CN-122021235-A - Data center frequency response capability modeling method based on data-driven inverse optimization

Abstract

A data center frequency response modeling method based on data-driven inverse optimization belongs to the technical field of power system dispatching and data center energy management, and comprises the steps of carrying out mathematical modeling on load response behaviors of a data center, constructing comprehensive optimization objective functions and constraint conditions comprising transferable loads, interruptible calculation loads and cooling equipment, aggregating high-dimensional mixed integer constraints of the data center into a low-order model with reduced dimensions based on ALF adjustable load group concepts, constructing a data-driven training framework based on inverse optimization, learning ALF adjustable load group model parameters from historical operation data, and realizing dynamic dispatching and frequency response optimization decision of the data center based on the trained reduced-dimension model. The invention can remarkably reduce the calculation complexity while keeping the key physical constraint, and improves the dispatching efficiency and modeling accuracy of the data center participating in the auxiliary service of the power grid.

Inventors

• SUN RONGFU
• YU ANGUO
• XUE XIAOQIANG
• LU SHIHUA
• MIN RUI
• Wei Chengmei
• Zhao Kangda
• WANG FEI
• YUE HAO
• WANG GE
• LIU JUAN
• ZHANG JIAYI
• XIAO YUNPENG
• ZHU TIANBO
• LIU QINZHE

Assignees

• 国网冀北电力有限公司
• 华北电力大学
• 国网冀北电力有限公司经济技术研究院
• 西安交通大学

Dates

Publication Date

20260512

Application Date

20251205

Claims (10)

1. 1. The modeling method for the frequency response capability of the data center based on the data-driven inverse optimization is characterized by comprising the following steps: The method comprises the steps of 1, carrying out mathematical modeling on load response behaviors of a data center, and constructing comprehensive optimization objective functions and constraint conditions comprising transferable loads, interruptible calculation loads and cooling equipment; step 2, based on an ALF adjustable load group concept, aggregating high-dimensional mixed integer constraints of a data center into a low-order model with reduced dimensions; step 3, constructing a data driving training frame based on inverse optimization, and learning ALF adjustable load group model parameters from historical operation data; And 4, realizing dynamic scheduling and frequency response optimization decision of the data center based on the trained dimension reduction model.
2. 2. The method for modeling a data center frequency response capability according to claim 1, wherein the step 1 comprises: Step 1.1, constructing a comprehensive optimization objective function of a data center Step 1.2, transferable load modeling; Step 1.3, modeling of interruptible computational load; step 1.4, modeling the cooling equipment.
3. 3. Modeling method in accordance with claim 2, characterized in that in step 1.1, the comprehensive optimization objective function is: Wherein, C inv is equipment investment cost, C op is operation and maintenance cost, C energy is energy purchasing cost, Lambda is the carbon emission coefficient, C DR is the demand response benefit, and C Inc is the incentive benefit; In step 1.2, the transferable loads are modeled by a linear programming model: Wherein P n,t is the power of the task n in the period t, x n,t is a binary decision variable, indicating whether the task n can be executed in the period t; The constraint conditions are as follows: i.e. each task can only be performed during one time period; In step 1.3, the model of the interruptible computational load is expressed as: Wherein mu n,t is the recovery cost of the task n in the period t, y n,t is a binary decision variable of whether the task is recovered, and in step 1.4, the cooling equipment power is represented as follows by a regression model: P cooling ＝α·T env +β·L, Wherein T env is the ambient temperature, L is the load, and the alpha and beta regression coefficients characterize the influence of the ambient temperature and the load on the power consumption of the cooling equipment.
4. 4. The method of modeling data center frequency response capability according to claim 1, wherein in step 2, the adjustable load group model includes a plurality of adjustable load units (AL), each AL unit being composed of a plurality of servers grouped by functions, and the basic constraints of the model include: Power balance constraint: device power constraint P i min ≤p t,i ≤P i max ; Daily accumulated energy constraint: Wherein x t is the total power consumption of the system at time t, P t,i is the power of the AL unit i at time t, deltat is the time interval, P i min and P i max are the upper and lower power limits, Is the upper and lower energy limits.
5. 5. The data center frequency response capability modeling method of claim 4, wherein the adjustable load group model sets differential constraints for different types of data centers: AL1 corresponds to a core service data center, has the characteristic of non-migration, and has strict task execution sequence and power consumption following task intensity; AL2 corresponds to the cloud computing platform data center, has time flexible migration characteristics, and tasks can be migrated in a 4-hour window to carry out load balancing; AL3 corresponds to an edge data center, has a remote flexible migration characteristic, and allows power adjustment on the premise of ensuring service quality.
6. 6. Modeling method in accordance with claim 1, characterized in that step 3 comprises: step 3.1, constructing a data set, collecting data of each data center for the past N days to form the data set Pr (n) is the power consumption on the nth day, x (n) is the power adjustment amount, and Task (n) is the Task strength; step 3.2, constructing an inverse optimization objective function: wherein, I, I F is Frobenius norm, and x n is an optimal solution of the ALF model under the nth day condition; Step 3.3, solving by adopting a mixed optimization algorithm based on rule search, wherein the algorithm integrates three strategies of pattern search, grid search and intelligent direction search; And 3.4, adopting a batch training strategy, randomly selecting B data for training each generation, and adopting an exponential weighted moving average updating parameter theta (k+1) ＝(1-α)θ (k) +αθ * , wherein theta * is the optimal parameter of the current generation, and alpha is the learning rate.
7. 7. The method for modeling data center frequency response capability according to claim 6, wherein in step 3.3, the rule search based hybrid optimization algorithm belongs to a derivative-free optimization method, and specifically comprises: Pattern searching is used for systematic exploration along a standard axis; Grid searching is used for local fine searching; Intelligent direction searching predicts a promising searching direction based on historical information; the search efficiency and convergence stability are ensured by adopting self-adaptive step length adjustment and multi-criterion early stop system.
8. 8. The method of modeling data center frequency response capability of claim 1, wherein step 4 comprises: The model application environment deployment, namely deploying the trained model in a data center dispatching system, and acquiring environmental parameters such as load, electricity price, task intensity and the like in real time; Step 4.2, executing the collaborative decision, and generating a power adjustment decision by each data center through an ALF model based on the real-time state; Step 4.3, model iterative optimization, periodically collecting real-time scheduling data to evaluate model performance, and retraining the model based on new data when the environment is significantly changed; And 4.4, outputting the decision optimization result and participating in the frequency response service.
9. 9. The modeling method of claim 1, wherein the ALF model reduces the high-dimensional mixed integer constraint to a low-order linear constraint model comprising 12 consecutive parameters, the 12 consecutive parameters comprising a lower power limit, an upper power limit, a lower energy limit, and an upper energy limit for each of the 3 AL units.
10. 10. A data center frequency response capability modeling apparatus based on data-driven inverse optimization, comprising: the data center load response modeling module is used for carrying out mathematical modeling on the data center load response behavior and constructing a comprehensive optimization objective function and constraint conditions which comprise transferable load, interruptible calculation load and cooling equipment; The adjustable load group dimension reduction module is used for converging the high-dimension mixed integer constraint of the data center into a dimension reduction low-order model based on an adjustable load group concept; The inverse optimization training module is used for constructing a data driving training frame based on inverse optimization and learning ALF model parameters from historical operation data; And the dynamic scheduling and optimizing decision module is used for realizing dynamic scheduling and frequency response optimizing decision of the data center based on the trained dimension reduction model.

Description

Data center frequency response capability modeling method based on data-driven inverse optimization Technical Field The invention belongs to the technical field of power system dispatching and data center energy management, relates to a load modeling and optimizing dispatching method for a data center participating in power grid auxiliary service, and particularly relates to a data center frequency response capacity modeling method based on data driving inverse optimization. Background With the rapid development of a novel power system, renewable energy sources such as wind power, photovoltaic and the like are connected in a large scale, the intermittence and fluctuation of the output of the renewable energy sources bring challenges to the safe and stable operation of a power grid, and a traditional 'source follow-up' scheduling mode faces huge pressure. Under the background, the excavation of flexible resources on the demand side to participate in power grid regulation becomes an important way for guaranteeing the supply and demand balance of a power system. Data centers have been the core infrastructure in the information age, with electricity usage accounting for 1% -2% of the total electricity usage worldwide, and continue to grow at a rate of 10% -15% per year. The data center load has obvious flexibility potential through deep analysis of the operation characteristics, namely the time shifting characteristic of the IT load, the occupation ratio of batch processing tasks (such as data backup and model training) can reach 30% -50%, the batch processing tasks can be flexibly scheduled in a wider time window, the thermal inertia of a cooling system, the cooling load approximately accounts for 30% -40% of total energy consumption, the building heat capacity can be utilized to operate in advance or after, the energy storage characteristic of a standby power supply, the UPS storage battery pack and an independent energy storage system can participate in power grid peak regulation and frequency modulation, and the geographic distribution load balancing is realized, so that the cost optimization is realized through the dynamic migration of the working load among different data centers. Based on the above features, data centers have been identified as potentially flexible resources that can participate in ancillary services such as frequency adjustment. A great deal of research has explored the possibility of data centers participating in the auxiliary services market. Representative work has proposed a frequency-tuned optimization model that utilizes IT load flexibility and integrates energy storage, a mixed integer programming model is widely adopted to capture workload migration and logic constraints, and a double-layer optimization model can synergistically optimize flexible load and equipment scheduling. However, these refined modeling methods require the acquisition of detailed equipment parameters inside the data center, involve a large number of integer variables and nonlinear constraints, result in heavy computational burden and limited scalability of real-time scheduling. Another research direction focuses on energy consumption modeling, describing the coupling relationship of IT equipment, cooling systems and the environment. The modular simulation framework, rack-level airflow and cooling models, and the temperature and humidity considered polynomial IT power formulas all exhibit value, but are heavily dependent on equipment-level data. Data center operators are generally reluctant to disclose internal parameters for business confidentiality and data security considerations, which presents a hurdle to system operators. To improve the handleability, the dimension reduction and agent model is widely adopted. The Virtual Battery (VB) model approximates flexibility by presetting a prototype feasible region, including external approximation of temperature control load flexibility, multi-battery approximation of an electric vehicle fleet and the like. These approaches reduce complexity but may overcomplete operating constraints and impair interpretability, limiting straightforward application in system level scheduling. The shortcomings of the prior art scheme are mainly manifested in the following aspects of high complexity of refined modeling, namely that the traditional method needs to acquire detailed parameters such as a server start-stop sequence, a task dependency relationship, network bandwidth constraint and the like, and the parameters are difficult to acquire and relate to the data privacy problem. Meanwhile, the high-dimensional mixed integer constraint greatly increases the solving difficulty of the scheduling problem, and the timeliness requirement of real-time scheduling is difficult to meet. The accuracy of the proxy model is insufficient, namely, the prototype feasible region preset by the proxy model such as a virtual battery and the like can excessively simplify the actual operation con