An intelligent computing collaboration method and system

Abstract

The present application relates to a kind of intelligent calculation cooperation method and system.The method includes the calculation power state data of the heterogeneous computing power server, the power supply state data of power supply side equipment and the energy storage state data of energy storage equipment in physical resource layer;In execution control layer, heterogeneous equipment raw data is unified as CEC-Frame format's calculation power cooperation data frame, and is released to data cooperation layer to form current scheduling period data stream and historical scheduling period data sequence;In intelligent decision-making layer, power resource demand, power generation capacity, energy storage charge-discharge state and power are predicted, and based on power supply-load matching degree, cooperative scheduling strategy is generated;In execution control layer, equipment scheduling instruction is generated and executed;In blockchain credible layer, equipment scheduling instruction, execution result and green electricity consumption certificate's abstract information are chained, and chain on record information is written back to data cooperation layer.Calculation power data and power operation data are cooperated to process, and the double-target minimization of electricity cost and carbon emission is achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent computing technology, specifically relating to an intelligent computing collaboration method and system. Background Technology

[0002] With the rapid development of new technologies such as artificial intelligence and big data, intelligent computing centers have become a core infrastructure supporting digital transformation. At the same time, the computing power of intelligent computing centers is constantly expanding, energy consumption is rising sharply, and electricity expenditure is becoming an increasingly prominent proportion of their operating costs. Driven by dual carbon targets, intelligent computing centers face multiple pressures: improving energy efficiency, reducing carbon emissions, and controlling operating costs.

[0003] Existing intelligent computing center management systems typically design and operate computing resource scheduling and power energy management as two separate technical approaches. This existing architecture of separating computing and power consumption prevents the proactive adjustment of computing load to reduce electricity costs using time-of-use pricing signals; it also prevents the dynamic addition of computing tasks to absorb clean energy during periods of surplus green electricity; and peak electricity consumption by computing devices can drive up electricity prices without the energy system's ability to intervene in advance. Furthermore, the different interfaces of computing devices from various manufacturers and the diverse communication protocols of power equipment make data collection and integration difficult. More critically, the lack of reliable data exchange mechanisms and tamper-proof auditing and traceability measures hinders the automation and mutual trust of green electricity consumption certification, carbon emission reduction accounting, and cross-entity settlement when coordinating scheduling between multiple intelligent computing centers or between intelligent computing centers and the power grid or carbon trading platforms.

[0004] The main drawbacks of existing intelligent power dispatching technologies are severe data silos. Computational power monitoring systems and energy management systems operate independently, with incompatible data formats, making it impossible to form a unified decision-making view. There is a lack of a unified data model; even when computing power and electricity data are collected, there is no standardized data structure to express the spatiotemporal coupling relationship between them. Furthermore, there is a lack of prediction-based dynamic optimization capabilities, resulting in poor adaptability. There is no cross-entity trusted interaction mechanism; when multiple intelligent computing centers collaborate on dispatching, or when intelligent computing centers collaborate with the power grid / carbon exchange, data is easily tampered with, making it difficult to gain third-party trust in green electricity consumption and carbon emission reduction, and hindering automatic settlement. Simultaneously, the computing power load and equipment health status of intelligent computing centers are trade secrets, making it difficult for existing solutions to ensure data usability without visibility when participating in collaborative dispatching. Summary of the Invention

[0005] This invention aims to overcome the above-mentioned technical problems of existing computing power scheduling technology, and provides an intelligent computing collaboration method and system. By constructing a layered and decoupled system architecture, it realizes plug-and-play and unified data collection of heterogeneous computing power devices and power equipment, and achieves collaborative linkage of computing power status, power status and carbon emission data, thereby minimizing the dual objectives of electricity cost and carbon emission.

[0006] To achieve the above objectives, this invention provides an intelligent computing collaboration method, comprising: collecting raw data at the physical resource layer, the raw data including computing power status data of heterogeneous computing power servers, power supply status data of power supply-side equipment, and energy storage status data of energy storage equipment; unifying the raw data of heterogeneous equipment into a standard data frame at the execution control layer to obtain a computing-power collaboration data frame in CEC-Frame format; publishing the computing-power collaboration data frame to the data collaboration layer to form a current scheduling cycle data stream and a historical scheduling cycle data sequence; predicting computing power resource demand forecast curves, power generation capacity forecast curves, energy storage charging and discharging status and power based on the data streams and data sequences at the intelligent decision layer, and generating a collaborative scheduling strategy based on the power supply-load matching degree; generating and executing equipment scheduling instructions according to the collaborative scheduling strategy at the execution control layer, the equipment scheduling instructions including computing power scheduling instructions for heterogeneous computing power servers, output power instructions for power supply equipment, and charging and discharging instructions for energy storage equipment; and uploading the equipment scheduling instructions, execution results, and summary information of green electricity consumption certificates to the blockchain at the blockchain trust layer, and writing the returned on-chain evidence information back to the data collaboration layer.

[0007] Furthermore, the CEC-Frame format computing-electricity collaborative data frame includes: a frame header field, a computing power status field, an electricity status field, an environment and carbon emission field, a collaborative relationship field, a blockchain notarization field, and a privacy protection field; The frame header field includes a timestamp, region ID, frame sequence number, and data source signature; the computing power status field includes total computing power, utilization rate, power consumption, device details, and task priority; the power status field includes grid power, time-of-use electricity price, photovoltaic or wind power, energy storage capacity percentage, and energy storage health; the environment and carbon emission field includes temperature, grid carbon emission intensity, and carbon price; the collaborative relationship field includes power supply-load matching degree, economic adjustability margin, and recommended action vector; the blockchain notarization field includes transaction hash value, block height data, and signature; and the privacy protection field includes zero-knowledge proof string and commitment value.

[0008] Furthermore, raw data is collected at the physical resource layer, and the raw data from heterogeneous devices is unified into standard data frames at the execution control layer. This includes: devices at the physical resource layer have a built-in Trusted Execution Environment (TEE) or security chip and are configured with a national cryptographic algorithm library; each device collects its own status data according to a preset collection cycle, encrypts the status data using a symmetric encryption algorithm, and signs it using the device's private key; the execution control layer receives the encrypted and signed status data, decrypts and verifies the signature, and encapsulates it into a CEC-Frame format computing-electronics collaborative data frame after successful verification through device driver and protocol adaptation; the computing-electronics collaborative data frame is published to the corresponding data topic in the data collaboration layer according to the device type or data type.

[0009] Furthermore, before the intelligent decision-making layer makes predictions based on the data stream and data sequence, the process includes: obtaining a CEC-Frame sequence within a preset time window from the data collaboration layer and filling in any missing fields in the CEC-Frame sequence; when multiple intelligent computing centers work collaboratively, each intelligent computing center uses its own computing power consumption, green electricity supply, energy storage status, or carbon emission status as private inputs, and jointly calculates the global computing power-electricity matching degree through a secure multi-party computation protocol; generating a zero-knowledge proof string based on the private inputs and commitment values, and writing the zero-knowledge proof string into the privacy protection field of the computing-electricity collaboration data frame; and the data receiver verifying whether the global computing power-electricity matching degree meets the preset matching conditions based on the zero-knowledge proof string.

[0010] Furthermore, predictions are made at the intelligent decision-making layer based on the data stream and data sequence, including: pre-training the Informer model based on historical CEC-Frame data in the historical scheduling cycle data sequence; reconstructing the fused CEC-Frame sequence into a three-dimensional array, the three-dimensional array including batch dimension data, time step dimension data, and feature dimension data, the feature dimension data including numerical fields of computing power, electricity, and carbon emissions; inputting the three-dimensional array into the Informer model, and outputting the computing power resource demand prediction curve, the microgrid power generation capacity prediction curve, and the energy storage charging and discharging status and power within a future set time window.

[0011] Furthermore, at the intelligent decision-making layer, a collaborative scheduling strategy is generated based on the power supply-load matching degree, including: pre-setting the optimization objective function and constraints, and configuring the MPC solver; constructing an optimization problem with the goal of minimizing total cost; setting the maximum and minimum power consumption of computing devices, the energy storage SOC range, the maximum output power of the microgrid, and the upper limit of the power required by the grid as optimization constraints; using MPC to solve the optimal control sequence in the prediction time domain, and introducing a deep Q network to dynamically adjust the penalty weight coefficient in the optimization objective function using historical execution results as reward signals, thereby obtaining the collaborative scheduling strategy. The total cost is the sum of electricity purchase cost, carbon emission cost, service level agreement (SLA) breach penalty, and energy storage lifespan reduction cost; the electricity purchase cost is the cost of purchasing electricity from the grid, taking into account time-of-use pricing differences; the carbon emission cost is the product of the purchased electricity volume, the grid carbon emission intensity, and the carbon price; the service level agreement (SLA) breach penalty is the penalty incurred due to delays or suspensions of low-priority tasks; and the energy storage lifespan reduction cost is the lifespan reduction cost caused by energy storage charge-discharge cycles. Furthermore, at the execution control layer, device scheduling instructions are generated and executed according to the collaborative scheduling strategy, including: decomposing the collaborative scheduling strategy into computing power scheduling instructions, microgrid control instructions, and energy storage control instructions; the computing power scheduling agent executes at least one operation among Pod pause, Pod resume, task migration, or GPU downclocking according to the computing power scheduling instructions; the microgrid inverter adjusts the output power of the power supply-side equipment according to the microgrid control instructions; the energy storage battery management system interface controls the charging state, discharging state, and charging / discharging power of the energy storage device according to the energy storage control instructions; during the instruction execution process, the actual power and energy storage SOC are collected, and when the deviation between the actual power and the device scheduling instructions exceeds a preset threshold, the local PID controller is triggered for correction.

[0012] Furthermore, at the blockchain trusted layer, the summary information of equipment scheduling instructions, execution results, and green electricity consumption certificates is uploaded to the chain, and the returned on-chain evidence information is written back to the data collaboration layer. This includes: at the end of each scheduling cycle, packaging the summary information of the equipment scheduling instructions actually executed in this scheduling cycle, the actual electricity consumption, green electricity consumption, carbon emissions, and the hash value of the energy storage SOC change curve into on-chain transaction data; submitting the on-chain transaction data to the consortium blockchain node for verification; after the on-chain transaction data is verified, obtaining the transaction hash value and block height data returned by the consortium blockchain; and writing the transaction hash value and block height data into the blockchain evidence storage field of the subsequent computing and electricity collaboration data frame in the data collaboration layer.

[0013] Furthermore, it also includes: after accumulating actual execution data for a preset period, initiating a feedback learning process; calculating a prediction deviation based on the actual execution data, wherein the prediction deviation is the deviation between the predicted value output by the Informer model and the corresponding actual execution data; performing incremental learning on the Informer model based on the prediction deviation within the most recent preset time period; and using the difference between the actual total cost and the MPC prediction cost as a reward signal to update the network parameters of the deep Q network.

[0014] This invention also provides an intelligent computing collaborative system, comprising: a physical resource layer for collecting raw data, the raw data including computing power status data of heterogeneous computing power servers, power supply status data of power supply-side equipment, and energy storage status data of energy storage devices; an execution control layer, communicatively connected to the physical resource layer, for unifying the raw data of heterogeneous devices into standard data frames to obtain CEC-Frame format computing-power collaborative data frames, and generating and executing device scheduling instructions according to a collaborative scheduling strategy, the device scheduling instructions including computing power scheduling instructions for heterogeneous computing power servers, output power instructions for power supply equipment, and charging and discharging instructions for energy storage devices; and a data collaboration layer, which... The execution control layer communication connection is used to receive the computing power coordination data frame and form the current scheduling cycle data stream and the historical scheduling cycle data sequence; the intelligent decision layer, which is connected to the data coordination layer and the execution control layer, is used to predict the computing power resource demand prediction curve, the power generation capacity prediction curve, the energy storage charging and discharging status and power based on the data stream and data sequence, and to generate the coordinated scheduling strategy based on the power supply-load matching degree; the blockchain trust layer, which is connected to the intelligent decision layer and the data coordination layer, is used to put the summary information of the equipment scheduling instructions, execution results and green electricity consumption certificates on the blockchain, and write the returned on-chain evidence information back to the data coordination layer.

[0015] Compared with existing technologies, the intelligent computing collaborative method and system proposed in this invention have the following advantages: This invention uses CEC-Frame to uniformly encapsulate computing power, electricity, energy storage, and carbon emission data, breaking down data silos between heterogeneous devices and systems; it generates collaborative scheduling strategies through prediction, MPC optimization, and reinforcement learning, reducing electricity purchase costs and carbon emissions, and increasing green electricity consumption; and it achieves trusted cross-entity data interaction, green electricity traceability, and automatic settlement through privacy computing and blockchain notarization, thereby improving scheduling security, auditability, and continuous optimization capabilities. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent computing collaboration method provided in the embodiments of the present invention; Figure 2 This is a hierarchical structure diagram of the intelligent computing collaborative system provided in the embodiments of the present invention. Detailed Implementation

[0017] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0018] like Figure 1As shown, this embodiment provides an intelligent computing collaboration method, including: collecting raw data at the physical resource layer, the raw data including computing power status data of heterogeneous computing power servers, power supply status data of power supply-side equipment, and energy storage status data of energy storage equipment; unifying the raw data of heterogeneous equipment into a standard data frame at the execution control layer to obtain a computing-power collaboration data frame in CEC-Frame format; publishing the computing-power collaboration data frame to the data collaboration layer to form a current scheduling cycle data stream and a historical scheduling cycle data sequence; predicting the computing power resource demand forecast curve, power generation capacity forecast curve, energy storage charging and discharging status and power based on the data stream and data sequence at the intelligent decision layer, and generating a collaborative scheduling strategy based on the power supply-load matching degree; generating and executing equipment scheduling instructions according to the collaborative scheduling strategy at the execution control layer; and uploading the summary information of the equipment scheduling instructions, execution results, and green electricity consumption certificates to the blockchain trust layer, and writing the returned on-chain evidence information back to the data collaboration layer.

[0019] Specifically, the computing power status data is used to characterize the resource usage, task load, and energy consumption status of heterogeneous computing power servers during the current collection period, including: CPU utilization (unit: %, collection interface: IPMI / Redfish API, collection period: 10 seconds), GPU utilization (unit: %, collection interface: NVML / ROCm SMI, collection period: 10 seconds), NPU utilization (unit: %, collection interface: vendor-defined API, collection period: 10 seconds), memory / video memory usage (unit: GB, collection interface: IPMI or system call, collection period: 10 seconds), total memory / video memory (unit: GB, collection interface: IPMI or system call, collection period: 1 minute), task queue length (unit: number of tasks, collection interface: task scheduler interface, collection period: 10 seconds), and current power consumption (unit: W, collection interface: IPMI or power management interface, collection period: 30 seconds).

[0020] Power supply status data is used to characterize the power supply capacity, cost, and stability of the external power grid and local microgrid during the current data collection period. This includes: real-time grid power (unit: kW, data collection interface: smart meter or Modbus, data collection period: 1 minute), real-time grid electricity price (unit: yuan / kWh, data collection interface: electricity market API, data collection period: 15 minutes), real-time wind power (unit: kW, data collection interface: wind turbine SCADA interface, data collection period: 1 minute), green electricity ratio of wind power (unit: %, data collection interface: energy management system, data collection period: 1 minute), real-time photovoltaic power (unit: kW, data collection interface: inverter Modbus, data collection period: 1 minute), and power supply stability index (value range: 0 to 1, data collection or calculation method: calculated locally based on historical availability, data collection period: 5 minutes).

[0021] Energy storage status data is used to characterize the available energy and charging / discharging capabilities of energy storage devices, including: remaining energy storage capacity (SOC) (unit: %, acquisition interface: energy storage BMS or CAN bus, acquisition cycle: 30 seconds), energy storage charging / discharging power (unit: kW, acquisition interface: energy storage BMS, acquisition cycle: 30 seconds), energy storage capacity (unit: kWh, source: static configuration, update cycle: once a day), and energy storage uptime (unit: minutes, calculation method: estimated based on SOC and current load, acquisition cycle: 1 minute).

[0022] like Figure 2 As shown, the intelligent computing collaborative system corresponding to the above method adopts a five-layer collaborative architecture, including the physical resource layer (L1), execution control layer (L2), data collaboration layer (L3), intelligent decision-making layer (L4), and blockchain trust layer (L5). The physical resource layer (L1) includes heterogeneous computing power servers (such as GPU / NPU / DPU), microgrids (such as photovoltaic arrays and wind turbines), energy storage systems (such as lithium batteries and flow batteries), and external power grid access points.

[0023] The execution control layer (L2) includes a computing power scheduling agent (based on Kubernetes CRD (Custom Resource Definition) extension), a microgrid inverter controller, an energy storage battery management system (BMS) interface, and a programmable logic controller (PLC), which is used to translate upper-layer decisions into equipment scheduling instructions. These equipment scheduling instructions include computing power scheduling instructions for heterogeneous computing power servers, output power instructions for power supply equipment, and charging and discharging instructions for energy storage equipment.

[0024] The Data Collaboration Layer (L3) is built upon a distributed message queue and a time-series database, providing standardized data publish-subscribe services. The current scheduling cycle data stream refers to the continuously published CEC-Frame data within this scheduling cycle, while the historical scheduling cycle data sequence refers to a data sequence formed by multiple completed CEC-Frames in chronological order. The Intelligent Decision Layer (L4) includes a large model prediction engine (Informer), a model predictive control (MPC) optimizer, and a reinforcement learning feedback module. The Blockchain Trust Layer (L5) is deployed above the Intelligent Decision Layer, employing a consortium blockchain framework (such as Hyperledger Fabric).

[0025] In this embodiment, the current scheduling cycle data stream refers to a sequence of CEC-Frames continuously published within an unfinished scheduling cycle and available for near real-time subscription by the intelligent decision-making layer. The default scheduling cycle length is 5 minutes, which can be dynamically configured based on business latency requirements, data volume, and system load. Within each scheduling cycle, the execution control layer continuously publishes the CEC-Frames of each device to the distributed message queue, forming the current scheduling cycle data stream. Downstream subscribers process the data in ascending order according to the timestamp field in the CEC-Frame to ensure data time sequence consistency.

[0026] For duplicate frames corresponding to the same device_id and timestamp, the data coordination layer retains the latest valid frame and removes duplicate frames caused by network retransmissions or interface redundancy. The streaming data for the current period can be cached in memory for low-latency reading by the large model prediction engine. After each scheduling period ends, the data coordination layer writes the CEC-Frames of that period into the time-series database in batches and creates a time index by device_id and timestamp to form a historical scheduling period data sequence. Through the above sorting, deduplication, caching, and database storage processes, data time-series consistency can be improved, and the interference of out-of-order and duplicate data on scheduling decisions can be eliminated.

[0027] The interaction relationships between the layers are as follows: L1→L2: Device driver and protocol adaptation, raw data reporting; L2→L3: Encapsulate the raw data into a CEC-Frame and publish it to the corresponding data topic; L3→L4: Provide the data stream for the current scheduling cycle and the data sequence for the historical scheduling cycle; L4→L2: Generate and distribute the collaborative scheduling policy; L4→L5: Upload the device scheduling instructions and execution result summary to the blockchain; L5→L3: Write the on-chain evidence information back to the subsequent CEC-Frame in the data collaboration layer.

[0028] Building upon the aforementioned embodiments, to achieve unified expression and trusted interaction between the computing power domain and the power domain, this embodiment defines a standardized data frame based on Protocol Buffers, namely the CEC-Frame format computing-power collaborative data frame. This computing-power collaborative data frame includes a frame header field, a computing power status field, a power status field, an environment and carbon emission field, a collaborative relationship field, a blockchain notarization field, and a privacy protection field, as detailed in the table below: CooperativeRelation calculates the difference between the total power consumption and the green electricity supply at the current moment and provides recommended actions (such as increasing load, reducing load, and charging energy storage), providing direct input for upper-layer decision-making. BlockchainProof is used to match key data in each CEC-Frame (such as scheduling instructions and actual power consumption) with on-chain records. PrivacyPreserving is used to store zero-knowledge proofs to verify whether the current computing power load and green electricity matching degree meet the requirements without exposing specific task content or internal device indicators.

[0029] Furthermore, the CooperativeRelation module takes the total computing power consumption P_total, total green electricity supply G_total, energy storage SOC, real-time electricity price, adjustable load ratio, and node power supply stability index as inputs to calculate the power supply-load matching degree, economic adjustability margin, and recommended action vector. The power supply-load matching degree, in the energy aggregation dimension, can be represented as the normalized value of the difference between green electricity power and computing power consumption, for example, Match_E=(G_total-P_total) / max(P_total,G_total), used to reflect the degree of green electricity surplus or shortage. In the task allocation dimension, this matching degree is also used to characterize the degree of matching between the task's power supply reliability requirements and the candidate node's power supply stability index; the closer the task reliability requirements are to the node stability index, the higher the matching degree.

[0030] For example, if a certain computing task requires a power supply reliability of 0.95, node A has a power supply stability index of 0.98, and node B has a power supply stability index of 0.70, then node A has a higher power supply-load matching degree with the task than node B. The intelligent decision layer will prioritize scheduling the task to node A. The economic adjustability margin can be estimated based on the electricity price difference, available energy storage capacity, and load sensitivity coefficient to determine potential benefits. Recommended action vectors include load adjustment suggestions (increase, decrease, or maintain), energy storage charging and discharging suggestions, and the economic sensitivity coefficient. These are directly used by the intelligent decision layer as input to the MPC optimization and reinforcement learning modules to improve green electricity consumption, reduce electricity costs, and maintain SOC within a safe range.

[0031] Based on the two embodiments described above, this embodiment provides a detailed explanation of trusted data acquisition, data standardization, and topic publishing. The physical resource layer devices have a built-in Trusted Execution Environment (TEE) or security chip and are configured with the national cryptographic SM2 / SM4 algorithm library. Each device collects its own status data according to a preset acquisition cycle (e.g., 5 seconds), encrypts the status data using the SM4 symmetric encryption algorithm, and signs it with its private key (based on the SM2 algorithm). The encrypted and signed data is then reported through the adapter of the execution control layer.

[0032] During the security processing, SM4 is used to perform symmetric encryption on sensitive fields in the CEC-Frame, energy storage / power status data, and data that needs to be stored in the database. The device side or device gateway uses the system entropy source to generate a 128-bit random data encryption key, which is uniformly registered, distributed, and rotated by the Key Management System (KMS). After receiving the encrypted data, the execution control layer calls the corresponding data encryption key to decrypt it. SM2 is used for key pair generation, data signing, and signature verification for entities such as device gateways, scheduling services, and blockchain nodes: Before reporting, the device side uses the SM2 private key to sign the data digest or encrypted message digest, and the execution control layer uses the device public key to verify the signature. Only after the signature verification is successful does the field mapping and CEC-Frame encapsulation process begin.

[0033] The SM2 private key can be stored in the Hardware Security Module (HSM), Trusted Execution Environment (TEE), or encrypted configuration file; the SM4 data encryption key is stored in the KMS and can be updated at a preset cycle, such as every 30 days. When a device is re-registered after going offline, when user permissions are changed, or when a security incident occurs, the system manually triggers key revocation and regeneration to reduce the risk of key leakage and forged data injection.

[0034] The execution control layer receives its own state data after encryption and signature, decrypts and verifies the signature, and encapsulates it into a CEC-Frame format computing-electronics collaborative data frame through device driver and protocol adaptation after successful verification. During initial encapsulation, the Blockchain Proof and PrivacyPreserving fields can be empty; subsequently, the execution control layer publishes the computing-electronics collaborative data frame to the corresponding data topic in the data collaboration layer according to the device type or data type.

[0035] The mapping between raw fields and CEC-Frame fields uses a combination of direct mapping, computational mapping, and static configuration mapping: direct mapping is used when field names are the same or semantically consistent. For example, timestamp, device_id, CPU / GPU utilization, and current power consumption are mapped to the corresponding fields in the frame header field or computing power status field, respectively. Fields that require unit conversion, aggregation, or derivation use computational mapping. For example, the instantaneous power consumption of the server is integrated into the periodic power consumption, the real-time power of wind power and the real-time power of photovoltaic power are aggregated into the total green electricity supply, and the SOC and load are estimated into the energy storage support time. Static fields such as device ID, region ID, device type, and capacity rating are read from the configuration file or database and then written to the corresponding fields.

[0036] For missing or abnormal fields, the execution control layer performs validation before encapsulation: if required fields such as timestamp, device_id, power, and SOC are missing or have abnormal formats, the frame is marked as invalid and discarded, and an error log is recorded. Simultaneously, downstream calculations can use the previous valid value or a preset default value to maintain short-term continuity. If optional fields are missing, the field is ignored during mapping, and the corresponding score is lowered in data_quality. Through the above processing, the execution control layer achieves unified encapsulation of raw data from heterogeneous devices into a standard CEC-Frame, providing a stable, reliable, and real-time computing power-power data foundation for the upper-layer scheduling strategy hub.

[0037] The corresponding data topics can be named according to combinations of region, device type, and data type, or a wildcard subscription method can be used. For example, the intelligent decision-making layer can subscribe to topics such as cec / frame / + / power to pull power data in real time and perform MPC calculations. The data flow process includes: encryption and signing are completed at the L2 execution control layer, and the data is published to the message queue at the L3 data collaboration layer. The consumer side performs routing, signature verification, and decryption. After successful verification, the CEC-Frame is passed to the upper-layer business module. The execution result of the scheduling instruction is returned through the feedback topic (e.g., cec / frame / {region_id} / ack) for closed-loop verification.

[0038] This embodiment provides a specific implementation process for multi-source data fusion and privacy verification. Before the intelligent decision-making layer makes predictions based on the data stream and data sequence, the intelligent decision-making layer obtains the CEC-Frame sequence within a preset time window from the data collaboration layer, for example, pulling the CEC-Frame sequence of the most recent time window (such as the past 6 hours), and fills in the missing fields in the CEC-Frame sequence using linear interpolation.

[0039] When multiple intelligent computing centers collaborate, each center uses its own computing power consumption, green electricity supply, energy storage status, or carbon emission status as private inputs to jointly calculate the global computing power-electricity matching degree through a secure multi-party computation protocol. The data provider generates a zero-knowledge proof string based on the private inputs and commitment values, and writes the zero-knowledge proof string into the privacy protection field of the computing-electricity collaborative data frame. The receiver verifies the correctness of a claim based on privacy.zk_proof, that is, verifies whether the global computing power-electricity matching degree meets the preset matching conditions, without needing to obtain the original data.

[0040] During secure multi-party computation, participants can be multiple entities such as Computing Center A, Computing Center B, and Computing Center C. Each participant's private inputs include local current total power consumption, green electricity supply, energy storage status, and carbon emissions. The output of secure multi-party computation is the global computing power-electricity matching degree or its compliance conclusion. In the default implementation, only the conclusion of "meeting / not meeting preset matching conditions" is disclosed, without revealing the specific matching degree value, to reduce the risk of leakage of sensitive commercial data.

[0041] The communication process for secure multi-party computation includes: each center generates a random secret share and distributes it through a secure channel; each party calculates the addition share locally without disclosing the original input; the share of the maximum value or threshold comparison result is obtained through a secure comparison protocol; the share of the global computing power-electricity matching degree is obtained through a secure multiplication or division protocol; if the business requires the matching degree value to be disclosed, each party broadcasts the share and reconstructs the global computing power-electricity matching degree; if only proof of compliance is required, no reconstruction is performed and the process enters the zero-knowledge proof stage.

[0042] The statement proved by privacy.zk_proof is as follows: During this scheduling cycle, the global computing power-power matching degree of all participating intelligent computing centers meets the preset matching conditions, such as Match_Global≥match_threshold, and the private inputs of all centers (including power consumption, green power, etc.) are all derived from the values ​​that each center has previously committed to and put on the chain or stored.

[0043] The commitment value can be hidden from private inputs using Pedersen commitments, and then recorded on the blockchain as public input or submitted to the verifier. During proof generation, the data provider calculates the commitment value based on the private inputs and a random blind factor, and calls a zero-knowledge proof generation algorithm (such as Bulletproofs' `prove` function) to generate proof π. Proof π, along with the commitment value, is then written to the PrivacyPreserving field or submitted to the blockchain. Upon receiving proof π and the commitment value, the verifier first checks the format of the commitment value, then calls the verification algorithm and inputs a threshold for verification. If verification passes, it confirms that the computing power-green electricity matching degree of the current scheduling cycle meets the requirements, and for example, a green electricity consumption ratio of no less than 80% can be used as a condition for compliance.

[0044] When the intelligent decision-making layer makes predictions based on the data stream and data sequence, the Informer model is first pre-trained based on historical CEC-Frame data from the historical scheduling cycle data sequence. Specifically, after accumulating at least one week's worth of historical CEC-Frame data, it is used for the pre-training of the Informer model (a variant of Transformer).

[0045] The Informer model's training samples are constructed by extracting continuous time series from the historical CEC-Frame time series database, forming "input sequence, output sequence" sample pairs. The true values ​​of the output sequence are fields such as P_total, G_total, SOC, and energy storage charging and discharging power at future L_out time points. The prediction target is the model directly outputting the corresponding future sequence. The training loss function can be a weighted combination of multivariate, multi-step mean absolute error (MAE) and mean squared error (MSE), with different weights set according to the importance of prediction components such as computing power consumption, total green electricity power, and SOC.

[0046] The number of pre-training epochs can be set to 50-100, and an Early Stopping mechanism is adopted. If the validation set loss does not decrease for 10 consecutive epochs, training is stopped. Evaluation metrics include MAE, RMSE, and MAPE. Model updates can be achieved by combining periodic retraining with triggered updates, such as performing full training weekly or monthly using the latest accumulated historical data. Retraining is automatically triggered when the evaluation metrics are below the threshold for three consecutive prediction periods (e.g., MAE exceeds 10%).

[0047] The trained Informer model is deployed in the large-scale prediction engine of the intelligent decision layer. During inference, the prediction engine takes the CEC-Frame data from the most recent L_in time points, constructs feature vectors consistent with those used in the training phase, and incorporates time features and weather forecast features. The model outputs prediction results for the next L_out time points at a time. It analyzes the P_total component to obtain the computing power resource demand prediction curve, the G_total component to obtain the microgrid power generation capacity prediction curve, and the SOC and charging / discharging power components to obtain the future state of energy storage and its charging / discharging power, where positive charging / discharging power indicates discharging and negative power indicates charging. The prediction curves are sent to the MPC optimizer and reinforcement learning module in JSON or time-series data format for pre-scheduling during periods of abundant green electricity, energy storage charging or discharging strategies before future peak electricity prices, and carbon emission management. The prediction engine can perform inference every 5 minutes to maintain dynamic rolling predictions for the next 4 hours.

[0048] The fused CEC-Frame sequence is reconstructed into a three-dimensional array [batch, time_steps, features]. This three-dimensional array includes batch dimension data, time_step dimension data, and feature dimension data, where features contain numerical fields such as computing power, electricity, and carbon emissions. The reconstructed three-dimensional array is input into the Informer model, which outputs a predicted curve for computing power resource demand within a future set time window (e.g., 2 hours) (e.g., one data point every 2 minutes), a predicted curve for microgrid power generation capacity (PV output + wind power output), and the charging and discharging status and power of energy storage.

[0049] The three-dimensional array [batch, time_steps, features] is the standard input format for time series prediction models such as Informer. Here, `batch` represents the number of samples input for model training each time, which can be set to 32, 64, or 128 depending on the GPU memory or system memory capacity; when using a sliding window to generate samples, the samples within a batch are independent of each other. `time_steps` represents the length of the input time window for each sample, i.e., the number of consecutive time points contained in the input sequence; short-term prediction can use 12–48 time points (corresponding to 1–4 hours, with a sampling period of 5 minutes), and long-term prediction can use 96–192 time points (corresponding to 8–16 hours). `features` represents the number of features at each time step, determined by the number of selected CEC-Frame fields, including computing power, electricity, energy storage, environmental, and time features, which can be 20–50 dimensions after filtering.

[0050] During feature normalization, the mean μ and standard deviation σ of each feature are first calculated on the training set, or the min and max ranges are calculated. The same set of statistics is used for standardization or Min-Max normalization on the training, validation, and test sets. The statistics are not recalculated on the test set and online real-time data. Normalization parameters (e.g., JSON or pickle files) are saved during model deployment for real-time data stream processing. When the model output is a normalized predicted value, it is restored to its original dimensions through inverse transformation. During sample construction, the CEC-Frame sequence can be divided into sliding window segments with a 5-minute sampling period. For example, when the input window time_steps is 96, 96 consecutive time points are used as input, and the subsequent L_out time points are used as output labels.

[0051] When generating a collaborative scheduling strategy based on power supply-load matching degree at the intelligent decision-making layer, the optimization objective function and constraints are pre-defined, and an MPC solver (such as CasaADi or OSQP) is configured. By constructing an optimization problem, the total cost is minimized, i.e.: Minimize Σ(C_elec+C_carbon+C_penalty+C_battery_degradation). Where: C_elec: Cost of purchasing electricity from the grid (considering time-of-use pricing differences); C_carbon: Carbon emission cost = Electricity purchase × Grid carbon emission intensity × Carbon price; C_penalty: SLA breach penalty caused by delaying or suspending low-priority tasks; C_battery_degradation: The cost of lifespan reduction caused by energy storage charge-discharge cycles.

[0052] The optimization constraints include the maximum and minimum power consumption of computing devices, the energy storage SOC range (e.g., 20%~90%), the maximum output power of the microgrid, and the upper limit of the power required by the grid. MPC is used to solve for the optimal control sequence (e.g., time resolution of 15 minutes) within the prediction time domain (e.g., 2 hours). Simultaneously, a deep Q-network (DQN) is introduced, using historical execution results as reward signals to dynamically adjust the penalty weight coefficients in the optimization objective function, thus obtaining the cooperative scheduling strategy.

[0053] During the generation of collaborative scheduling strategies, the power supply-load matching degree is used to constrain the selection of candidate nodes and the order of task allocation: for tasks requiring high power supply reliability, the intelligent decision layer prioritizes nodes whose power supply stability index is closer to the task's reliability requirements; during periods of green electricity surplus, the recommended action vector tends to increase delayable or migrateable loads; during periods of green electricity shortage or peak electricity prices, the recommended action vector tends to reduce low-priority loads or call energy storage for discharge. This matching degree participates in solving the MPC optimization problem as a constraint or penalty term, enabling the scheduling strategy to improve green electricity consumption and power supply reliability while satisfying the computing power service level agreement.

[0054] When the execution control layer generates and executes equipment scheduling instructions based on the collaborative scheduling strategy, the execution control layer first establishes communication links with the computing power scheduler (K8s), microgrid inverter, and energy storage BMS, and decomposes the collaborative scheduling strategy into computing power scheduling instructions, microgrid control instructions, and energy storage control instructions.

[0055] The computing power scheduling agent executes at least one of the following operations according to the computing power scheduling instructions: Pod pause, Pod resume, task migration, or GPU frequency reduction; the microgrid inverter adjusts the output power of the power supply-side equipment according to the microgrid control instructions; and the energy storage battery management system interface controls the charging status, discharging status, and charging / discharging power of the energy storage equipment according to the energy storage control instructions.

[0056] During instruction execution, the execution control layer collects the actual power and energy storage SOC; when the deviation between the actual power and the equipment scheduling instruction exceeds a preset threshold, or when the energy storage SOC approaches the preset upper or lower limit, the local PID controller is triggered to make corrections.

[0057] The local PID controller is set at the execution control layer L2. When the deviation between the actual power and the device scheduling command exceeds a preset threshold (e.g., ±5%), or when the energy storage SOC approaches the preset upper or lower limits (e.g., below 20% or above 90%), closed-loop correction is triggered. The inputs of the PID controller include the actual total power P_actual, the scheduling command power P_cmd, the actual energy storage SOC, the lower SOC threshold, and the upper SOC threshold; the output control objects include the server CPU / GPU frequency, the deferred task queue, and the energy storage charging and discharging power.

[0058] The primary objective of PID control is to ensure that the actual total power P_actual follows the dispatch command P_cmd, with a steady-state error not exceeding 2%. The secondary objective is to prevent the energy storage SOC from exceeding the limit, i.e., SOC_low ≤ SOC_actual ≤ SOC_high. PID control is activated when the power deviation exceeds the threshold (e.g., |e_P| > 5kW or relative deviation > 5%) and persists for more than 2 seconds. If the SOC is close to the limit (e.g., SOC < 25% and the dispatch command requires discharge), the PID control adds a SOC correction term and prioritizes the protection of the energy storage.

[0059] The PID output value u(t) can be normalized to the interval [-1, 1] and mapped to the adjustment amount of the corresponding controlled object. The incremental control formula is u(t) = u(t-1) + Kp·[e(t) - e(t-1)] + Ki·e(t) + Kd·[e(t) - 2e(t-1) + e(t-2)], where e(t) is the power deviation, Kp can be configured from 0.5 to 1.2 to respond to the deviation amplitude; Ki can be configured from 0.1 to 0.3 to eliminate steady-state error; and Kd can be configured from 0.05 to 0.2 to suppress overshoot and oscillation. Through local PID correction, the execution control layer can quickly, smoothly, and safely track scheduling commands, ensuring the closed-loop stability of the intelligent computing collaborative system.

[0060] When uploading the summary information of equipment scheduling instructions, execution results, and green electricity consumption certificates to the blockchain's trusted layer, a consortium blockchain network is first deployed, with each participating party (such as a smart computing center, power grid company, and carbon exchange) set as a consortium blockchain node. After each scheduling cycle (e.g., 15 minutes), the summary information of the equipment scheduling instructions actually executed in this scheduling cycle, the actual electricity consumption, the amount of green electricity consumed, the carbon emissions, and the hash value of the energy storage SOC change curve are packaged into on-chain transaction data.

[0061] The field structure of on-chain transaction data includes: period_id, a string representing a unique identifier for the scheduling period, in the format YYYYMMDDHHMM, obtained by aggregating the timestamp of the CEC-Frame; device_list, an array representing a list of device IDs participating in the current period, obtained by deduplicating the device_id in each CEC-Frame within the period; and schedule_instructions_hash, a string representing a summary of device scheduling instructions, obtained by hashing the JSON serialized scheduling instructions issued by the intelligent decision layer, and can correspond to the instruction_id field in the extended field of the CEC-Frame.

[0062] On-chain transaction data also includes: actual_power_kwh, a number representing the total actual electricity consumption in the current period, in kWh, calculated by integrating the power_draw_w of each server over time; green_consumption_kwh, a number representing the actual green electricity consumption in the current period, in kWh, which can be obtained by multiplying green_pct by the actual electricity consumption, or by integrating wind power and photovoltaic power; and carbon_emission_kg, a number representing the actual carbon emissions in the current period, in kgCO2, which can be calculated by combining the actual electricity consumption with the carbon emission factor of each node. If CEC-Frame directly collects the carbon_emission field, it will be summed over the period.

[0063] On-chain transaction data also includes: soc_curve_hash, a string representing the hash value of the energy storage SOC change curve, obtained by serializing and hashing all battery_soc and timestamps within the period; match_degree, a number representing the average power supply-load matching degree for this period, which can be obtained by averaging (G_total-P_total) / max(P_total,G_total) at each time point within the period, or by combining the node power supply stability index and actual load; data_quality, a number representing the comprehensive data quality score for this period, obtained by weighted averaging of data_quality in each CEC-Frame; and timestamp, an integer representing the transaction generation time, which can use a Unix millisecond timestamp.

[0064] The on-chain hash algorithm can uniformly adopt the Chinese national cryptographic standard SM3, outputting a 256-bit hexadecimal string; the signature can use the Chinese national cryptographic standard SM2 to sign the original transaction data. The signature fields include signature, public_key, and key_id, where signature is the submitter's SM2 private key signing the original transaction, public_key is the submitter's SM2 public key, and key_id is the key identifier; if the submitter's identity has already been registered on the chain, public_key can be omitted. Consortium chain nodes use the submitter's SM2 public key to verify whether the signature matches the original transaction; if the verification fails, the transaction is rejected.

[0065] The correspondences with the CEC-Frame fields are as follows: actual_power_kwh is obtained by integrating power_draw_w within the period; green_consumption_kwh can be calculated from green_pct, wind_power_kw, and solar_power_kw; carbon_emission_kg can be obtained by summing the carbon_emission field or by multiplying actual_power_kwh by carbon_factor; soc_curve_hash is generated from battery_soc and the timestamp sequence; match_degree is calculated from aggregate fields such as G_total and P_total; data_quality is obtained by averaging the data_quality fields of each frame; and period_id is obtained by aligning the first timestamp according to the period.

[0066] The on-chain transaction data is submitted to the consortium blockchain node for verification. After the on-chain transaction data is verified, the transaction hash value and block height data returned by the consortium blockchain are obtained. The transaction hash value and block height data are written into the blockchain storage field of subsequent computing and electricity collaboration data frames in the data collaboration layer, thereby binding the data frame with the on-chain record. The smart contract automatically executes the settlement operation according to preset rules (such as automatically issuing carbon credits when the green electricity consumption rate exceeds 80%).

[0067] The smart contract's preset rules use the scheduling cycle as the settlement granularity. A trigger condition is that the ratio of `green_consumption_kwh` to `actual_power_kwh` reaches or exceeds 80% within a given scheduling cycle. When the trigger condition is met, the smart contract automatically distributes carbon credits or corresponding rewards to the participants in that scheduling cycle. The settlement result, along with the transaction hash value and block height data returned by the consortium blockchain, can be written into the BlockchainProof field of the subsequent CEC-Frame. Transactions that fail on-chain verification will not receive credits. For duplicate submissions with the same `period_id` and `schedule_instructions_hash`, idempotency checks can be performed based on existing on-chain records to avoid duplicate settlements.

[0068] This embodiment provides a detailed explanation of feedback learning and model updates. After accumulating actual execution data for a preset period (e.g., 24 hours), the feedback learning process is initiated. First, the prediction deviation is calculated based on the actual execution data. This prediction deviation is the discrepancy between the predicted value output by the Informer model and the corresponding actual execution data, such as the discrepancy between predicted values ​​for computing power demand, photovoltaic output, etc., and the actual execution data.

[0069] The prediction deviation calculation is performed on a scheduling cycle basis. The prediction fields involved in the calculation include the total power consumption P_total, total green power G_total, energy storage SOC, energy storage charging and discharging power, and carbon emissions when necessary, output by Informer. The corresponding actual fields come from the CEC-Frame after execution or on-chain transaction data, including the actual power integral of each server, the actual power integral of wind power / solar power, battery_soc, actual charging and discharging power of energy storage, and carbon_emission_kg, etc.

[0070] At the end of each scheduling cycle, the system collects actual execution data. After accumulating a preset cycle (e.g., 24 hours, or 96 15-minute cycles), a feedback learning process is initiated, and statistical values ​​such as absolute error, relative error, MAE, RMSE, and MAPE are calculated. If the prediction deviation exceeds a preset threshold, for example, if the MAE exceeds 10% and persists for multiple prediction cycles, the system retrains the Informer model using the latest accumulated historical data (usually from the past 2-4 weeks). The training method can be full training or incremental fine-tuning. If the performance improvement of the new model on the validation set exceeds 5%, the production model is replaced; otherwise, the old model is retained and the trigger threshold is appropriately increased. The system also writes the trigger reason, deviation statistics, and model version to the blockchain audit log.

[0071] Then, based on the prediction deviation within the most recent preset time period, such as the deviation data of the last 7 days, the Informer model is incrementally learned; the difference between the actual total cost and the MPC predicted cost is then used as a reward signal to update the network parameters of the deep Q network, where the actual total cost includes electricity costs, carbon costs, and service level agreement (SLA) default penalties. Through the above steps, a closed-loop optimization system of prediction, decision-making, execution, feedback, and optimization is formed.

[0072] The reward signal is calculated based on the difference between the actual total cost and the MPC predicted cost. The MPC predicted cost, C_mpc, is derived from the electricity purchase cost, carbon emission cost, SLA default penalty, and energy storage lifetime loss cost calculated by the MPC solver when generating dispatch instructions. The actual total cost, C_actual, is derived from the actual electricity consumption, electricity price, carbon emission factor, actual SLA default situation, and energy storage cycle loss statistics after the end of the dispatch cycle. The reward can be normalized using r=(C_mpc-C_actual) / max(|C_mpc|,ε) and can be clipped to the [-1,1] interval to maintain consistency in the reward scale across different cycles.

[0073] The DQN parameter update process includes: the Informer outputs prediction results periodically; the MPC generates scheduling instructions and prediction cost C_mpc based on the prediction results; the control layer executes the instructions and collects actual data; at the end of the period, the actual cost C_actual and the reward r are calculated; the state, action, reward, and next state (s,a,r,s′) are written into the experience pool; the DQN is trained periodically and the penalty weight coefficients λ, α, β, γ, and δ are updated; the new weights participate in the construction of the MPC optimization objective for the next period. Simultaneously, the accumulated prediction bias triggers incremental learning by the Informer to improve subsequent prediction accuracy.

[0074] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent computing collaboration, characterized in that, include: Raw data is collected at the physical resource layer, including computing power status data of heterogeneous computing power servers, power supply status data of power supply-side equipment, and energy storage status data of energy storage equipment. At the execution control layer, the raw data from heterogeneous devices are unified into a standard data frame, resulting in a CEC-Frame format computing and power collaboration data frame. The computing and power coordination data frame is published to the data coordination layer to form the current scheduling cycle data stream and the historical scheduling cycle data sequence; At the intelligent decision-making layer, based on the data stream and data sequence, the computing power resource demand prediction curve, the power generation capacity prediction curve, the energy storage charging and discharging status and power are predicted, and a collaborative scheduling strategy is generated based on the power supply-load matching degree. The execution control layer generates and executes device scheduling instructions based on the collaborative scheduling strategy. The device scheduling instructions include: computing power scheduling instructions for heterogeneous computing power servers, output power instructions for power supply equipment, and charging and discharging instructions for energy storage equipment. In the blockchain trusted layer, the summary information of equipment scheduling instructions, execution results and green electricity consumption certificates are uploaded to the chain, and the returned on-chain evidence information is written back to the data collaboration layer.

2. The intelligent computing collaboration method according to claim 1, characterized in that, The CEC-Frame format computing-electricity collaborative data frame includes: a frame header field, a computing power status field, an electricity status field, an environment and carbon emission field, a collaborative relationship field, a blockchain notarization field, and a privacy protection field; The frame header field includes a timestamp, region ID, frame sequence number, and data source signature; the computing power status field includes total computing power, utilization rate, power consumption, device details, and task priority; the power status field includes grid power, time-of-use electricity price, photovoltaic or wind power, energy storage capacity percentage, and energy storage health; the environment and carbon emission field includes temperature, grid carbon emission intensity, and carbon price; the collaborative relationship field includes power supply-load matching degree, economic adjustability margin, and recommended action vector; the blockchain notarization field includes transaction hash value, block height data, and signature; and the privacy protection field includes zero-knowledge proof string and commitment value.

3. The intelligent computing collaboration method according to claim 1, characterized in that, Raw data is collected at the physical resource layer, and the raw data from heterogeneous devices is unified into a standard data frame at the execution control layer. This includes: devices at the physical resource layer have a built-in Trusted Execution Environment (TEE) or security chip and are configured with a national cryptographic algorithm library; each device collects its own status data according to a preset collection cycle, encrypts the status data using a symmetric encryption algorithm, and signs it using the device's private key; the execution control layer receives the encrypted and signed status data, decrypts and verifies the signature, and encapsulates it into a CEC-Frame format computing-electronics collaborative data frame after successful verification through device driver and protocol adaptation; the computing-electronics collaborative data frame is published to the corresponding data topic in the data collaboration layer according to the device type or data type.

4. The intelligent computing collaboration method according to claim 2, characterized in that, Before the intelligent decision-making layer makes predictions based on the data stream and data sequence, the process includes: obtaining a CEC-Frame sequence within a preset time window from the data collaboration layer and filling in any missing fields in the CEC-Frame sequence; when multiple intelligent computing centers work collaboratively, each intelligent computing center uses its own computing power consumption, green electricity supply, energy storage status, or carbon emission status as private inputs, and jointly calculates the global computing power-electricity matching degree through a secure multi-party computation protocol; generating a zero-knowledge proof string based on the private inputs and commitment values, and writing the zero-knowledge proof string into the privacy protection field of the computing-electricity collaboration data frame; and the data receiver verifying whether the global computing power-electricity matching degree meets the preset matching conditions based on the zero-knowledge proof string.

5. The intelligent computing collaboration method according to claim 1, characterized in that, The intelligent decision-making layer performs predictions based on the data stream and data sequence, including: pre-training the Informer model based on historical CEC-Frame data in the historical scheduling cycle data sequence; reconstructing the fused CEC-Frame sequence into a three-dimensional array, the three-dimensional array including batch dimension data, time step dimension data and feature dimension data, the feature dimension data including numerical fields of computing power, electricity and carbon emissions; inputting the three-dimensional array into the Informer model, and outputting the computing power resource demand prediction curve, the microgrid power generation capacity prediction curve, and the energy storage charging and discharging status and power within a future set time window.

6. The intelligent computing collaboration method according to claim 1, characterized in that, The intelligent decision-making layer generates a collaborative scheduling strategy based on power supply-load matching degree, including: pre-setting the optimization objective function and constraints, and configuring the MPC solver; constructing an optimization problem with the goal of minimizing total cost; setting the maximum and minimum power consumption of computing devices, the energy storage SOC range, the maximum output power of the microgrid, and the upper limit of the power required by the grid as optimization constraints; using MPC to solve the optimal control sequence in the prediction time domain, and introducing a deep Q network to dynamically adjust the penalty weight coefficient in the optimization objective function with historical execution results as reward signals, thereby obtaining the collaborative scheduling strategy. The total cost is the sum of electricity purchase cost, carbon emission cost, service level agreement (SLA) breach penalty, and energy storage lifespan reduction cost; the electricity purchase cost is the cost of purchasing electricity from the grid, taking into account time-of-use pricing differences; the carbon emission cost is the product of the purchased electricity volume, the grid carbon emission intensity, and the carbon price; the service level agreement (SLA) breach penalty is the penalty incurred due to delays or suspensions of low-priority tasks; and the energy storage lifespan reduction cost is the lifespan reduction cost caused by energy storage charge-discharge cycles.

7. The intelligent computing collaboration method according to claim 1, characterized in that, The execution control layer generates and executes device scheduling instructions based on the collaborative scheduling strategy, including: decomposing the collaborative scheduling strategy into computing power scheduling instructions, microgrid control instructions, and energy storage control instructions; the computing power scheduling agent executes at least one operation among Pod pause, Pod resume, task migration, or GPU downclocking according to the computing power scheduling instructions; the microgrid inverter adjusts the output power of the power supply-side equipment according to the microgrid control instructions; the energy storage battery management system interface controls the charging status, discharging status, and charging / discharging power of the energy storage device according to the energy storage control instructions; during instruction execution, the actual power and energy storage SOC are collected, and when the deviation between the actual power and the device scheduling instructions exceeds a preset threshold, the local PID controller is triggered for correction.

8. The intelligent computing collaboration method according to claim 1, characterized in that, In the blockchain trusted layer, the summary information of equipment scheduling instructions, execution results, and green electricity consumption certificates is uploaded to the chain, and the returned on-chain evidence information is written back to the data collaboration layer. This includes: at the end of each scheduling cycle, packaging the summary information of the equipment scheduling instructions actually executed in this scheduling cycle, the actual electricity consumption, green electricity consumption, carbon emissions, and the hash value of the energy storage SOC change curve into on-chain transaction data; submitting the on-chain transaction data to the consortium blockchain node for verification; after the on-chain transaction data is verified, obtaining the transaction hash value and block height data returned by the consortium blockchain; and writing the transaction hash value and block height data into the blockchain evidence storage field of the subsequent computing and electricity collaboration data frame in the data collaboration layer.

9. The intelligent computing collaboration method according to claim 1, characterized in that, Also includes: Once the actual execution data for the preset period has been accumulated, the feedback learning process will be initiated. The prediction bias is calculated based on the actual execution data. The prediction bias is the deviation between the predicted value output by the Informer model and the corresponding actual execution data. Based on the prediction bias within the most recent preset time period, the Informer model is incrementally learned. The difference between the actual total cost and the MPC prediction cost is used as a reward signal to update the network parameters of the deep Q network.

10. A smart computing collaborative system, characterized in that, include: The physical resource layer is used to collect raw data, including computing power status data of heterogeneous computing power servers, power supply status data of power supply-side equipment, and energy storage status data of energy storage devices. The execution control layer is communicatively connected to the physical resource layer and is used to unify the raw data of heterogeneous devices into standard data frames to obtain CEC-Frame format computing-power coordination data frames. It also generates and executes device scheduling instructions according to the coordination scheduling strategy. The device scheduling instructions include computing power scheduling instructions for heterogeneous computing power servers, output power instructions for power supply equipment, and charging and discharging instructions for energy storage devices. The data coordination layer, communicating with the execution control layer, is used to receive the computing and power coordination data frames and form a data stream for the current scheduling cycle and a data sequence for the historical scheduling cycle. The intelligent decision-making layer, communicating with the data coordination layer and the execution control layer, is used to predict the computing power resource demand forecast curve, the power generation capacity forecast curve, the energy storage charging and discharging status and power based on the data stream and data sequence, and to generate the coordinated scheduling strategy based on the power supply-load matching degree. The blockchain trust layer, communicating with the intelligent decision-making layer and the data coordination layer, is used to put the summary information of the equipment scheduling instructions, execution results and green electricity consumption certificates on the blockchain, and write the returned on-chain evidence information back to the data coordination layer.

Images