Computing and power coordination scheduling method, device, equipment, storage medium and program product

Abstract

The application relates to the technical field of artificial intelligence, and provides an algorithm and power collaborative scheduling method and device, equipment, a storage medium and a program product. The method comprises the following steps: obtaining algorithm and power data of a target node; the algorithm and power data comprises algorithm data and power data; performing federated learning training based on the algorithm and power data to obtain an algorithm and power federated model; constructing a multi-agent system based on the algorithm and power data and the algorithm and power federated model; generating an algorithm and power scheduling strategy based on a power agent, an algorithm agent and a resource coordination agent, and performing algorithm and power collaborative scheduling based on the algorithm and power scheduling strategy; wherein the power agent, the algorithm agent and the resource coordination agent communicate through a communication network based on a distributed hash table. In the foregoing manner, the multi-agent can be applied to the scene of algorithm and power resource collaborative scheduling, the collaborative communication between the agents is smooth, the situation that information sharing between the agents is not timely is avoided, timely algorithm and power collaborative scheduling is facilitated, and resource waste is avoided.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a computing-electricity collaborative scheduling method, device, equipment, storage medium, and program product. Background Technology

[0002] In scenarios where computing power and power resources need to be coordinated and scheduled, how to achieve coordinated scheduling of the two has become an urgent problem to be solved.

[0003] In some related technologies, choosing to use multi-agent systems for computing network task scheduling and computing network resource allocation decisions can effectively reduce system latency and achieve joint optimization of load balancing among heterogeneous nodes.

[0004] However, there is currently a lack of methods for collaborative scheduling of computing and power resources using multiple agents. In addition, existing multi-agent systems suffer from poor communication and collaboration among agents, resulting in untimely information sharing and difficulty in timely collaborative scheduling, which may lead to resource waste and affect system stability. Summary of the Invention

[0005] This application provides a computing power collaborative scheduling method, apparatus, device, storage medium, and program product to solve the current lack of a method for collaborative scheduling of computing power resources using multiple agents, and the technical problem that existing multi-agent systems have poor cooperation and communication between agents.

[0006] In a first aspect, embodiments of this application provide a computing-power collaborative scheduling method, comprising: acquiring computing-power data of a target node; the computing-power data including computing power data and power data; performing federated learning training based on the computing-power data to obtain a computing-power federated model; constructing a multi-agent system based on the computing-power data and the computing-power federated model; the multi-agent system including a power agent, a computing power agent, and a resource coordination agent; generating a computing-power scheduling strategy based on the power agent, the computing power agent, and the resource coordination agent, and performing computing-power collaborative scheduling based on the computing-power scheduling strategy; wherein the power agent, the computing power agent, and the resource coordination agent communicate through a communication network based on a distributed hash table.

[0007] In one embodiment, the computing power data of the target node is stored in a blockchain network. Before acquiring the computing power data of the target node, the method further includes: identifying multiple data acquisition nodes; determining the weight of each data acquisition node based on a preset algorithm; conducting consensus election based on the weight of each data acquisition node to select the target node from all data acquisition nodes; and uploading the computing power data of the target node to the blockchain to construct a blockchain network.

[0008] In one embodiment, the preset algorithm is either a Practical Byzantine Fault Tolerance algorithm or a Proof-of-Stake variant algorithm. Based on the preset algorithm, the weight of each data acquisition node is determined, including: if the preset algorithm is a Practical Byzantine Fault Tolerance algorithm, the reputation value of each data acquisition node is determined based on its historical performance and data accuracy, and the weight of each data acquisition node is determined based on its reputation value; if the preset algorithm is a Proof-of-Stake variant algorithm, the weight of each data acquisition node is determined based on its real-time contribution and network load.

[0009] In one embodiment, the number of target nodes is at least two, each target node is a participant in multi-party secure computation, and all target nodes communicate through an unintentional transport protocol. Federated learning training is performed based on computational power data to obtain a federated computational model, including: performing data sample analysis and feature analysis on the computational power data of all target nodes to obtain analysis results; if the analysis results show that the data feature dimensions of each participant are similar but the data samples are different, then horizontal federated learning training is performed based on the computational power data of each participant to obtain a federated computational model; during the horizontal federated learning training process, the computational power data of each participant is encrypted using a homomorphic encryption method; if the analysis results show that the data samples of each participant are similar but the data feature dimensions are different, then vertical federated learning training is performed based on the computational power data of each participant to obtain a federated computational model; during the vertical federated learning training process, the computational power data of each participant is encrypted using a secret sharing method.

[0010] In one embodiment, a multi-agent system is constructed based on power data and a power federation model, including: constructing a power agent based on power data; constructing a computing power agent based on computing power data; generating correlation information between power data and computing power data through the power federation model based on power data and computing power data; and constructing a resource coordination agent based on the correlation information between power data and computing power data.

[0011] In one embodiment, the computing power scheduling strategy includes a power scheduling scheme, a computing power allocation scheme, and a resource allocation strategy. Based on a power intelligent agent, a computing power intelligent agent, and a resource coordination intelligent agent, the computing power scheduling strategy is generated, and computing power collaborative scheduling is performed based on the computing power scheduling strategy. This includes: generating a power scheduling scheme through a power intelligent agent based on real-time power data of the target node; generating a computing power allocation scheme through a computing power intelligent agent based on real-time computing power data of the target node; generating a resource allocation strategy through a resource coordination intelligent agent based on real-time power data and real-time computing power data; and performing computing power collaborative scheduling based on the power scheduling scheme, the computing power allocation scheme, and the resource allocation strategy.

[0012] Secondly, embodiments of this application provide a computing power collaborative scheduling device, comprising: an acquisition module for acquiring computing power data of a target node; the computing power data includes computing power data and power data; a federated learning module for performing federated learning training based on the computing power data to obtain a computing power federated model; a multi-agent construction module for constructing a multi-agent system based on the computing power data and the computing power federated model; the multi-agent system includes a power agent, a computing power agent, and a resource coordination agent; and a computing power collaborative scheduling module for generating a computing power scheduling strategy based on the power agent, computing power agent, and resource coordination agent, and performing computing power collaborative scheduling based on the computing power scheduling strategy; wherein the power agent, computing power agent, and resource coordination agent communicate through a communication network based on a distributed hash table.

[0013] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements any of the above-described computer-computer cooperative scheduling methods.

[0014] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described computer-power cooperative scheduling methods.

[0015] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements any of the above-described computer-computer collaborative scheduling methods.

[0016] The computing power collaborative scheduling method, apparatus, device, storage medium, and program product provided in this application first perform federated learning training based on the computing power data of the target node to obtain a computing power federated model. Then, based on the computing power data and the computing power federated model, a multi-agent system including a power agent, a computing power agent, and a resource coordination agent is constructed. A computing power scheduling strategy is generated based on the power agent, computing power agent, and resource coordination agent, and computing power collaborative scheduling is performed according to the computing power scheduling strategy, thereby applying the multi-agent system to the scenario of computing power resource collaborative scheduling. In addition, the power agent, computing power agent, and resource coordination agent communicate with each other through a communication network based on a distributed hash table. The communication network based on a distributed hash table can achieve decentralized, efficient, and reliable transmission through an efficient routing algorithm, thus ensuring smooth collaborative communication between the agents and effectively avoiding untimely information sharing among the agents. This is conducive to timely computing power collaborative scheduling, avoiding resource waste, and improving system stability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts of the computing and power-coordinated scheduling method provided in the embodiments of this application.

[0019] Figure 2 This is the second flowchart of the computing and power-coordinated scheduling method provided in the embodiments of this application.

[0020] Figure 3 This is a schematic diagram of the blockchain data structure provided in the embodiments of this application.

[0021] Figure 4 This is one of the schematic diagrams illustrating the computation process of the practical Byzantine fault-tolerant algorithm provided in the embodiments of this application.

[0022] Figure 5 This is the second schematic diagram of the calculation process of the practical Byzantine fault-tolerant algorithm provided in the embodiments of this application.

[0023] Figure 6 This is a schematic diagram illustrating the calculation process of the proof-of-rights variant algorithm provided in this application embodiment.

[0024] Figure 7 This is one of the flowcharts illustrating the horizontal federated learning training provided in the embodiments of this application.

[0025] Figure 8 This is the second schematic diagram of the horizontal federated learning training process provided in the embodiments of this application.

[0026] Figure 9 This is one of the flowcharts illustrating the vertical federated learning training provided in the embodiments of this application.

[0027] Figure 10 This is the second schematic diagram of the vertical federated learning training process provided in the embodiments of this application.

[0028] Figure 11 This is a schematic diagram of the design process of the multi-agent system provided in the embodiments of this application.

[0029] Figure 12 This is a schematic diagram of the structure of the multi-agent system provided in the embodiments of this application.

[0030] Figure 13 This is a schematic diagram of the structure of the computing and power coordination scheduling device provided in the embodiments of this application.

[0031] Figure 14 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Please see Figure 1 and Figure 2 , Figure 1 This is one of the flowcharts illustrating the computer-computer collaborative scheduling method provided in the embodiments of this application. Figure 2 This is the second flowchart illustrating the computing-powered collaborative scheduling method provided in this application. Figure 1 As shown in the embodiments of this application, the computing power collaborative scheduling method includes steps S110 to S140, and the specific steps are as follows: S110: Obtain the power calculation data of the target node.

[0034] Computational data includes computing power data and power data.

[0035] In this embodiment, the computational power data of the target node is stored in a blockchain network. Understandably, there are at least two target nodes, and each target node corresponds to a block in the blockchain network; therefore, the blockchain network comprises multiple blocks. For each target node, its computational power data can be stored in the block corresponding to that target node.

[0036] When federated learning training is required, the computing power data of the target node can be obtained directly from the blockchain network.

[0037] S120: Federated learning training is performed based on computing power data to obtain a computing power federated model.

[0038] Specifically, such as Figure 2 As shown, after obtaining computing power data from the blockchain network, the computing power federation model can be obtained by selecting either horizontal federated learning (HFL) training or vertical federated learning (VFL) training based on the characteristics of the computing power data of each target node.

[0039] S130: Construct a multi-agent system based on computing power data and a computing power federation model.

[0040] Multi-Agent Systems (MAS) include power agents, computing agents, and resource coordination agents.

[0041] S140: Based on the power intelligent agent, computing intelligent agent and resource coordination intelligent agent, generate computing power scheduling strategy, and perform computing power collaborative scheduling based on the computing power scheduling strategy.

[0042] Among them, the power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent communicate through a communication network based on a distributed hash table.

[0043] The computing power collaborative scheduling method provided in this application first performs federated learning training based on the computing power data of the target node to obtain a computing power federated model. Then, based on the computing power data and the computing power federated model, a multi-agent system including a power agent, a computing power agent, and a resource coordination agent is constructed. A computing power scheduling strategy is generated based on the power agent, computing power agent, and resource coordination agent, and computing power collaborative scheduling is performed according to the computing power scheduling strategy. Thus, the multi-agent system is applied to the scenario of computing power resource collaborative scheduling. In addition, the power agent, computing power agent, and resource coordination agent communicate with each other through a communication network based on a distributed hash table. The communication network based on a distributed hash table can achieve decentralized, efficient, and reliable transmission through an efficient routing algorithm. Therefore, it can ensure smooth collaborative communication between the agents, effectively avoid the situation of untimely information sharing between the agents, facilitate timely computing power collaborative scheduling, avoid resource waste, and thus improve system stability.

[0044] In some embodiments, the computing power data of the target node is stored in a blockchain network; before acquiring the computing power data of the target node, the method further includes: determining multiple data acquisition nodes; determining the weight of each data acquisition node based on a preset algorithm; performing consensus election based on the weight of each data acquisition node to select the target node from all data acquisition nodes; and uploading the computing power data of the target node to the blockchain to construct a blockchain network.

[0045] Please continue reading. Figure 2 ,like Figure 2 As shown, since the computing power data of the target node needs to be stored in the blockchain network, it is necessary to build a blockchain network and put the data on the chain before using the computing power data of the target node for federated learning training.

[0046] In this embodiment, the computing power collaborative scheduling method is applicable to distributed scenarios, that is, to achieve collaborative scheduling of computing power resources among multiple distributed nodes. Therefore, there are multiple data acquisition nodes, and each data acquisition node has its corresponding computing power data. However, since the data quality of different data acquisition nodes is different, and there may be some malicious or low-quality nodes, the computing power data of these malicious or low-quality nodes may interfere with the subsequent training of the computing power federated model and affect the quality of the computing power federated model. Therefore, before performing federated learning training, it is necessary to select high-quality nodes as target nodes and use the computing power data of the target nodes for federated learning training to ensure the quality and performance of the computing power federated model.

[0047] Please see Figure 3 , Figure 3 This is a schematic diagram of the blockchain data structure provided in the embodiments of this application.

[0048] Specifically, such as Figure 3 As shown, depending on the applicable scenario, the Practical Byzantine Fault Tolerance (PBFT) algorithm or the Proof of Stake (PoS) variant algorithm based on a reputation mechanism can be used to dynamically adjust the weight of each data collection node and define a unique blockchain data structure. In addition to the block header and block body, each block also has an additional data quality identifier field to record the accuracy, completeness and other evaluation information of the corresponding data collection node's historical computing power data and historical electricity data, which facilitates the filtering and verification of data when it is used later.

[0049] Furthermore, based on the weight of each data acquisition node, a consensus election is conducted to select at least two high-quality target nodes from all data acquisition nodes, and the computing power data of all target nodes are uploaded to the blockchain to build a blockchain network.

[0050] Optionally, zero-knowledge proof technology can be used to upload the computing data of all target nodes to the blockchain, proving the authenticity and compliance of the data without disclosing the specific content of the data.

[0051] For example, electricity data can include electricity transaction data. For electricity transaction data of a target node, its legitimacy can be proven to other target nodes through zero-knowledge proof technology, while protecting the privacy information of both parties in the electricity transaction.

[0052] Optionally, the historical power data, historical computing power data, transaction information, etc. of each target node are subjected to multi-layer encryption processing, and a combination of symmetric encryption and asymmetric encryption is used to encrypt the data to ensure data security.

[0053] For example, a symmetric encryption algorithm is first used to quickly encrypt the computing data of the target node, and then an asymmetric encryption algorithm is used to encrypt the symmetric key for transmission, ensuring the security of the data during transmission and storage.

[0054] Optionally, a precise timestamp based on the Blockchain Timestamp Service (BTS) can be added to the computing data of each target node to ensure the temporal order and immutability of the data.

[0055] Optionally, the smart contract functionality of blockchain can be used to automatically perform data verification, storage, and update operations, thereby improving the automation and standardization of data management.

[0056] In some embodiments, the preset algorithm is a Practical Byzantine Fault Tolerance (PBT) algorithm or a Proof-of-Stake (PoS) variant algorithm. Determining the weight of each data acquisition node based on the preset algorithm includes: if the preset algorithm is a PBT algorithm, determining the reputation value of each data acquisition node based on its historical performance and data accuracy, and determining the weight of each data acquisition node based on its reputation value; if the preset algorithm is a PoS variant algorithm, determining the weight of each data acquisition node based on its real-time contribution and network load.

[0057] Optionally, for scenarios with small node size or low energy consumption sensitivity, the preset algorithm can be the Practical Byzantine Fault Tolerance (PBFT) algorithm based on a reputation mechanism.

[0058] Specifically, in the practical Byzantine fault-tolerant algorithm based on the reputation mechanism, a node reputation evaluation system is introduced. The reputation value of each data acquisition node is evaluated based on factors such as the historical performance and data accuracy of each data acquisition node, and the weight of each data acquisition node is determined based on the reputation value of each data acquisition node.

[0059] Among them, data collection nodes with high reputation values ​​have higher weight in the consensus election process. Data collection nodes with higher weights are more likely to be elected as target nodes in the consensus election process, which can effectively reduce the interference of malicious or low-quality nodes on the consensus and improve the efficiency and reliability of the consensus.

[0060] Please see Figure 4 and Figure 5 , Figure 4 This is one of the schematic diagrams illustrating the computation process of the practical Byzantine fault-tolerant algorithm provided in the embodiments of this application. Figure 5 This is the second schematic diagram of the calculation process of the practical Byzantine fault-tolerant algorithm provided in the embodiments of this application.

[0061] like Figure 4 and Figure 5 As shown, the Practical Byzantine Fault Tolerance Algorithm is an algorithm based on state machine replica replication. Each replica of the state machine (i.e., each data acquisition node) saves its service state and also implements the operation of all legitimate client requests. It can guarantee that, while satisfying the liveness and security of the distributed system, it allows... One data acquisition node malfunctioned (e.g., data loss at the data acquisition node, data acquisition node not working, etc.), among which This represents the number of data collection nodes participating in the consensus election process in a distributed system. Practical Byzantine fault-tolerant algorithms can guarantee that the distributed system... Even if a data collection node malfunctions or is maliciously manipulated, it can still correctly achieve distributed consensus.

[0062] It's important to note that in the practical Byzantine fault tolerance algorithm, the nodes storing replicas are all part of a view's rotation process. For example, in view number v, one replica node is the primary node, and the other replica nodes are backup nodes; the primary node is mainly used to receive request messages sent by clients, and its function is determined by the formula... Calculate and select, This indicates the number of storage replica nodes. If the primary node fails, a view replacement is initiated, the current view number v is changed, and then a new primary node is selected from all data acquisition nodes using the formula described above.

[0063] Optionally, for scenarios with a large number of nodes and high energy consumption sensitivity, the preset algorithm can use a variant of the Proof of Stake (PoS) algorithm.

[0064] Specifically, in the Proof-of-Stake variant algorithm, the weight of each data collection node can be dynamically adjusted based on its real-time contribution and network load to ensure the fairness and efficiency of the blockchain network.

[0065] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating the calculation process of the proof-of-rights variant algorithm provided in this application embodiment.

[0066] The Proof-of-Stake (PoS) variant primarily balances safety and liveness. Safety refers to the blockchain's ability to avoid forks or ensure a very low probability of forks, thus safely achieving finality (a state where blocks cannot be changed or rolled back). Liveness refers to the blockchain network's ability to process data from data collection nodes and package it into blocks in a timely manner, propelling the blockchain forward. A Verifiable Random Function (VRF) generates verifiable pseudo-random numbers. Given a data collection node's private key... Random input Running VRF will produce a fixed output. , These are generated random numbers. It is proof. and One-to-one correspondence, and irreversible (i.e., satisfying the condition) VRF can guarantee It is uniformly distributed and unpredictable; The public key of the corresponding dataset node is ,pass The randomness can be verified. Is it by the data acquisition node through input? Generated. Therefore, this design has two advantages: first, it can verify the identity of the random number producer; second, it can verify whether the random number matches the random input. Random number The value of the random number determines whether the data collection node is eligible to participate in this round of federated learning training. In If so, then in that round, that collection node can participate in model training.

[0067] Introducing the VRF method into a blockchain network, its consensus mechanism is called PpoS (Pure Proof-of-Stake). This consensus mechanism has the following characteristics: the confirmation process of data collection nodes is random; any data collection node only needs to hold historical electricity data and historical computing power data to have the opportunity to use it for model training, possessing complete openness and freedom. Figure 6 As shown, this consensus election process can be divided into the following four steps: (1) Block proposal phase: All data collection nodes run their own VRF. If a data collection node is selected as training data, it outputs its corresponding proposed block and VRF.

[0068] (2) Screening and Voting Phase: All data collection nodes run their own VRF. If a data collection node is selected, it can participate in the voting. There are usually multiple blocks from eligible data collection nodes. At this time, the block with the smallest VRF hash value can be selected. If a block reaches the quorum for consensus election, then the block can be used as the sole confirmed block.

[0069] (3) Confirmation voting stage: The block selected in the screening voting stage is finally verified. If more than the number of data collection nodes vote in favor, the block is submitted to the ledger and the corresponding data collection node becomes the target node.

[0070] (4) Add the computing data of the target node to the existing blockchain network for federated learning training.

[0071] The computing and power collaborative scheduling method provided in this application transforms the data structure of traditional blockchain networks. It reconstructs the blockchain network using the PBFT algorithm and a variant of the PoS algorithm, and dynamically adjusts the weight of data acquisition nodes based on their historical performance. This effectively reduces the interference of malicious or low-quality nodes on subsequent model construction, ensuring the fairness and efficiency of the blockchain network.

[0072] In some embodiments, the number of target nodes is at least two, each target node is a participant in multi-party secure computation, and all target nodes communicate through an unintentional transport protocol; federated learning training is performed based on computational power data to obtain a computational power federated model, including: performing data sample analysis and feature analysis on the computational power data of all target nodes to obtain analysis results; if the analysis results show that the data feature dimensions of each participant are similar but the data samples are different, then horizontal federated learning training is performed based on the computational power data of each participant to obtain a computational power federated model; during the horizontal federated learning training process, the computational power data of each participant is encrypted using a homomorphic encryption method; if the analysis results show that the data samples of each participant are similar but the data feature dimensions are different, then vertical federated learning training is performed based on the computational power data of each participant to obtain a computational power federated model; during the vertical federated learning training process, the computational power data of each participant is encrypted using a secret sharing method.

[0073] Please continue reading. Figure 2 ,like Figure 2 As shown, after obtaining the computing power data of the target node, a federated learning framework needs to be built to train the computing power federated model. Federated learning is a machine learning algorithm framework, whose main features include: (1) Multi-party participation: There are two or more participants who collaborate to build a shared federated learning model. Each participant has a portion of the data used for model training.

[0074] (2) No data exchange: During the federated learning training process, no original data of any participant will leave the participant, that is, the data will not be directly exchanged and collected, which can guarantee the security and privacy of the data.

[0075] (3) Protecting transmitted information: During the federated learning training process, the data necessary for training must be protected before it can be transmitted between the participants, so that the participants cannot infer the data of other participants based on the transmitted information.

[0076] (4) Approximately lossless: The performance of the federated learning model should be close to that of the ideal model (i.e., the model trained by each participant by directly merging data).

[0077] In this embodiment, the federated learning training process innovatively combines Secure Multi-Party Computation (MPC) technology with Oblivious Transfer (OT) protocol to achieve joint computation of computing power data and power data in an encrypted state, thereby enabling efficient data interaction among participants while ensuring data privacy. Each target node is a participant in the Secure Multi-Party Computation, and all target nodes can communicate through the Oblivious Transfer protocol.

[0078] Based on this, data sample analysis and feature analysis are performed on the power calculation data of all target nodes.

[0079] If the data feature dimensions of each participant are similar but the data samples are different, then horizontal federated learning training is performed based on the computing power data of each participant to obtain a computing power federated model. During the horizontal federated learning training process, homomorphic encryption technology is used to encrypt the computing power data of each participant, so that each participant can train the model on ciphertext and avoid training data leakage caused by plaintext data transmission.

[0080] Among them, data feature dimension similarity refers to the overlap of sample features of different data samples being greater than a preset threshold.

[0081] Please see Figure 7 and Figure 8 , Figure 7 This is one of the flowcharts illustrating the horizontal federated learning training provided in this application embodiment. Figure 8 This is the second schematic diagram of the horizontal federated learning training process provided in the embodiments of this application.

[0082] like Figure 7 and Figure 8As shown, the essence of horizontal federated learning training is the joint use of samples. It is suitable for scenarios where data collection nodes have similar or identical business types, but different samples (i.e., many overlapping features but few overlapping samples). In this case, by splitting the sample data of each participant according to the horizontal sample dimension, the data with the same but not completely identical sample features can be extracted for training. The main characteristics of this data distribution are: consistent feature space and different sample IDs. Horizontal federated learning training creates a more comprehensive and accurate federated computing model by aggregating different feature information from different data collection nodes, without directly sharing the original data.

[0083] If the data samples of each participant are similar but the data feature dimensions are different, then based on the computing power data of each participant, vertical federated learning training is carried out to obtain the computing power federated model. During the vertical federated learning training process, the computing power data of each participant is encrypted through a secret sharing method, and the data is divided into multiple shares for collaborative computing among different participants.

[0084] Among them, data sample similarity means that the number of identical data samples among all data samples of each participant is greater than a preset threshold.

[0085] Please see Figure 9 and Figure 10 , Figure 9 This is one of the schematic diagrams of the vertical federated learning training process provided in the embodiments of this application. Figure 10 This is the second schematic diagram of the vertical federated learning training process provided in the embodiments of this application.

[0086] like Figure 9 and Figure 10 As shown, vertical federated learning training is another form of federated learning, suitable for situations where different participants have the same samples but different feature spaces. Each participant's dataset overlaps in the sample dimension but differs in the feature dimension. The main purpose of vertical federated learning training is to build a more powerful model by combining these complementary features while ensuring data privacy. Its data distribution characteristics are: identical sample IDs but different feature dimensions. By aggregating different feature information from different participants, vertical federated learning training can create a more comprehensive and accurate federated computing model without directly sharing the original data.

[0087] Optionally, a blockchain-based federated learning network can be constructed, leveraging the immutability and traceability of blockchain to record parameter updates and participant contributions during model training, thereby incentivizing active participation from all parties in model training.

[0088] In the federated learning training process, participants can utilize cutting-edge deep learning models, such as time series forecasting models based on the Transformer architecture, to deeply mine local data. For example, power suppliers use multi-source information such as power generation, equipment status, and meteorological data as input, employing attention mechanisms to capture the impact of different factors on power supply, and using power supply forecasts as output for model training. Computing service providers use computing power demand, resource utilization, and business type as input, extracting features through convolutional neural networks (CNNs), and using computing power task allocation optimization as output for model training. In the parameter aggregation phase, an adaptive aggregation strategy is adopted, dynamically adjusting aggregation weights based on the quality and stability of each participant's data. For example, participants with high-quality data and low volatility are assigned higher aggregation weights to ensure the accuracy and stability of the global model. Through multiple iterations of training, the model performance is continuously optimized, enabling it to accurately capture the complex patterns and potential correlations in power and computing power data.

[0089] The computing-power collaborative scheduling method provided in this application innovatively combines MPC technology and OT protocol in terms of model training. Under the premise of ensuring the privacy of computing power and power data, it reasonably uses HFL or VFL architecture and encryption technology according to the feature dimensions and samples of the data, so that computing power data can be trained efficiently on encrypted data, so that the trained computing-power federated model can accurately capture the complex patterns and potential correlations between power data and computing power data.

[0090] In some embodiments, a multi-agent system is constructed based on power data and a power federation model, including: constructing a power agent based on power data; constructing a computing power agent based on computing power data; generating correlation information between power data and computing power data through a power federation model based on power data and computing power data; and constructing a resource coordination agent based on the correlation information between power data and computing power data.

[0091] Please see Figure 11 , Figure 11 This is a schematic diagram of the design process of the multi-agent system provided in the embodiments of this application.

[0092] like Figure 11 As shown, in this embodiment, the design process of a multi-agent system mainly includes the following steps: (1) Definition of intelligent agent architecture: Specifically, all intelligent agents in this implementation adopt a hierarchical decision-making architecture, which includes three core modules: perception module, decision-making module, and execution module.

[0093] The sensing module is responsible for collecting power data (such as voltage, frequency, and load) of the power system in each node and computing power data (such as task queues and computing resource utilization) of the computing power system in each node.

[0094] The decision module is used to generate corresponding computing and power scheduling strategies based on deep reinforcement learning (DRL) and game theory algorithms.

[0095] The execution module is used to convert the decision results into specific computing power scheduling instructions and power control signals.

[0096] (2) Vector definition and function design: Specifically, the state space S is defined to contain the power data of the power system (used to characterize the state of the power system), such as grid power. ,Voltage ,frequency The state space S also contains computing power data of the computing system (used to characterize the state of the computing system), such as CPU utilization. GPU utilization Task queue length The state space S also includes environmental states, such as renewable energy output forecasts. Real-time electricity price wait.

[0097] Based on this, the state vector of state space S It can be represented as: .

[0098] Specifically, the action space A is defined to include computing power scheduling and power control. Computing power scheduling includes task migration, resource allocation, task priority adjustment, etc., while power control includes load reduction, energy storage charging and discharging, demand response, etc.

[0099] Based on this, the action vectors of action space A It can be represented as: in, Represents load power. Represents storage power.

[0100] Specifically, taking into account power system stability, computing efficiency, and economic benefits, a reward function R is defined, which can be expressed as: ; Among them, power system stability reward It can be represented as: ; Computing efficiency reward It can be represented as: ; Economic benefit reward It can be represented as: ; in, , , , , , , , , and These are all weight coefficients, which can be dynamically adjusted through reinforcement learning; This represents the transmission cost.

[0101] (3) Agent training and optimization: Historical data from power systems and computing systems are normalized, and training samples are generated using a sliding window method.

[0102] The formula for normalization can be expressed as: .

[0103] Optionally, the agent can be trained using the PPO (Proximal Policy Optimization) algorithm, the process of which can be represented as follows: ; in, It is the strategy update ratio. It is the estimate of the dominance function. It's a clipping range hyperparameter. The objective function is denoted as .

[0104] Optionally, the advantage function estimate can be calculated using the GAE (Generalized Advantage Estimation) algorithm: ; in, It is TD error. It is a GAE hyperparameter.

[0105] Optionally, a multi-agent game model can be constructed by optimizing the strategy using dynamic game theory: ; Furthermore, the Nash-Q learning algorithm is used to solve for the equilibrium strategy of the multi-agent game model: .

[0106] (4) Agent execution: Specifically, state observation is performed on the constructed computing-related intelligent agents (i.e., power intelligent agents, computing power intelligent agents, and resource coordination intelligent agents), real-time state data of the power system and computing power system are collected, the data is preprocessed and feature extracted, and the input state vector is used. The trained DRL model outputs action vectors. According to the action vector Adjust the allocation of computing resources and the execution results of power load monitoring and feedback to the corresponding agents to calculate reward values. Update the agent's policy and value function.

[0107] (5) Architecture performance evaluation: Specifically, the stability of the power system of the constructed multi-agent system, such as frequency deviation, is evaluated. Voltage deviation Evaluate the computational efficiency of the constructed multi-agent system, such as resource utilization. Task completion rate Evaluate the economic benefits of the constructed multi-agent system, such as total cost. The system compares the performance of the agent with traditional scheduling methods (such as round-robin scheduling and static priority scheduling) to verify the agent's adaptability and robustness in dynamic environments.

[0108] Understandably, in designing different intelligent agents, it is also necessary to consider the functional differences between the agents and design different functional modules according to different needs.

[0109] After understanding the design process of multi-agent systems, power agents can be constructed based on power data.

[0110] Specifically, the power intelligent agent should have the ability to perceive the entire state of the power system in real time. It should not only monitor conventional data such as power generation, transmission line load, and power inventory, but also collect micro-state data such as vibration, temperature, and humidity of power generation equipment in real time through IoT sensors. It should use fault prediction models to predict equipment failure risks in advance. At the initialization stage, the power intelligent agent should set an initial power generation plan and power allocation strategy based on historical power data and prediction models, and set up an emergency response plan to deal with sudden power events.

[0111] Therefore, the design goal of the power intelligent agent is to monitor power generation and consumption in real time; predict power system faults and optimize power dispatch (the accuracy of data prediction can be improved by Kalman filtering and LSTM models, or fault prediction can be performed by SVM to avoid risks in advance); ensure the stable operation of the power system, optimize power dispatch, and reduce power generation costs.

[0112] Based on this, the core functional modules of the power intelligent agent include: (1) Power generation monitoring module: This module is used to collect generator output data in real time to ensure that power generation matches demand and avoids overload or underload. It also uses Kalman filtering to denoise the data and improve data accuracy. Its workflow can be represented as follows: ; in, This is the Kalman gain, used for dynamic adjustment to optimize prediction accuracy.

[0113] (2) Transmission line load monitoring module: This module is used to monitor the load of transmission lines in real time, predict future load trends, avoid line overload, and monitor the line load rate. , It can be represented as: ; in, For actual load, This represents the maximum carrying capacity of the line.

[0114] Optionally, an LSTM model can be used to perform time series analysis on the load data to predict future load trends, and a load rate threshold can be preset. An alarm is triggered when the load rate exceeds the threshold. The prediction process of the LSTM model can be represented as follows: .

[0115] (3) Fault prediction module: This module is used to predict possible faults in the power system based on historical power data and real-time power data.

[0116] Optionally, fault classification is performed based on Support Vector Machine (SVM), and a fault prediction model is generated by combining fault feature vectors. The fault prediction model can be represented as: ; in, For kernel function, Lagrange multipliers (4) Power Dispatch Module: This module is used to optimize generator output based on power generation monitoring and load monitoring results, ensure stable operation of the power system, construct an optimization model, minimize power generation costs and output adjustment penalties, and use linear programming or dynamic programming to solve for the optimal power dispatch scheme. This process can be expressed as: ; in, For the cost of electricity generation, Adjust the penalty items to compensate for the effort.

[0117] Furthermore, based on computing power data, a computing power intelligent agent is constructed.

[0118] The computing power agent needs to comprehensively collect multi-dimensional information about computing power tasks, including not only task type, priority, and estimated execution time, but also real-time progress, dynamic resource consumption, and other computing power data. During initialization, the computing power agent needs to formulate an initial computing power allocation plan based on historical computing power demand data and resource allocation experience, and establish a dynamic task priority adjustment mechanism to adapt to constantly changing computing power demand scenarios.

[0119] Therefore, the design goal of the computing power agent is to dynamically allocate computing resources to meet task requirements; use a PID controller to track task progress in real time and optimize task scheduling; and improve the utilization rate of computing resources.

[0120] Based on this, the core functional modules of the computing power intelligent agent include: (1) Task classification module: This module is used to classify tasks according to their type (computation-intensive, I / O-intensive, etc.), extract task feature vectors, generate classification models, and provide a basis for subsequent resource allocation. Its workflow can be represented as follows: ; in, As a task category, This represents the task feature vector.

[0121] (2) Priority allocation module: This module is used to allocate priorities based on task type. User level and deadline Factors such as these determine the priority of tasks. It uses a weighted priority algorithm to dynamically adjust the weight coefficients to adapt to different scenario requirements.

[0122] The priority calculation formula is as follows: ; in, , , All are weighting coefficients.

[0123] (3) Computing power allocation module: This module is used to dynamically allocate computing power resources according to task priority and computing power resource status, construct dynamic programming model, and optimize computing power allocation scheme.

[0124] Optionally, a greedy algorithm or a genetic algorithm can be used to find the optimal computing power allocation scheme. The process of finding the optimal computing power allocation scheme can be represented as follows: ; in, Let be the utility function of the task. To allocate computing power, Total computing power.

[0125] (4) Task progress tracking module: This module is used to monitor the task progress in real time, dynamically adjust the computing power allocation, ensure that the task is completed on time, use the PID controller to adjust the computing power allocation, ensure that the task progress meets expectations, calculate the task completion rate in real time, and generate a progress report.

[0126] The expression for the progress report is as follows: .

[0127] The process of adjusting computing power allocation using a PID controller can be represented as follows: ; .

[0128] Furthermore, based on power data and computing power data, a federated computing model is used to generate correlation information between power data and computing power data, and a resource coordination intelligent agent is constructed based on the correlation information between power data and computing power data.

[0129] The resource coordination agent, acting as a bridge between the power agent and the computing agent, can utilize a high-efficiency communication network based on a distributed hash table (DHT) to achieve fast and reliable message transmission between agents, possessing deep data analysis and intelligent decision-making capabilities. During initialization, the resource coordination agent can load a federated computing model trained on historical data, enabling it to quickly and accurately analyze the correlation between power and computing data, generate correlation information between power and computing data, and formulate resource allocation strategies.

[0130] Optionally, a communication fault tolerance mechanism can be established among multiple agents to ensure that basic coordination functions can still be maintained in the event of communication anomalies.

[0131] Specifically, the design goal of the resource coordination agent is to globally optimize the allocation of power resources and computing resources, maximize the utilization efficiency of power resources, and meet computing power requirements; realize the correlation analysis of power and computing data; achieve resource collaborative optimization through the correlation analysis of power and computing data; use the NSGA-II algorithm to solve multi-objective optimization problems and balance power costs and computing power requirements; and at the same time, ensure efficient and reliable information transmission through a DHT-based communication network.

[0132] Based on this, the core functional modules of the resource coordination intelligent agent include: (1) Power data correlation analysis module: This module is used to analyze the correlation between power data and computing power data, providing a basis for resource scheduling. It uses Pearson correlation coefficient analysis to generate correlation information between power data and computing power data, and combines the correlation analysis results to provide data support for generating resource allocation strategies.

[0133] The correlation information (i.e., the relationship) between power data and computing power data can be represented as follows: ; in, For power data, This is computing power data.

[0134] (2) Resource Allocation Strategy Formulation Module: This module is used to formulate a global resource allocation strategy based on the results of the correlation analysis of computing data, construct a multi-objective optimization model, balance power costs and computing power requirements, and use the NSGA-II algorithm to solve for the Pareto optimal solution, thereby obtaining the optimal resource allocation strategy. Its optimization objective can be expressed as: ; in, For electricity costs, This refers to the utilization rate of computing power.

[0135] (3) Global Scheduling Module: This module coordinates the operation of power agents and computing agents according to resource allocation strategies. It uses a DHT-based communication network to obtain the state information of power agents and computing agents, and uses a distributed consensus algorithm (such as Raft) to ensure the consistency of scheduling decisions. Its distributed consensus process can be represented as: .

[0136] (4) Dynamic adjustment module: This module is used to dynamically adjust the resource allocation strategy based on real-time data to adapt to environmental changes.

[0137] Optionally, a reinforcement learning algorithm can be used to dynamically adjust the resource allocation strategy, and an adjustment plan can be generated by combining real-time data. The process can be represented as follows: ; in, This is the adjustment function based on reinforcement learning.

[0138] It should be noted that, in this embodiment, the DHT-based communication network must meet the following three core functional requirements: (1) Information storage and retrieval: Consistent hashing is used to distribute data to a DHT-based communication network to achieve efficient information storage and retrieval.

[0139] Specifically, a hash function is used to map data to DHT network nodes, and redundant storage and replication mechanisms are set up to ensure data reliability, i.e., satisfying the following formula: ; in, This represents the number of nodes.

[0140] (2) Message passing: Use the Gossip protocol for efficient information transmission to ensure timely information delivery. Set the transmission probability and transmission rounds to control the scope of information transmission. Combine message priority to optimize the information transmission order, i.e., satisfy the following formula: ; in, For the probability of propagation, For the propagation rounds.

[0141] (3) Fault tolerance mechanism: Use redundant storage and replication mechanisms to ensure data reliability and prevent data loss. For example, set the number of replicas to ensure that data is stored on multiple nodes, and use a heartbeat mechanism to detect node status and handle faulty nodes in a timely manner, that is, satisfy the following formula: ; in, For the number of copies, This represents the number of nodes.

[0142] Optionally, the multi-agent system can be optimized periodically. For example, the decision-making strategies of the computing federated model and the multi-agent system can be optimized periodically. Transfer learning techniques can be used to transfer the parameters of excellent models in other similar scenarios to the current model, thereby accelerating model convergence and improving the model's generalization ability.

[0143] In some embodiments, the computing power scheduling strategy includes a power scheduling scheme, a computing power allocation scheme, and a resource allocation strategy. Based on a power intelligent agent, a computing power intelligent agent, and a resource coordination intelligent agent, a computing power scheduling strategy is generated, and computing power collaborative scheduling is performed based on the computing power scheduling strategy. This includes: generating a power scheduling scheme through a power intelligent agent based on real-time power data of the target node; generating a computing power allocation scheme through a computing power intelligent agent based on real-time computing power data of the target node; generating a resource allocation strategy through a resource coordination intelligent agent based on real-time power data and real-time computing power data; and performing computing power collaborative scheduling based on the power scheduling scheme, the computing power allocation scheme, and the resource allocation strategy.

[0144] Specifically, the power dispatch module of the power intelligent agent can optimize generator output based on real-time power data such as power generation monitoring and load monitoring results, ensure the stable operation of the power system, and build an optimization model to minimize power generation costs and output adjustment penalties. The optimal power dispatch scheme can be obtained by using linear programming or dynamic programming.

[0145] Furthermore, the computing power allocation module of the computing power agent can dynamically allocate computing power resources based on real-time computing power data such as task priority and computing power resource status, construct a dynamic programming model, and optimize to obtain the optimal computing power allocation scheme.

[0146] Furthermore, the resource allocation strategy formulation module of the resource coordination agent can formulate a global resource allocation strategy based on real-time power data, real-time computing power data, and the correlation analysis results of computing power data, construct a multi-objective optimization model, balance power costs and computing power requirements, and use the NSGA-II algorithm to solve for the Pareto optimal solution, thereby obtaining the optimal resource allocation strategy.

[0147] Furthermore, based on power dispatching schemes, computing power allocation schemes, and resource allocation strategies, collaborative scheduling of computing and power is carried out.

[0148] For ease of understanding, this section combines... Figure 12 This section describes the computational and power-coordinated scheduling process of the multi-agent system in this embodiment. Please refer to [link / reference]. Figure 12 , Figure 12 This is a schematic diagram of the structure of the multi-agent system provided in the embodiments of this application.

[0149] like Figure 12 As shown, the data support layer of the multi-agent system includes a federated learning model (i.e., the power federated model) and a blockchain network. The power federated model uses MPC technology, Transformer, and other algorithms, and through continuous training, it provides data support for power agents (i.e.,...) Figure 12 The power supply intelligent agent and the computing power intelligent agent (i.e.) Figure 12 It provides accurate data prediction and decision-making basis for the computing power demand agent and the resource coordination agent.

[0150] For example, the federated computing model can provide power supply forecasts for power agents to help them formulate power generation plans; and provide optimization schemes for the allocation of computing power tasks for computing power agents.

[0151] Blockchain networks are used to store historical data and transaction information. They employ algorithms such as PBFT or POS to provide different intelligent agents with immutable historical data and transaction information, ensuring that the agents have a reliable data foundation for decision-making.

[0152] For example, when analyzing the correlation between electricity data and computing power data, resource coordination agents can refer to historical data on the blockchain network to make decisions more scientific.

[0153] The power intelligent agent includes a power generation monitoring module, a transmission line load monitoring module, and a fault prediction module; the computing power intelligent agent includes a task classification module, a priority allocation module, a computing power allocation module, and a task progress tracking module; the resource coordination intelligent agent includes a computing power data correlation analysis module, a resource allocation strategy formulation module, a global scheduling module, and a dynamic adjustment module, and is responsible for communicating with other intelligent agents and handling communication anomalies.

[0154] The communication lines between the various agents are all based on a distributed hash table (DHT) communication network. When exchanging messages, the message transmission from the power agent to the computing power agent may include power supply shortage messages (such as region, degree, time, etc.), and the messages from the computing power agent back to the power agent may include computing power adjustment plan feedback. Similar messages are also used to transmit between the resource coordination agent and the other two agents, such as "power and computing power correlation analysis results" and "coordination strategy instructions".

[0155] For example, when a power intelligence agent detects a power shortage in a certain area (e.g., area A), it sends a power shortage message to a computing intelligence agent. Assume the area with the power shortage is A, and the power shortage is [missing information]. , Monitoring formulas within the power intelligent body The calculation shows that, This is the real-time electricity demand for region A. This refers to the available power supply in region A. Upon receiving the message, the computing power agent first analyzes the data to determine the areas with power shortages and their expected duration. Then, it evaluates the computing power tasks within that region, assuming the set of computing power tasks in region A is [value missing]. The set of adjustable low-priority tasks is selected by using a task priority scoring function. The computing power agent filters adjustable low-priority tasks from the task priority queue. For example, it discovers that some data backup tasks (low priority) are running in region A, and these tasks do not have high real-time requirements. Assume the computing power requirement of the data backup tasks is... Its computing resources are The computing agent can then decide to suspend the execution of these low-priority tasks and reallocate the computing resources occupied by these tasks to other regions with sufficient power supply and computing demand. For example, the computing resources of the suspended data backup task in region A can be allocated to the ongoing artificial intelligence training task in region B to improve the overall system's resource utilization efficiency and alleviate the power supply pressure on region A. Assume the increase in computing demand for the artificial intelligence training task in region B is... ,and The computing power agent will then transmit the adjusted computing power allocation plan to the relevant computing power nodes via messages, and each computing power node will execute tasks according to the new allocation plan. During this process, the power agent and the computing power agent maintain the stable operation of the system through continuous message interaction and cooperation. After each message interaction, both the power agent and the computing power agent can update their own strategies based on the reward function. For example, if the reward value is positive after a certain collaboration, the power agent is more likely to adopt similar power generation adjustment or power allocation strategies in subsequent decisions, and the computing power agent is also more likely to adopt similar computing power allocation adjustment strategies.

[0156] Optionally, the agent can continuously update its strategy based on the feedback from reinforcement learning. Through interaction with the environment, the agent can obtain reward values ​​and adjust its strategy based on the reward values ​​to maximize long-term cumulative rewards.

[0157] Optionally, the reward value can be updated using a Q-learning algorithm, whereby the agent selects the optimal action based on the current state s and action a. The formula for updating the Q value is: ; in, It's the learning rate. This is the current reward value. It is a discount factor. This is the next state.

[0158] Optionally, the federated computing model can be retrained periodically, incorporating new power and computing power data into the training and updating model parameters to adapt to dynamic changes in the system. Simultaneously, the agent's decisions can be analyzed and optimized based on historical data and transaction information on the blockchain network; for example, by analyzing historical power supply anomalies, the agent's strategies for dealing with such situations can be adjusted.

[0159] Optionally, a system monitoring mechanism can be established to monitor indicators such as power supply, computing power demand, and system stability in real time.

[0160] For example, power supply thresholds and computing power demand thresholds can be set. When the power supply is below the threshold or the computing power demand is above the threshold, an alarm can be issued, and the system can be optimized based on the monitoring results.

[0161] For example, when it is found that a certain agent's decisions frequently lead to system instability, its reward function or parameter settings can be adjusted to guide it to make more reasonable decisions. At the same time, the communication mechanisms and cooperation strategies of the agents can be continuously improved to enhance the overall performance and stability of the system.

[0162] The computing and power collaborative scheduling method provided in this application relies on federated learning models and blockchain data to construct a multi-agent system based on reinforcement learning, including a power agent, a computing agent, and a resource coordination agent. It utilizes a DHT-based communication network for collaborative communication and dynamically adjusts the computing and power scheduling strategy according to real-time data, thereby ensuring system stability.

[0163] This application also provides a computing power collaborative scheduling device. Please refer to [link to relevant documentation]. Figure 13 , Figure 13 This is a schematic diagram of the structure of the computing and power-coordinated scheduling device provided in an embodiment of this application. In this embodiment, the computing and power-coordinated scheduling device includes an acquisition module 1310, a federated learning module 1320, a multi-agent construction module 1330, and a computing and power-coordinated scheduling module 1340.

[0164] The acquisition module 1310 is used to acquire the power calculation data of the target node.

[0165] Computational data includes computing power data and power data.

[0166] Federated learning module 1320 is used to perform federated learning training based on computing data to obtain a computing federated model.

[0167] The multi-agent building module 1330 is used to build multi-agent systems based on computing data and computing federation models.

[0168] Multi-agent systems include power agents, computing agents, and resource coordination agents.

[0169] The computing power collaborative scheduling module 1340 is used to generate computing power scheduling strategies based on power intelligent agents, computing power intelligent agents and resource coordination intelligent agents, and to perform computing power collaborative scheduling based on computing power scheduling strategies.

[0170] Among them, the power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent communicate through a communication network based on a distributed hash table.

[0171] In some embodiments, the computing power data of the target node is stored in a blockchain network.

[0172] The acquisition module 1310 is used to determine multiple data acquisition nodes; determine the weight of each data acquisition node based on a preset algorithm; conduct consensus election based on the weight of each data acquisition node to select a target node from all data acquisition nodes; and upload the computing power data of the target node to the blockchain to build a blockchain network.

[0173] In some embodiments, the default algorithm is a practical Byzantine fault-tolerant algorithm or a proof-of-stake variant algorithm.

[0174] The acquisition module 1310 is used to determine the reputation value of each data acquisition node based on its historical performance and data accuracy, and to determine the weight of each data acquisition node based on its reputation value, if the preset algorithm is the Practical Byzantine Fault Tolerance Algorithm; if the preset algorithm is the Proof-of-Stake Variant Algorithm, to determine the weight of each data acquisition node based on its real-time contribution and network load, and to obtain the weight of each data acquisition node based on the Proof-of-Stake Variant Algorithm.

[0175] In some embodiments, the number of target nodes is at least two, each target node is a participant in a multi-party secure computation, and all target nodes communicate via an unintentional transport protocol.

[0176] The Federated Learning Module 1320 is used to perform data sample analysis and feature analysis on the computing power data of all target nodes to obtain analysis results. If the analysis results show that the data feature dimensions of each participant are similar but the data samples are different, then horizontal federated learning training is performed based on the computing power data of each participant to obtain a computing power federated model. During the horizontal federated learning training process, the computing power data of each participant is encrypted using a homomorphic encryption method. If the analysis results show that the data samples of each participant are similar but the data feature dimensions are different, then vertical federated learning training is performed based on the computing power data of each participant to obtain a computing power federated model. During the vertical federated learning training process, the computing power data of each participant is encrypted using a secret sharing method.

[0177] In some embodiments, the multi-agent construction module 1330 is used to construct a power agent based on power data; construct a computing power agent based on computing power data; generate correlation information between power data and computing power data through a computing-power federation model based on power data and computing power data; and construct a resource coordination agent based on the correlation information between power data and computing power data.

[0178] In some embodiments, the computing power scheduling strategy includes a power scheduling scheme, a computing power allocation scheme, and a resource allocation strategy.

[0179] The computing and power collaborative scheduling module 1340 is used to generate a power scheduling scheme through a power intelligent agent based on the real-time power data of the target node; generate a computing power allocation scheme through a computing power intelligent agent based on the real-time computing power data of the target node; generate a resource allocation strategy through a resource coordination intelligent agent based on the real-time power data and real-time computing power data; and perform computing and power collaborative scheduling based on the power scheduling scheme, computing power allocation scheme, and resource allocation strategy.

[0180] This application also provides an electronic device. Figure 14 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 14As shown, the electronic device may include a processor 1410, a communications interface 1420, a memory 1430, and a communication bus 1440. The processor 1410, communications interface 1420, and memory 1430 communicate with each other via the communication bus 1440. The processor 1410 can call logical instructions from the memory 1430 to execute a computing-electrical cooperative scheduling method.

[0181] Furthermore, the logical instructions in the aforementioned memory 1430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the computer-power cooperative scheduling method provided by the above methods.

[0183] This application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the computer-computer cooperative scheduling method provided by the above methods.

[0184] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0185] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A computer-based collaborative scheduling method, characterized in that, include: Acquire the computing power data of the target node; the computing power data includes computing power data and power data; Federated learning training is performed based on the aforementioned computing power data to obtain a computing power federated model; Based on the power computing data and the power computing federation model, a multi-agent system is constructed; the multi-agent system includes a power agent, a computing power agent, and a resource coordination agent. Based on the power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent, a computing power scheduling strategy is generated, and computing power collaborative scheduling is performed based on the computing power scheduling strategy. The power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent communicate through a communication network based on a distributed hash table.

2. The computing-coordinated scheduling method according to claim 1, characterized in that, The computing power data of the target node is stored in the blockchain network; Before acquiring the power calculation data of the target node, the process also includes: Identify multiple data acquisition nodes; The weight of each data acquisition node is determined based on a preset algorithm; Based on the weight of each data acquisition node, a consensus election is performed to select the target node from all the data acquisition nodes. The computing data of the target node is uploaded to the blockchain to construct the blockchain network.

3. The computer-computer collaborative scheduling method according to claim 2, characterized in that, The preset algorithm is either a practical Byzantine fault-tolerant algorithm or a proof-of-stake variant algorithm. The determination of the weight of each data acquisition node based on a preset algorithm includes: If the preset algorithm is a practical Byzantine fault-tolerant algorithm, then based on the historical performance and data accuracy of each data acquisition node, the reputation value of each data acquisition node is determined according to the practical Byzantine fault-tolerant algorithm, and based on the reputation value of each data acquisition node, the weight of each data acquisition node is determined. If the preset algorithm is a Proof-of-Stake variant algorithm, then the weight of each data collection node is determined based on the real-time contribution and network load of each data collection node using the Proof-of-Stake variant algorithm.

4. The computer-based collaborative scheduling method according to claim 1, characterized in that, The number of target nodes is at least two, each of which is a participant in a multi-party secure computation, and all target nodes communicate through an unintentional transport protocol. The step of training a federated learning model based on the computing power data to obtain a computing power federated model includes: Data sample analysis and feature analysis are performed on the power calculation data of all the target nodes to obtain the analysis results; If the analysis results show that the data feature dimensions of each participant are similar but the data samples are different, then based on the computing power data of each participant, horizontal federated learning training is performed to obtain the computing power federated model; during the horizontal federated learning training process, the computing power data of each participant is encrypted using a homomorphic encryption method. If the analysis results show that the data samples of each participant are similar but the data feature dimensions are different, then based on the computing power data of each participant, longitudinal federated learning training is performed to obtain the computing power federated model; during the longitudinal federated learning training process, the computing power data of each participant is encrypted using a secret sharing method.

5. The computer-based collaborative scheduling method according to claim 1, characterized in that, The construction of a multi-agent system based on the computing data and the computing federation model includes: Based on the power data, the power intelligent agent is constructed; Based on the computing power data, the computing power intelligent agent is constructed; Based on the power data and the computing power data, the association information between the power data and the computing power data is generated through the computing federation model; Based on the correlation information between the power data and the computing power data, the resource coordination intelligent agent is constructed.

6. The computer-based collaborative scheduling method according to claim 1, characterized in that, The computing power scheduling strategy includes power scheduling schemes, computing power allocation schemes, and resource allocation strategies; The step of generating a computing power scheduling strategy based on the power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent, and performing computing power collaborative scheduling based on the computing power scheduling strategy, includes: Based on the real-time power data of the target node, the power dispatching scheme is generated by the power intelligent agent; Based on the real-time computing power data of the target node, the computing power allocation scheme is generated by the computing power agent; Based on the real-time power data and the real-time computing power data, the resource allocation strategy is generated by the resource coordination agent. Based on the power dispatching scheme, the computing power allocation scheme, and the resource allocation strategy, a collaborative scheduling of computing and power is carried out.

7. A computer-aided collaborative scheduling device, characterized in that, include: The acquisition module is used to acquire the computing power data of the target node; the computing power data includes computing power data and power data. The federated learning module is used to perform federated learning training based on the computing power data to obtain a computing power federated model; A multi-agent construction module is used to construct a multi-agent system based on the power computing data and the power computing federation model; the multi-agent system includes a power agent, a computing power agent, and a resource coordination agent; The computing power collaborative scheduling module is used to generate a computing power scheduling strategy based on the power intelligent agent, the computing power intelligent agent and the resource coordination intelligent agent, and to perform computing power collaborative scheduling based on the computing power scheduling strategy. The power intelligent agent, the computing power intelligent agent, and the resource coordination intelligent agent communicate through a communication network based on a distributed hash table.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the computer-computer collaborative scheduling method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the computer-powered collaborative scheduling method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the computer-powered collaborative scheduling method as described in any one of claims 1 to 6.

Images