, Constrained to Within the closed interval. For the aforementioned price tolerance minimum value. Minimum tolerance value for green electricity ratio The specific value setting logic can be obtained by system administrators through offline static calibration based on the statistical variance data of the system's historical network communication delay and their experience.

[0088] While electricity pricing and green electricity verification are conducted simultaneously, computing power service provider nodes delve into the underlying layers to probe whether the computing resources of the target data center are truly idle; security nodes, on the other hand, use cryptographic principles to verify the validity of security level matching signatures. For the physical probe of the true idle status of underlying computing resources and the cryptographic verification of security level matching signatures, those skilled in the art can call the native resource monitoring interface of the data center operating system and the signature verification function of the asymmetric encryption algorithm. The underlying resource activation and signature verification execution logic is well-known technology in this field and will not be elaborated upon here.

[0089] After completing local physical verification of the multi-source heterogeneous nodes, to ensure that the cluster as a whole can still reach a consensus on resource allocation decisions even when some network nodes experience high communication latency or Byzantine hardware failures, a practical Byzantine fault-tolerant consensus process is initiated. During the pre-preparation and setup phases, each network node broadcasts its independent verification results to the entire peer-to-peer network. For full broadcast and inter-node communication mechanisms in peer-to-peer networks, those skilled in the art can implement them using standard P2P network transmission components combined with the Gossip protocol. The underlying network routing and message broadcasting principles are well-known technologies in this field and will not be elaborated here.

[0090] Upon entering the submission phase, to establish a quantifiable global trust assessment system and uniformly execute consensus judgments, the system aggregates and statistically calculates the decision credibility of the continuous state based on verification details. Its mathematical formula is defined as follows: ;in, This indicates the reliability of the decision based on the calculated output. This represents the total number of nodes in the consortium blockchain network participating in this consensus process. This is a mathematical operator that performs a cyclic accumulation on all nodes participating in the verification. Represents a mathematical characteristic function, when the internal state equation If true, the output constant is 1; otherwise, the output constant is 0. Indicates the first The actual verification results broadcast by each verification node in the network, whose discrete state value set is strictly defined as... . A valid status identifier indicating that the verification logic has passed completely. This indicates an invalid state that failed verification. At least tolerate (One malicious node).

[0091] The rule for determining consensus is: when the credibility of the decision is high. Greater than or equal to the configured preset trust threshold When (i.e., satisfying) At the physical level, the cluster determines the current scheduling decision to be verified. An absolute consensus was reached, and this was defined as a globally trustworthy state. In this embodiment, a preset trust threshold is used. The value of is determined by combining the Byzantine fault tolerance theory with the safety boundary setting, and it is automatically defined as . The molecular structure is mathematically rigorously derived using the floor operator in the practical Byzantine fault tolerance mechanism. Required under the system The minimum honest node consensus lower bound, a constant. ,constant With constant Together, a fault tolerance ratio boundary was constructed. This represents the integer rounding operator. Administrators can also increase this threshold in advance in the background based on higher security requirements, but its upper limit must be strictly constrained not to exceed the theoretical maximum proportion of currently alive and healthy nodes in the network.

[0092] Based on the established globally trusted state, the underlying smart contract triggers state machine transition instructions, automatically updating the scheduling decisions to be verified. Transform and encapsulate into effective scheduling instructions The system invokes a cryptographic hash function, performs a combined hash mapping on the serialized data set of the effective scheduling instruction along with the current physical timestamp, generates a transaction hash certificate with unique time-series characteristics across the entire network, and then issues the effective scheduling instruction. The main data and credentials are pushed together into the new block structure. For the byte concatenation of the underlying serialized data and the collision-resistant processing of the standard hash function, those skilled in the art can rely on mature distributed ledger component libraries; the cryptographic encapsulation process is well-known in the field and will not be elaborated here. With the successful broadcast and appending of this new block structure to the entire network's distributed ledger, cross-domain computing power scheduling instructions are officially endowed with immutable physical attributes, thus constructing a fully traceable data execution closed loop in an untrusted network environment.

[0093] See attached document Figure 4 In step S7, the feedback module receives the effective scheduling instruction with the transaction hash certificate and executes the scheduling execution monitoring and system self-closed loop feedback optimization program.

[0094] In this embodiment, the target data center's control system monitors new block events generated in the consortium blockchain network in real time through an underlying event monitoring daemon. Once a valid scheduling instruction containing a unique identifier and a transaction hash certificate is detected on the blockchain, the target data center's hardware controller immediately parses the metadata carried by the instruction, allocates predetermined central processing unit and memory resources in the physically isolated computing cluster, and initiates the actual processing of the computing task. During the execution of the computing task, the underlying system strictly locks and deducts the committed green electricity consumption share according to the instruction.

[0095] The feedback module constructs a dual feedback closed-loop mechanism for monitoring prediction errors and enhancing actual utility. Simultaneously, the feedback module initiates a real-time monitoring mechanism for the underlying physical state, invoking a hardware-level monitoring agent deployed within the target data center to periodically capture actual operational metrics. The feedback module accurately collects the actual completion latency of computing tasks, the actual settlement price returned by the power grid settlement terminal during the computing task execution period, and the actual green electricity consumption statistics from physical meters. For the underlying hardware sensor network's use of out-of-band management interfaces to read central processing unit operating parameters and actual power consumption, those skilled in the art can employ standard system-level liveness detection and resource acquisition components. The hardware-level performance data reading and network transmission processes are well-known technologies in the field and will not be elaborated upon here.

[0096] After the computing task is completed and the calculation results are returned, the feedback module summarizes all actual operating metrics and performs data closed-loop optimization according to the following two dimensions: Supervised optimization of the perception layer based on prediction bias. The feedback module cross-compares the actual operating indicators with the estimated operating parameters generated in the early multi-objective optimization phase of the system. Before performing the numerical comparison, the feedback module performs time axis alignment based on the globally unique identifier of the computing power task and the standard timestamp protocol, and performs unified normalization processing based on extreme values ​​on the extracted actual indicators and estimated parameters. Subsequently, the feedback module performs multi-dimensional comprehensive bias feedback signal calculation based on squared error. The calculation formula for this multi-dimensional comprehensive bias feedback signal is: ; in, This represents the scalar value of the calculated multidimensional integrated deviation feedback signal. This represents the actual completion delay of the computing task after normalization. This represents the system's estimated total end-to-end execution latency. This represents the normalized value of the actual cost incurred after the computing task is completed. This represents the normalized value of the system's estimated consumption cost. This represents the time delay deviation penalty coefficient set statically by the system. This represents the cost deviation penalty coefficient set statically by the system. and The threshold values ​​for the penalty coefficient are all constrained to the real number range of zero to one, and their sum is strictly equal to a constant of one. Regarding the method for obtaining the penalty coefficient threshold, management personnel obtain it through offline static calibration based on the relative importance of business time tolerance and cost sensitivity, combined with historical scheduling data. The offline grid search optimization process for the aforementioned penalty coefficient threshold can be implemented by those skilled in the art using conventional hyperparameter tuning methods; the calibration of its model hyperparameter thresholds is a well-known technique in this field and will not be elaborated upon here.

[0097] The decision-making layer reinforces rewards based on actual utility. For the dynamic multi-objective optimization weight vector output by the meta-learning module, the feedback module recalculates the actual comprehensive business utility scalar value generated by this scheduling in the real environment based on the collected actual physical indicators, and uses this as the feedback signal of the actual comprehensive business utility of reinforcement learning. The formula for calculating this actual comprehensive business utility feedback signal is as follows: ; in, This represents the calculated scalar value of the actual comprehensive business utility. , , , , These represent the weight coefficients for electricity price economy, green electricity consumption, time delay satisfaction, task value, and data security output by the meta-learning basis model, respectively. This represents the economic incentive value of the actual electricity price calculated based on the actual settlement price and the actual electricity consumption. This represents the actual green electricity consumption reward value derived from statistics based on physical electricity meters. This represents the actual delay satisfaction reward value calculated based on the actual task completion delay. This represents the task value reward value determined based on the task's multidimensional attributes. This represents the data security reward value determined based on the matching degree between the node's physical isolation attributes and task requirements.

[0098] After calculating and obtaining the above signals, the complete context of this scheduling is encapsulated into a standard Markov decision process experience quadruple, which contains the current comprehensive environment feature vector, the execution action (dynamic multi-objective optimization weight vector), the reward feedback (i.e. the actual comprehensive business utility feedback signal), and the comprehensive environment feature vector at the next moment, and stored in the experience replay pool at the bottom layer of the system.

[0099] Periodically, the meta-learning module extracts batches of Markov decision process experience quadruples from the experience replay pool and executes a continuous optimization loop based on meta-learning. In the specific model training fine-tuning steps, the meta-learning module performs multi-task gradient updates: on the one hand, it uses the multi-dimensional comprehensive bias feedback signal as a supervision signal to update the meta-learning base model parameters; on the other hand, based on reinforcement learning algorithms, it uses the actual comprehensive business utility feedback signal as a reward signal to guide policy evolution, optimizing the policy output network with the goal of maximizing accumulated real business utility. The meta-learning module calculates the meta-gradient of the composite loss function for neurons within each hidden layer of the underlying neural network using the backpropagation algorithm, and employs an adaptive moment estimation optimizer, combined with parameters set within a very small positive number interval (the specific parameter threshold is set according to empirical rules). to The parameters are fine-tuned by adjusting the local learning rate step size within the range of the interval. The mathematical implementation details of the adaptive moment estimation optimizer and the parameter update process of backpropagation chain differentiation can be easily achieved by those skilled in the art through calling the built-in functions of existing deep learning frameworks. The gradient derivation and update operations at the model's underlying level are well-known techniques in this field and will not be elaborated upon here.

[0100] By deeply integrating the closed-loop mechanism of environmental prediction correction and real benefit feedback, the subsequent dynamic multi-objective optimization weight vector generation strategy is continuously improved, thus building a robust automated computing and power collaborative scheduling optimization platform that continuously pursues the maximization of real business utility in unreliable external power and communication environments.

[0101] See attached document Figure 5 and Figure 6 In this embodiment, the computing power application terminal is specifically a 5G edge computing node for intelligent connected vehicles deployed in the Beijing area, which initiates a hybrid computing task of autonomous driving trajectory prediction and visual rendering with a concurrent scale of 500 megabytes.

[0102] The underlying network probe of the data processing module captures the computing task data stream sent by the 5G edge computing node of the intelligent connected vehicle in real time through a bound multi-source data interface (the specific application layer communication protocol adopts the Message Queue Telemetry Transport Protocol (MQTT) based on the publish-subscribe model, with the access port set to 1883). The data processing module extracts basic task information, obtaining the task identity identifier as AD-Task-9527, the task type as high-frequency interactive inference, the business type as vehicle-to-everything (V2X) strong real-time control, and the data size as precisely measured at 500MB. The data processing module simultaneously extracts additional attributes, parsing out a data sensitivity identifier value of 0.85 and a business urgency index of 0.92. Through a standard Hypertext Transfer Protocol (HTTP) application programming interface, it sends a request message with a time-series digital signature to the State Grid external power trading center node to retrieve real-time electricity prices and wind and solar power output forecasts for Beijing and surrounding areas (such as the data center clusters in Ulanqab, Inner Mongolia, and Zhangjiakou, Hebei). Since the meteorological forecast data is updated once per hour, a forward filling algorithm based on a zero-order hold is adopted to maintain the state of the wind speed forecast value (8.5 m / s) for the Zhangjiakou area acquired in the most recent hour and map it to the current millisecond-level computing power task timestamp node.

[0103] For all continuous numerical input variables, the data processing module calls the min-max normalization algorithm to perform dimensionless processing. Taking the real-time network bandwidth utilization rate of nodes in the Zhangjiakou area as an example, the actual measured values ​​are obtained. for Historical lower limit extreme value for Historical extreme values for The first minimum positive offset is set. for Substitute into the formula The normalized absolute value is obtained. for The preprocessed input dataset is serialized into a multidimensional matrix data structure and directly written to an independent physical memory segment based on a shared memory mechanism.

[0104] The task classification module mounts and reads the preprocessed input dataset from the shared memory region via inter-process communication. It extracts the task type, business type, and data size, and performs latency sensitivity quantitative analysis. A pre-built type mapping table includes the feature functions of high-frequency interactive inference tasks. Mapped to 0.9, the characteristic function of the vehicle-to-everything (V2X) real-time control service type is... The mapping is 0.95. Data size characteristic function. Calculate using logarithmic scaling logic, substituting the current data size. 500MB, historical maximum and minimum values For 10000MB, 1MB, the second smallest positive offset Pick Calculation Approximately 0.69. This is combined with the weighting coefficients determined by expert static calibration. , , Substitute into the time delay sensitivity index calculation formula Calculate the time delay sensitivity index It is 0.878.

[0105] The system's built-in inverse delay mapping function (specifically a linear interpolation inverse model) maps this exponent to the expected response time. Based on the threshold truncation rule, this task is forcibly locked by the system and discretized into a latency-sensitive type. The task classification module calls a multilayer perceptron secure mapping model with a sigmoid activation function. Given an input data sensitivity flag value of 0.85, a data size of 500MB, and a task type scalar, the output data security level score is 0.82. Since 0.82 is greater than the first preset classification threshold... The task was coded as high-security (corresponding to physical isolation requirements). The resulting structured task dataset and its type encoding (latency-sensitive, high-value, high-security) were persistently stored in the master-slave constraint table of the underlying relational database (using the MySQL InnoDB storage engine).

[0106] The meta-learning module extracts environmental feature components from the current North China power grid region, finding that the Beijing area is experiencing peak electricity consumption, leading to a significant increase in electricity prices, while the Zhangjiakou area is affected by strong winds at night, resulting in a surge in green power output and redundancy. The combined environmental feature vector is then propagated forward through a multilayer perceptron basis model and output through a Softmax layer to generate a dynamic multi-objective optimized weight vector. The output weight components at this point are: electricity price economic weight. Green electricity consumption weight Delay satisfaction weight Task value weight Data security weight The dynamic multi-objective optimization weight vector When a shift occurs, green electricity is proactively assigned an extremely high scheduling priority.

[0107] The matching module performs candidate filtering based on spatiotemporal network topology. Given that the task is designated for high security, it retrieves the static attribute database of data center nodes and directly performs hard filtering of ordinary commercial public cloud nodes that are not configured with independent physical server cluster isolation schemes. This addresses latency requirements. After deducting the estimated internal calculation delay of 15ms within the node, the remaining usable communication delay is 50ms. The maximum maximum distance is then calculated based on the fiber optic propagation speed coefficient. The distance is approximately 5000 kilometers. Both the Ulanqab and Zhangjiakou nodes meet this physical distance constraint and are retained in the candidate data center set. The matching module calculates the comprehensive matching weight for each node. Taking a candidate data center in Inner Mongolia as an example, cosine similarity is calculated based on the feature vectors of demand and supply capacity. And combine the balance between electricity price and network cost and green electricity ratio The matching weights are calculated using a weighted algorithm, and the results are ranked in descending order to determine the top performers. position (for) Figure 5 The discrete sample size in The initial scheduling candidate subset is set to 50. Each candidate entity in this subset is identified as a candidate data center node.

[0108] The optimization module performs five-dimensional independent utility reward calculations for 50 candidate data center nodes in the initial scheduling candidate subset. For example, for a fully naturally cooled green data center in Zhangjiakou, where the actual green energy supply ratio reaches 95%, the calculated green electricity consumption reward... The value is 0.95, and its estimated end-to-end total execution latency is 45ms, which is less than the task latency requirement of 65ms, thus meeting the requirement. Conditions, latency satisfaction reward Assigned a constant of 1, its physical security level is 3, which is greater than or equal to the task requirement of 3, and a data security bonus is awarded. The output constant is 1. The optimization module multiplies the independent rewards of these five dimensions by dynamic weights. to The weighted total utility score is obtained by weighted summation. .

[0109] See attached document Figure 5 The above 50 candidate data center nodes (for Figure 5The hollow circle (a point in the graph) is projected onto the multi-objective optimization space, where the horizontal axis represents the normalized network and computational latency cost, and the vertical axis represents the normalized total execution cost. A non-dominated sorting algorithm is used to automatically delineate the Pareto front non-dominated solution set within this space, marked by squares and connected by dashed lines. Subsequently, the optimization module searches for a weighted total utility score within the front solution set. Maximize the intersection points and establish Figure 5 The optimal execution target is marked with a black pentagram. Based on this optimal target, the underlying layer encapsulates a verifiable scheduling decision containing a globally unique identifier and an RSA digital signature. .

[0110] The audit module broadcasts the scheduling decisions to be verified to the consortium blockchain verification nodes deployed at various computing power hubs via a P2P network. The distributed smart contract deployed on-chain (written in Go and running in a Docker container) then calls... and The verification function compares the on-chain message data with the oracle price feed data; the absolute difference between the two does not exceed the set minimum tolerance value. and The cluster initiates a practical Byzantine fault tolerance mechanism, based on the total number of consensus nodes. In this environment, three valid votes were received, and the credibility of the decision was calculated. Preset confidence threshold Calculated according to the formula If the decision meets the criteria, the decision is classified as a globally trustworthy state. The underlying smart contract calls the SHA-256 hash function to generate a TxHash credential, which is then packaged to generate an effective scheduling instruction. .

[0111] The control system of the Zhangjiakou data center detected the new block with TxHash, allocated the corresponding GPU computing resources through the underlying system-level application programming interface, and locked the wind power quota at the physical level. After the task was completed, the hardware-level monitoring agent reported the actual completion latency as 48ms (estimated 45ms). The feedback module calculated the squared error between the actual indicator and the estimated parameter, generating a comprehensive deviation feedback signal. Figure 6 As shown in (a), the comprehensive deviation feedback signal converges rapidly with the increase of the horizontal axis network iteration update period. Based on the actual settlement data returned by the physical meters and the grid interface, the feedback module recalculates the actual comprehensive utility of this scheduling. Figure 6 As shown in (b), the actual overall utility exhibits a robust increase over the iteration cycle. The data from the complete process are encapsulated into Markov decision process empirical quadruples and pushed into the empirical replay pool. Using an adaptive moment estimation optimizer, the local learning rate step size is set to... The gradient is calculated based on backpropagation of the multi-task composite loss function, and the weights of the deep multilayer perceptron network are fine-tuned.

[0112] See attached document Figure 5 The massive number of heterogeneous candidate data center nodes exhibit a nonlinear divergence in the physical limits of normalized total execution cost and normalized network and computation latency costs. The system not only reduces dimensionality to eliminate... Figure 5 The large number of high-latency, high-cost redundant nodes in the upper right corner rely on a five-dimensional multi-objective weight vector dynamically injected through meta-learning to anchor the physical intersection of the unique optimal solution interval within the Pareto front non-dominated solution set formed by square markers. Figure 5 The optimal execution target.

[0113] Combined with appendix Figure 6 The data performance of the dual feedback closed-loop characteristics further demonstrates that the traditional architecture causes the failure of prediction parameters due to cross-domain physical network delays and sudden changes in power grid load, resulting in misalignment of the underlying reinforcement learning credit allocation. Figure 6 (a) shows that after approximately 40 network iterations, the supervision signal constructed based on the squared error enabled the model feature extraction network to accurately align with the actual physical loss, and its comprehensive bias feedback signal quickly fell below the baseline of 0.1; at this point, as Figure 6 As shown in (b), the actual physical utility of the underlying implementation then shows a stable ramp-up and eventually converges. The data clearly reveals that, relying on the prediction error supervision and actual utility enhancement parallel update mechanism with rigorous closed-loop logic, the underlying multilayer perceptron-based model is able to escape the local optimum trap of prediction distortion. Ultimately, it spontaneously masters a flexible collaborative control strategy that can continuously maximize the real benefits of business in unreliable physical environments such as external power grid fluctuations and network jitter, fully supporting the absolute security and efficient scheduling of massive heterogeneous resources in the deep integration environment of the new generation computing network and power grid.

Claims

1. A computing-powered collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity, characterized in that, Includes the following steps: Acquire multi-source input data, including at least computing power task data streams, electricity market data streams, and network topology data streams; Based on the computing power task data stream, the latency sensitivity index and comprehensive value score of the computing power task are evaluated, and the computing power task is classified in multiple dimensions by combining the latency sensitivity index and the comprehensive value score to generate task type code. Based on the power market data stream, the current market environment characteristics are extracted and a multi-objective optimization weight vector is generated. Based on the network topology data stream, multiple data center nodes are obtained and data center nodes that do not meet the constraints are removed to generate a candidate data center set. Combining the task type encoding and the multi-objective optimization weight vector, the comprehensive collaborative utility of the data center nodes in the candidate data center set is calculated, and a scheduling decision to be verified is generated. The scheduling decision to be verified is automatically verified, and an effective scheduling instruction is generated after the verification is passed. Monitor the execution of the computing power tasks according to the effective scheduling instructions, collect actual indicator parameters, fine-tune the parameters based on the actual indicator parameters, and complete the computing power collaborative scheduling.

2. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, The multi-source input data also includes security and value assessment data streams; After acquiring the multi-source input data, the following steps are performed: performing dimensionless mapping on the continuous values ​​in the multi-source input data, and concatenating them to generate a preprocessed input dataset.

3. The computing-powered collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, The specific steps for generating task type codes are as follows: The data security level of the computing power task is determined by discretizing the latency sensitivity index, the comprehensive value score, and the data security level. Based on the time-delay sensitivity index, the comprehensive value score, and the data security level after discretization, the multi-dimensional classification process is performed according to the feature cross-combination logic to map the computing power task into a structured task dataset and the task type code.

4. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 3, characterized in that, The specific steps for evaluating the latency sensitivity index of the computing power task are as follows: extract the task type, business type, and data size of the computing power task; sum the products of the task type feature function corresponding to the task type and the corresponding first static weight coefficient, the business type feature function corresponding to the business type and the corresponding second static weight coefficient, and the data size feature function corresponding to the data size and the corresponding third static weight coefficient to obtain the latency sensitivity index. The specific steps for evaluating the comprehensive value score of the computing power task are as follows: extract the business value score, user impact score, and data value score of the computing power task; sum the products of the business value score and the corresponding first global weight coefficient, the user impact score and the corresponding second global weight coefficient, and the data value score and the corresponding third global weight coefficient to obtain the comprehensive value score. The specific steps for determining the data security level of the computing power task are as follows: extracting the data sensitivity identifier value of the computing power task, using the data sensitivity identifier value, the data size, and the task type as joint judgment input variables, calling a preset security level evaluation function to perform security evaluation calculation on the joint judgment input variables, and outputting the data security level.

5. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 4, characterized in that, The data size characteristic function corresponding to the data size is calculated using logarithmic scaling logic. The calculation process is as follows: Obtain the preset minimum value and the preset maximum value of the data scale, and obtain the first preset constant and the second preset constant; Calculate the logarithm of the sum of the data size and the first preset constant, and subtract the logarithm of the sum of the preset minimum data size and the first preset constant to obtain the first difference; Calculate the logarithm of the sum of the preset maximum data size and the first preset constant, and subtract the logarithm of the sum of the preset minimum data size and the first preset constant to obtain the denominator value; Calculate the ratio of the first difference to the denominator value, and subtract the ratio from the second preset constant to obtain the value of the data scale characteristic function.

6. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, The specific steps for dynamically generating a multi-objective optimization weight vector based on the current market environment characteristics are as follows: The current market environment characteristics are used to construct a current comprehensive environment feature vector, and historical environment scene feature vectors are extracted from the historical environment feature library. Calculate the inner product of the current integrated environment feature vector and the historical environment scene feature vector; calculate the product of the magnitude of the current integrated environment feature vector and the magnitude of the historical environment scene feature vector. Calculate the ratio of the inner product value to the product value to obtain a similarity score. Based on the maximum similarity score, extract the starting point of the corresponding pre-trained meta-model parameters from the historical environment feature database. Obtain sample data of the current electricity market environment, a preset local learning rate step size, and a preset loss function; Based on the sample data, calculate the partial derivative gradient of the loss function with respect to the starting point of the pre-trained meta-model parameters, calculate the product of the local learning rate step size and the partial derivative gradient to obtain the parameter gradient update amount, and subtract the parameter gradient update amount from the starting point of the pre-trained meta-model parameters to obtain the fine-tuned new model parameters; The current electricity market state is inferred by forward propagation using the fine-tuned new model parameters, and the multi-objective optimized weight vector is output.

7. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, After generating the candidate data center set, the process also includes calculating network transmission costs, specifically: Obtain the normalized physical distance parameter, normalized real-time bandwidth utilization, normalized network latency fluctuation parameter, and bandwidth instability parameter of each data center node in the candidate data center set, and obtain the preset first allocation weight coefficient, preset second allocation weight coefficient, preset third allocation weight coefficient, and preset fourth allocation weight coefficient. Calculate the first product of the physical distance parameter and the first allocation weight coefficient, the second product of the real-time bandwidth utilization rate and the second allocation weight coefficient, the third product of the network latency fluctuation parameter and the third allocation weight coefficient, and the fourth product of the bandwidth instability parameter and the fourth allocation weight coefficient; The network transmission cost is obtained by summing the first product, the second product, the third product, and the fourth product.

8. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, The specific steps for generating the scheduling decision to be verified are as follows: Calculate the electricity price economic reward, green power consumption reward, latency satisfaction reward, task value reward, and data security reward for each data center node in the candidate data center set in relation to the computing power task; The comprehensive synergistic effect of each data center node is obtained by weighting and summing the electricity price economic incentive, the green power consumption incentive, the latency satisfaction incentive, the task value incentive, and the data security incentive with the corresponding weight coefficients in the multi-objective optimization weight vector. The Pareto front of the candidate data center set on multiple reward dimensions is calculated based on a multi-objective optimization algorithm, and the data center nodes that meet the non-dominated solution conditions are extracted to form a non-dominated solution set. Select the target data center node with the greatest comprehensive collaborative utility from the set of non-dominated solutions, and generate the scheduling decision to be verified.

9. The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity according to claim 1, characterized in that, The specific steps for generating an effective scheduling instruction after successful verification are as follows: sending the scheduling decision to be verified to the consortium blockchain network for automated verification and consensus determination. The automated verification process specifically involves extracting the quoted electricity price data from the scheduling decision to be verified, obtaining real-time external real electricity price data and a preset minimum price tolerance value, and having the transaction center node in the consortium blockchain network calculate the absolute difference between the quoted electricity price data and the real-time external real electricity price data. When the absolute difference is less than the minimum price tolerance value, the automated verification is deemed to have passed. The consensus nodes in the consortium blockchain network are triggered to execute the consensus determination, and after the consensus determination is passed, the effective scheduling instruction is generated. The specific steps for fine-tuning parameters based on the actual indicator parameters are as follows: based on the actual indicator parameters, extract the multi-dimensional comprehensive deviation feedback signal and the actual comprehensive business utility feedback signal, and combine the current comprehensive environment feature vector, the multi-objective optimization weight vector, the actual comprehensive business utility feedback signal and the comprehensive environment feature vector at the next moment to construct an empirical quadruple; The multidimensional comprehensive deviation feedback signal is used as a supervision signal, and the empirical quadruple is input into the preset meta-learning network to update the model parameters, and the parameter fine-tuning operation is performed.

10. A computer-computer collaborative scheduling system based on multi-dimensional classification of task value and latency sensitivity, characterized in that: The computing-electricity collaborative scheduling method based on multi-dimensional classification of task value and latency sensitivity as described in any one of claims 1-9 includes: The data processing module is used to acquire multi-source input data, including at least computing power task data streams, electricity market data streams, and network topology data streams; The task classification module is used to evaluate the latency sensitivity index and comprehensive value score of the computing power task based on the computing power task data stream, and to perform multi-dimensional classification processing on the computing power task by combining the latency sensitivity index and the comprehensive value score to generate task type code. The meta-learning module is used to extract current market environment features based on the electricity market data stream and dynamically generate a multi-objective optimization weight vector based on the current market environment features. A matching module is used to obtain multiple data center nodes based on the network topology data flow, and remove data center nodes that do not meet the constraints to generate a candidate data center set. An optimization module is used to combine the task type encoding and the multi-objective optimization weight vector to calculate the comprehensive collaborative utility of each data center node in the candidate data center set and generate a scheduling decision to be verified. The audit module is used to automatically verify the scheduling decision to be verified, and generate an effective scheduling instruction after the verification is passed. The feedback module is used to monitor the execution process of the computing power task according to the effective scheduling instruction, collect actual indicator parameters, and perform parameter fine-tuning based on the actual indicator parameters to realize computing power collaborative scheduling.

Images