Information interaction method and system for collaborative scheduling of data center computing power and power

Abstract

The application relates to the technical field of data center interaction, and specifically discloses a data center computing power and electric power coordinated scheduling information interaction method and system, which integrates multi-source heterogeneous data and constructs a time-space correlation model, realizes accurate prediction and dynamic modeling of data center resource demand, utilizes the real-time learning ability of DRL and the discrete decision advantage of MIP, realizes synchronous optimization of multiple targets, significantly improves the energy efficiency and economy of the data center, the multi-target optimization can realize "computing power-electric power-heat" coupling, through the blockchain technology, the data is ensured to be non-falsifiable, transparent and credible, the safety and auditability of the system are improved, carbon footprint tracking and green electricity certification are supported, the PUE index dynamically optimizes the energy efficiency, directly encourages the use of green electricity, reduces the carbon footprint and operation cost, heat data is used for cooling control, electric power data is used for task scheduling, resource conflicts are avoided, the limitation of traditional single resource scheduling is broken through, and the response speed is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data center interaction, in particular to a data center computing power and power collaborative scheduling information interaction method and system. BACKGROUND

[0002] At present, with the continuous expansion of the scale of data centers, the collaborative demand of computing power scheduling and power management is increasingly urgent, and the traditional independent scheduling mode has been unable to meet the efficient operation demand. The existing system usually adopts the architecture of separating the computing power monitoring module and the power management module.

[0003] However, in the prior art, the data center has high energy consumption and low green electricity proportion, which does not meet the low carbon emission demand, and cannot dynamically match the resource supply of computing power and power, resulting in low resource utilization. SUMMARY

[0004] The purpose of the present application is to provide a data center computing power and power collaborative scheduling information interaction method and system, which aims to solve the technical problems of high energy consumption, low green electricity proportion, and low carbon emission demand in the prior art, which cannot dynamically match the resource supply of computing power and power, resulting in low resource utilization.

[0005] To achieve the above purpose, a data center computing power and power collaborative scheduling information interaction method is adopted, which includes the following steps:

[0006] The power load curve, computing power task type, thermal system state, and carbon emission intensity data of the data center are collected, calculated, and integrated by multi-modal, and a space-time correlation model is constructed;

[0007] A multi-objective optimization algorithm MIP combining improved deep reinforcement learning DRL and mixed integer programming is used to simultaneously optimize the computing power cost, carbon emission data, and thermal data of the data center;

[0008] The power source, computing power task intensity, carbon emission intensity, and thermal system state data of the data center are recorded by a blockchain distributed ledger;

[0009] The data center dynamically adjusts energy efficiency and encourages green electricity consumption according to PUE;

[0010] Through data interaction, the computing power, power, and thermal power of the data center are collaboratively scheduled and coupled optimized.

[0011] The multi-modal computing and integration of the power load curve, computing power task type, thermal system state, and carbon emission intensity data of the data center, and the construction of the space-time correlation model include the following steps:

[0012] The real-time power of the data center, time-of-use electricity price, new energy output proportion, and power grid congestion signal are collected;

[0013] Collect the task priority, calculation accuracy, data dependency and migratability of the data center;

[0014] Collect the inlet and outlet air temperature, cooling water flow, PUE value, meteorological data and waste heat recovery efficiency of the data center;

[0015] Collect the carbon emission factor of the data center regional power grid, green electricity consumption ratio, IT equipment carbon footprint and indirect emission data of the refrigeration system;

[0016] Unify asynchronous data streams to the same time granularity, and use a sliding window mechanism to process delayed data;

[0017] Design a dynamic space-time correlation matrix to quantify the coupling strength between nodes;

[0018] Through a multi-modal feature fusion architecture, the power load curve, computing power task sequence, heat and carbon emission data are fused.

[0019] Among them, the power load curve processes the spatial dependence and temporal dynamics of power data through a spatio-temporal graph convolutional network (ST-GCN), and the computing power task sequence is processed through a Mamba-SSM state space model to model long-range dependencies with linear complexity.

[0020] Among them, the upper algorithm architecture adopts DRL, and the lower algorithm architecture adopts MIP. In DRL, Soft Actor-Critic is used to optimize the dynamic strategy. The PUE real-time value, green electricity available ratio, task queue length, cooling system COP, task migration, cooling pump frequency adjustment and green electricity procurement quantity interact with the environment. The data after each step and environmental interaction are pushed into the experience pool to train the network. Finally, only the recommended action is output, but it is not directly executed by DRL. The upper algorithm architecture DRL is used as the initial solution of the lower algorithm architecture MIP. The lower algorithm architecture MIP processes the mixed integer programming through the Gurobi solver. The objective function is:

[0021]

[0022] In the formula, , , are multi-objective weight coefficients; is the total electricity cost, unit yuan;

[0023] is the period carbon emission, unit kg; is the heat constraint violation, unit ;

[0024] The constraints include power balance and heat constraints;

[0025] Thermal constraints:

[0026]

[0027] wherein, Tin is the temperature of the cabinet inlet, unit ℃;

[0028] Power balance:

[0029]

[0030] wherein, Pgrid is the power purchased from the grid, unit kW; Ppv is the real-time photovoltaic power generation, unit kW; Pbat is the battery charging and discharging power, unit kW; Pit is the total power of IT equipment, unit kW; Pcool is the total power of the cooling system, unit kW.

[0031] Wherein, through the Hyperledger Fabric private chain, the power source, the computing task intensity, the carbon emission intensity, and the thermal system state are uploaded with a frequency of 15 minutes, the green electricity certificate is verified to match the real-time power generation, the computing task intensity is uploaded with a frequency of 5 minutes, the task resource request is verified to be consistent with the scheduling record, the carbon emission intensity is uploaded with a frequency of 60 minutes, the IPCC coefficient is calculated and cross-verified, and the thermal system state is uploaded with a frequency of 1 minute, and the temperature overrun triggers the early warning contract.

[0032] Wherein, the real-time data center power utilization efficiency is calculated as:

[0033]

[0034] wherein, PUE is the real-time power utilization efficiency of the data center; Ptotal is the total input active power of the data center, unit kW; Pit is the total active power of the IT equipment, unit kW, wherein, The UPS, PDU, and chiller unit data are collected through Modbus TCP.

[0035] Wherein, when PUE>1.3, the frequency of the cooling pump is triggered to reduce the current frequency by 10%, the chilled water temperature set value is increased, and the set value is increased by 1℃ each time;

[0036] When PUE<1.2, green electricity is preferentially consumed, and the battery is discharged to 20% SOC.

[0037] Wherein, the green electricity premium = benchmark electricity price x (1-green electricity ratio) x carbon price coefficient; when green electricity is surplus, start the delay tolerant task, and compensate 0.02-0.1 yuan / kWh.

[0038] Wherein, the power and computing power interact: the grid dispatching system sends a price signal and a carbon signal, and the data center feeds back the load reduction and delay tolerant task list;

[0039] Computing power and heat power interact: when the cabinet temperature exceeds the preset value, the scheduler preferentially migrates tasks to low-temperature cabinets;

[0040] Heat and power interact: when the cooling system COP decreases, the load rate of IT equipment is reduced.

[0041] The application also provides an information interaction system for collaborative scheduling of data center computing power and power, according to the information interaction method for collaborative scheduling of data center computing power and power described above, comprising an FPGA, a multi-modal acquisition module, a collaborative optimization module, a blockchain scheduling module and an energy efficiency management module;

[0042] The multi-modal acquisition module is used to acquire computing power, power and heat parameters, and construct a space-time correlation model;

[0043] The collaborative optimization module is used to synchronize the optimization of computing power cost, carbon emission data and heat data of the data center through DRL and MIP;

[0044] The blockchain scheduling module is used to realize the cross-domain scheduling and credibility of the power source, computing power task intensity, carbon emission intensity and heat system state;

[0045] The energy efficiency management module is used to dynamically adjust energy efficiency and encourage green electricity consumption;

[0046] The FPGA is used to process real-time data face tasks, realize instant update of the scheduling algorithm, and is used to connect and integrate the multi-modal acquisition module, the collaborative optimization module, the blockchain scheduling module and the energy efficiency management module.

[0047] The data center computing power and power collaborative scheduling information interaction method and system of the application integrates multi-source heterogeneous data and constructs a space-time correlation model, realizes accurate prediction and dynamic modeling of data center resource demand, and the space-time model can capture the load changes of the data center at different times and spaces, thereby providing a basis for subsequent optimization, using the real-time learning ability of DRL and the discrete decision advantage of MIP, realizing the synchronous optimization of multiple objectives, significantly improving the energy efficiency and economy of the data center, and the multi-objective optimization can realize the coupling of "computing power-power-heat", through the blockchain technology to ensure that the data is tamper-proof, transparent and credible, improve the security and auditability of the system, support carbon footprint tracking and green electricity certification, PUE index dynamic optimization energy efficiency, and directly encourage the use of green electricity, reduce carbon footprint and operating costs, the coupling optimization of computing power, power and heat eliminates the island effect, heat data is used for cooling control, and power data is used for task scheduling, avoiding resource conflicts, breaking through the limitations of traditional single resource scheduling, and the hybrid algorithm of DRL and MIP improves the optimization efficiency of multiple objectives and improves the response speed. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0049] Figure 1 is a step flow chart of a data center computing power and power collaborative scheduling information interaction method of the present application.

[0050] Figure 2 is a method step flow chart of step S1 in the present application. Figure 1

[0051] Figure 3 is a system diagram of a data center computing power and power collaborative scheduling information interaction system of the present application.

[0052] 1-FPGA, 2-multimodal acquisition module, 3-collaborative optimization module, 4-blockchain scheduling module, 5-energy efficiency management module. DETAILED DESCRIPTION

[0053] Please refer to Figure 1 and Figure 2 , the present application provides a data center computing power and power collaborative scheduling information interaction method, comprising the following steps:

[0054] ​S1: Collect, calculate, and integrate the power load curve, computing task type, thermal system state, and carbon emission intensity data of the data center, and build a spatiotemporal correlation model;

[0055] S2: Use a multi-objective optimization algorithm MIP combining an improved deep reinforcement learning DRL and mixed integer programming to simultaneously optimize the computing cost, carbon emission data, and thermal data of the data center;

[0056] S3: Use a blockchain distributed ledger to record the power source, computing task intensity, carbon emission intensity, and thermal system state data of the data center;

[0057] S4: The data center dynamically adjusts energy efficiency and encourages green power consumption according to PUE;

[0058] S5: Through data interaction, realize the coordinated scheduling and coupled optimization of the computing, power, and thermal networks of the data center.

[0059] In this embodiment, by integrating multi-source heterogeneous data and building a spatiotemporal correlation model, precise prediction and dynamic modeling of the resource demand of the data center are realized. The spatiotemporal model can capture the load changes of the data center at different times and spaces, thereby providing a basis for subsequent optimization. The real-time learning ability of DRL and the discrete decision advantage of MIP are used to realize the simultaneous optimization of multiple objectives, significantly improving the energy efficiency and economy of the data center. Multi-objective optimization can realize the coupling of "computing power, power, and heat", and through the use of blockchain technology, the data is ensured to be tamper-proof, transparent, and trustworthy, improving the security and auditability of the system, supporting carbon footprint tracking and green power certification, and directly encouraging the use of green power, reducing carbon footprint and operating costs. Coupled optimization of computing power, power, and heat eliminates the island effect, and thermal data is used for cooling control, while power data is used for task scheduling, avoiding resource conflicts and breaking through the limitations of traditional single resource scheduling. The hybrid algorithm of DRL and MIP improves the optimization efficiency of multiple objectives and improves the response speed.

[0060] Further, the multi-modal calculation and integration of the power load curve, computing task type, thermal system state, and carbon emission intensity data of the data center, and the construction of a spatiotemporal correlation model, includes the following steps:

[0061] S101: Collect real-time power, time-of-use electricity price, new energy output proportion, and power grid congestion signals of the data center;

[0062] S102: Collect the task priority, computing accuracy, data dependency, and migratability of the data center;

[0063] S103: Collect the inlet and outlet air temperature, cooling water flow, PUE value, meteorological data, and waste heat recovery efficiency of the data center;

[0064] S104: Collect data of carbon emission factor of regional power grid, green power consumption ratio, carbon footprint of IT equipment, and indirect emission of refrigeration system;

[0065] S105: Unify asynchronous data streams to the same time granularity, and process delayed data by using a sliding window mechanism;

[0066] S106: Design a dynamic space-time correlation matrix to quantify the coupling strength between nodes;

[0067] S107: Fuse power load curve, computing power task sequence, heat and carbon emission data through a multi-modal feature fusion architecture.

[0068] In the embodiment, by synchronously collecting power, time-of-use electricity price, new energy output ratio, and power grid congestion signal, the power grid state can be dynamically perceived to provide millisecond-level response basis for computing power scheduling. The task priority, data dependency, and migratability can be collected to optimize the task allocation logic. The real-time monitoring of inlet and outlet air temperature, PUE value, and waste heat recovery efficiency, combined with the regional power grid carbon emission factor and green power consumption ratio, can realize the collaborative optimization of energy efficiency and carbon emission. Through real-time, coupling quantization, and heterogeneous fusion, the shortcomings of traditional data center "power-computing-heat" fragmented optimization are solved, and the core scheduling capability is provided for the data center.

[0069] Further, the power load curve processes the spatial dependence and temporal dynamics of power data through a space-time graph convolution network ST-GCN, and the computing power task sequence processes the high-dimensional computing power task sequence through a Mamba-SSM state space model to model long-range dependencies with linear complexity.

[0070] In the embodiment, the fusion architecture of ST-GCN and Mamba-SSM cooperates through the division of space topology modeling and long sequence efficient learning, and the power-computing-heat coupling optimization drives the energy efficiency improvement and carbon emission reduction of the data center.

[0071] Further, the upper algorithm architecture adopts DRL, and the lower algorithm architecture adopts MIP. In the DRL, a SoftActor-Critic is used to optimize the dynamic strategy. The PUE real-time value, green power availability ratio, task queue length, cooling system COP, task migration, cooling pump frequency adjustment, and green power procurement amount interact with the environment. The data after each step of interaction with the environment is pushed into the experience pool to train the network. Finally, only the recommended action is output, but it is not directly executed by the DRL. The upper algorithm architecture DRL serves as the initial solution of the lower algorithm architecture MIP. The lower algorithm architecture MIP processes the mixed integer programming through a Gurobi solver. The objective function is:

[0072]

[0073] wherein, , , are multi-objective weight coefficients; is total electricity cost, unit yuan;

[0074] is period carbon emission, unit kg; is heat constraint violation, unit ;

[0075] The constraints include power balance and heat constraints;

[0076] The heat constraints are:

[0077]

[0078] wherein, is cabinet inlet temperature, unit ℃;

[0079] The power balance is:

[0080]

[0081] wherein, is power purchased from the grid, unit kW; is on-site photovoltaic real-time power generation, unit kW; is battery charging and discharging power, unit kW; is total IT equipment power, unit kW; is total cooling system power, unit kW.

[0082] In the embodiment, dynamic response is achieved, real-time cost reduction and efficiency improvement are achieved, based on real-time data, the DRL algorithm dynamically generates recommended actions for cooling frequency adjustment and task migration, reduces power cost by 12-15%, automatically migrates tasks to green electricity rich nodes during the peak of green electricity output, improves green electricity consumption rate, reduces carbon emission cost, the lower algorithm MIP strictly checks the recommendations of the upper algorithm DRL to avoid device overload and downtime, ensures system safety, synchronously optimizes PUE, carbon emission, and task efficiency, solves the contradiction between "carbon reduction" and "performance preservation", realizes continuous self-evolution, and achieves long-term stable operation.

[0083] Further, through the Hyperledger Fabric private chain, the power source, the computing power task intensity, the carbon emission intensity, and the heat system state are uploaded at a frequency of 15 minutes, the green electricity certificate is verified to match the real-time power generation, the computing power task intensity is uploaded at a frequency of 5 minutes, the task resource request is verified to be consistent with the scheduling record, the carbon emission intensity is uploaded at a frequency of 60 minutes, the IPCC coefficient is calculated and cross-verified, and the heat system state is uploaded at a frequency of 1 minute, and the temperature exceeding the limit triggers a warning contract.

[0084] In this embodiment, the green electricity certificate is verified every 15 minutes to match the actual power generation, false green electricity transactions are prevented, the real transaction volume of green electricity is realized, the task resource application is verified every 5 minutes to match the actual scheduling record, false occupation of computing power resources is prevented, and resource utilization is improved, the IPCC coefficient is automatically called every hour to calculate carbon emissions, cross-verification by multiple agencies is performed to reduce calculation errors, temperature is monitored every minute and early warning is triggered to improve the speed of accident impact and reduce the risk of equipment downtime.

[0085] Further, the real-time calculation data center electric energy utilization efficiency is:

[0086]

[0087] In the formula, is the real-time electric energy utilization efficiency of the data center; is the total input active power of the data center, in kW; is the total active power of the IT equipment, in kW, wherein, The UPS, PDU, and chiller data are collected through Modbus TCP.

[0088] In this embodiment, by calculating the data center electric energy utilization efficiency, the energy efficiency can be accurately monitored and optimized, and for every 0.1 reduction in PUE, the operating cost is reduced, invalid carbon emissions are reduced, green electricity consumption and heat recovery are promoted, low-PUE data centers are more easily matched with green electricity volatility, the consumption rate is improved, and the waste heat recovery efficiency is optimized in conjunction with PUE, which can create secondary benefits.

[0089] Further, when PUE>1.3, the cooling pump frequency is changed to reduce the current frequency by 10%, the chilled water temperature set value is increased, and the set value is increased by 1℃ each time;

[0090] When PUE<1.2, green electricity is preferentially consumed, and the battery is discharged to 20% SOC.

[0091] In this embodiment, the cooling system energy consumption is directly reduced, thereby reducing the overall PUE, avoiding excessive cooling and equipment damage, prolonging the service life of the equipment, and discharging the state of charge of the energy storage battery from a high value to 20%. Discharging the battery to 20% SOC can optimize the utilization rate of energy storage, avoid damage caused by deep discharge of the battery, and at the same time provide peak shaving services for the power grid, use stored green electricity to support IT loads, directly reduce carbon emissions through green electricity consumption, and help achieve the "double carbon" goal. Based on the PUE value, dynamic adjustment is realized to improve energy efficiency and maximize the use of clean energy, forming a virtuous cycle of "energy saving-carbon reduction-compliance".

[0092] Further, the green electricity premium = benchmark electricity price x (1-green electricity ratio) x carbon price coefficient; when there is excess green electricity, start the delay tolerance task and compensate 0.02~0.1 yuan / kWh.

[0093] In the embodiment, the environmental value of green electricity is internalized as economic value through the premium mechanism, and the compensation mechanism can smooth supply and demand fluctuations, significantly reduce the system operation cost of users, accurately quantify the premium of green electricity, improve transaction efficiency, reduce transaction friction cost, improve the liquidity of the green electricity market, effectively promote green electricity consumption, reduce carbon emissions, reduce dependence on fossil energy, help achieve the "double carbon" goal, and improve the flexibility and synergy efficiency of the energy system through the delay tolerant task compensation mechanism.

[0094] Further, power and computing power interact: the grid dispatching system sends price signals and carbon signals, and the data center feeds back a list of tasks that can be reduced and delay-tolerant tasks;

[0095] Computing power and heat interact: when the cabinet temperature exceeds the preset value, the dispatcher migrates tasks to low-temperature cabinets first;

[0096] Heat and power interact: when the cooling system COP decreases, the load rate of IT equipment is reduced.

[0097] In the embodiment, the data center feeds back a list of tasks that can be reduced and a list of delay-tolerant tasks, so that the grid dispatching can dynamically adjust the load according to real-time electricity / carbon price signals; the carbon price signal drives the data center to run high-load tasks during low-carbon periods; when the cabinet temperature exceeds the safety threshold, the dispatcher migrates tasks to low-temperature areas, and task migration reduces local cooling demand; when the cooling system COP is lower than the threshold, the load rate of IT equipment is automatically reduced; the computing power-power-heat of the data center is coupled, and the control variables of the three systems are controlled, so as to optimize and improve energy efficiency; the data center adjusts through delay-tolerant tasks, so as to improve the matching degree of the load curve and the wind-solar output.

[0098] Please refer to Figure 3 The application also provides an information interaction system for collaborative dispatching of data center computing power and power, which comprises an FPGA1, a multi-modal acquisition module2, a collaborative optimization module3, a blockchain dispatching module4, and an energy efficiency management module5 according to the information interaction method for collaborative dispatching of data center computing power and power described above.

[0099] The multi-modal acquisition module2 is used to acquire computing power, power, and heat parameters and construct a space-time correlation model.

[0100] The collaborative optimization module3 is used to synchronize the optimization of the computing power cost, carbon emission data, and heat data of the data center through DRL and MIP.

[0101] The blockchain dispatching module4 is used to realize cross-domain dispatching and credibility of the power source, computing power task intensity, carbon emission intensity, and heat system state.

[0102] The energy efficiency management module 5 is used for dynamically adjusting energy efficiency and encouraging green electricity consumption;

[0103] The FPGA 1 is used for processing real-time data plane tasks, realizing instant update of scheduling algorithms, and connecting and integrating the multi-modal acquisition module 2, the collaborative optimization module 3, the blockchain scheduling module 4 and the energy efficiency management module 5.

[0104] In the embodiment, the FPGA 1 collects cabinet / chip temperature in real time through an on-chip temperature sensor, predicts heat dissipation demand, dynamically calibrates cooling strategies, splits the FPGA 1 into virtual functions by using PCIe SR-IOV technology, monitors the calculation load and delay tolerance of each operator in a fine-grained manner, directly analyzes carbon intensity signals and real-time electricity prices of the power grid, realizes signal decoding and priority sorting through a hardware logic circuit, directly connects sensors and actuators through the logic circuit, eliminates operating system scheduling overhead, realizes 5μs-level closed-loop control, meets high-priority task requirements, and becomes a core engine of computing power-power-heat power collaborative scheduling through the trinity architecture of "hardware reconfigurable x real-time control x energy efficiency closed loop". Through multi-module deep coupling and hardware acceleration innovation, a closed-loop system of "sensing-decision-making-scheduling-execution-verification" is constructed, and breakthrough improvement in economy, environmental protection, reliability and response speed is realized.

[0105] The above only discloses one preferred embodiment of the present application, and of course cannot limit the scope of the rights of the present application. Those skilled in the art can understand that all or part of the above-mentioned embodiments are implemented, and equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.

Claims

1. A method for information interaction of data center computing power and power collaborative scheduling, characterized in that, Comprising the following steps: Through multi-modal acquisition, calculation and integration of data center power load curve, computing power task type, heat system state, carbon emission intensity data, and construction of space-time correlation model, the power load curve is processed by space-time graph convolution network ST-GCN to handle the spatial dependence and temporal dynamics of power data, and the computing power task sequence is processed by Mamba-SSM state space model to handle high-dimensional computing power task sequence to model long-distance dependence with linear complexity; A multi-objective optimization algorithm MIP combining improved deep reinforcement learning DRL and mixed integer programming is used to simultaneously optimize the computing power cost, carbon emission data and heat data of the data center, the upper algorithm architecture adopts DRL, the lower algorithm architecture adopts MIP, Soft Actor-Critic is used in DRL to optimize dynamic strategy, the PUE real-time value, green electricity available proportion, task queue length, cooling system COP, task migration, cooling pump frequency adjustment and green electricity procurement quantity are interacted with the environment, the data after each step and environment interaction is pushed into the experience pool to train the network, and finally only the recommended action is output, but not directly executed by DRL, the upper algorithm architecture DRL is used as the initial solution of the lower algorithm architecture MIP, the lower algorithm architecture MIP processes mixed integer programming through Gurobi solver, and the objective function of the mixed integer programming is: In the formula, , , are multi-objective weight coefficients; is the total power consumption cost, unit yuan; is the period carbon emission, in kg; is the thermal constraint violation, in ; The constraints include power balance and heat constraints; Heat constraints: In the formula, Tin is the temperature of the air entering the cabinet, in °C. Power balance: wherein, Pgrid is the power purchased from the grid, in kW; Ppv is the real-time PV power, in kW; Pbat is the battery charge / discharge power, in kW; Pit is the total IT equipment power, in kW; Pcool is the total cooling system power, in kW; The blockchain is used to record the power source, computing power task intensity, carbon emission intensity and heat system state data of the data center, the blockchain adopts Hyperledger Fabric private chain, the power source is uploaded to the chain at a frequency of 15 minutes, the green electricity certificate is verified to match the real-time power generation, the computing power task intensity is uploaded to the chain at a frequency of 5 minutes, the task resource request is verified to be consistent with the scheduling record, the carbon emission intensity is uploaded to the chain at a frequency of 60 minutes, the IPCC coefficient is calculated and cross-verified, and the heat system state is uploaded to the chain at a frequency of 1 minute, and the temperature overrun triggers the warning contract; The data center adjusts the energy efficiency dynamically and encourages green electricity consumption according to PUE; Through data interaction, the computing power, power and heat of the data center are cooperatively scheduled and coupled optimized.

2. The information interaction method for collaborative scheduling of computing power and power in a data center as described in claim 1, wherein the multimodal calculation integrates the data center's power load curve, computing task type, thermal system status, and carbon emission intensity data, and constructs a spatiotemporal correlation model, characterized in that... Comprising the following steps: Collecting real-time power, time-of-use electricity price, new energy output proportion and power grid congestion signals of the data center; Collecting task priority, calculation accuracy, data dependency and migratability of the data center; Collecting data center inlet and outlet air temperature, cooling water flow, PUE value, meteorological data and waste heat recovery efficiency; Collecting data center regional power grid carbon emission factor, green electricity consumption proportion, IT equipment carbon footprint and indirect emission data of the refrigeration system; Uniformly processing asynchronous data streams to the same time granularity, and using a sliding window mechanism to process delayed data; Designing a dynamic space-time correlation matrix to quantify the coupling strength between nodes; Through a multi-modal feature fusion architecture, the power load curve, computing power task sequence, heat and carbon emission data are fused.

3. The information interaction method for data center computing power and power collaborative scheduling according to claim 1, wherein the data center dynamically adjusts energy efficiency and encourages green power consumption according to PUE. Real-time calculation of data center electric energy utilization efficiency: In the formula, PDC is the real-time power utilization efficiency of the data center; PDC is the total input active power of the data center, in kW; PIT is the total active power of the IT equipment, in kW, wherein, The UPS, PDU, and chiller data are collected through Modbus TCP.

4. The information interaction method for data center computing power and power collaborative scheduling according to claim 3, wherein when PUE>1.3, the frequency of the cooling pump is changed to reduce the current frequency by 10%, and the chilled water temperature setting value is increased by 1℃ each time; when PUE<1.2, green power is preferentially consumed, and the battery is discharged to 20% SOC.

5. The information interaction method for data center computing power and power collaborative scheduling according to claim 3, wherein green power premium=reference price×(1-green power ratio)×carbon price coefficient; when green power is abundant, delay-tolerant tasks are started to compensate 0.02~0.1 yuan / kWh.

6. The information interaction method for data center computing power and power collaborative scheduling according to claim 1, wherein the computing power, power, and heat of the data center are collaboratively scheduled and coupled optimized through data interaction, and the method comprises the following steps: power and computing power interaction: the power grid dispatching system sends price signals and carbon signals, and the data center feeds back the reducible load and the delay-tolerant task list; computing power and heat interaction: when the cabinet temperature exceeds the preset value, the scheduler preferentially migrates tasks to low-temperature cabinets; heat and power interaction: when the cooling system COP decreases, the load rate of the IT equipment is reduced.

7. An information interaction system for data center computing power and power collaborative scheduling, which is used to execute the information interaction method for data center computing power and power collaborative scheduling according to any one of claims 1-6, and comprises the following steps: FPGA, multi-modal acquisition module, collaborative optimization module, blockchain scheduling module, and energy efficiency management module; the multi-modal acquisition module is used to acquire computing power, power, and heat parameters, and construct a space-time correlation model; the collaborative optimization module is used to synchronously optimize the computing power cost, carbon emission data, and heat data of the data center through DRL and MIP; the blockchain scheduling module is used to realize cross-domain scheduling and credibility of the power source, computing power task intensity, carbon emission intensity, and heat system state; the energy efficiency management module is used to dynamically adjust energy efficiency and encourage green power consumption; the FPGA is used to process real-time data plane tasks, realize instant update of the scheduling algorithm, and connect and integrate the multi-modal acquisition module, the collaborative optimization module, the blockchain scheduling module, and the energy efficiency management module. ​ ​

Images