Digital native model energy prediction method and system of physically-driven computing power center

Abstract

The invention discloses a digital native model energy prediction method and system of a physically-driven computing power center. The method comprises the steps of obtaining time sequence mass flow, temperature data and geometric physical parameters required by operation of the cold storage tank, and setting prior information such as an initial to-be-identified parameter set and a flow direction to obtain input data and parameters; on the basis of energy conservation, constructing a layered ordinary differential equation model for describing the temperature change of each layer by using the input data and the parameters; using an adaptive step ODE solver to predict temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model according to input data and parameters so as to obtain predicted temperature; constructing a loss function by comparing the predicted temperature with an actual measurement value; a gradient is calculated based on the loss function, and parameters are adjusted by an optimization algorithm to minimize the loss function. By implementing the method provided by the invention, the model dimension can be reduced and the calculation can be simplified while the main physical mechanism is reserved.

Description

TECHNICAL FIELD

[0001] The present application relates to digital twinning, and more particularly to a digital native model energy prediction method and system for a physically driven computing power center. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, high-performance computing facilities represented by AIDC (Artificial Intelligence Data Center) are expanding in scale. These data centers are characterized by continuous increase in rack power density and overall installed capacity, which puts higher demands on the cooling system: high cooling capacity, high reliability, and low energy consumption. However, traditional data center cooling modes, such as direct cooling methods like water chillers, cooling towers, end air conditioners, or liquid cooling, are difficult to meet the needs of peak load management, safety redundancy guarantee, and life cycle cost optimization at the same time when facing new constraints such as time-of-use electricity prices, renewable energy access, and grid demand response.

[0003] As a key equipment, the cold storage tank stores cold energy during periods of low electricity prices or high energy efficiency, and releases it during peak or failure periods, achieving time dimension peak load shifting of power load. This not only helps to reduce the pressure on the cooling machine and power distribution system, improves the flexibility of the cooling system, but also improves the proportion of local consumption of renewable energy. Therefore, in the AIDC scenario, the cold storage tank has become an indispensable important infrastructure on the cold source side. Although the application of cold storage tanks is widespread, in existing engineering practice, the capacity selection and operation strategy of cold storage tanks still mainly rely on experience and simple rules, and lack the ability to accurately depict the charging and discharging dynamic process under complex working conditions. Most methods only use rough static energy balance or black box data fitting, which cannot accurately assess the impact of these processes on unit efficiency, PUE (Power Usage Effectiveness), and cooling safety. In addition, there is currently a lack of high-fidelity, differentiable cold storage tank models and simulation tools that are deeply integrated with load forecasting, electricity / carbon price forecasting, optimal scheduling, and model predictive control algorithms, limiting the collaborative optimization potential between cold storage and chillers, new energy, and electricity prices, hindering the further cost reduction and efficiency improvement of AIDC energy systems and low-carbon transformation.

[0004] The closest prior art solution is based on the PDE (Partial Differential Equation) based hierarchical tank physics modeling method. This method regards the cold storage tank as a one-dimensional continuous medium, uses the temperature distribution T(z, t) as the state variable, and establishes the PDE control equation by combining the factors such as the convection formed by the inflow and outflow water, the heat conduction between adjacent water layers, and the heat exchange between the tank wall and the environment, to describe the change of the temperature field in the cold storage tank with time and space. However, this method has the following disadvantages: since the model is in the form of distributed parameters, a large number of spatial nodes need to be introduced after discretization, forming a high-dimensional state and a large-scale algebraic equation system, the calculation amount of a single simulation is huge, and it is difficult to meet the real-time requirements of long-time, multi-condition, high-frequency simulation in the AIDC scenario. The PDE model contains many local parameters such as interlayer thermal conductivity, mixing parameters, and tank wall heat exchange coefficient, which are mostly dependent on empirical settings and are difficult to effectively identify and continuously correct under the condition of limited sensors, resulting in poor adaptability of the model to actual working condition changes. The discrete solving process of PDE is not conducive to efficient gradient calculation in the automatic differentiation framework, limiting the end-to-end coupling ability with load prediction, electricity / carbon price prediction, optimization scheduling and model predictive control algorithms.

[0005] Therefore, it is necessary to design a new method that retains the main physical mechanism while reducing the model dimension and simplifying the calculation, so that it can not only efficiently simulate the charging and discharging process of the cold storage tank, but also be easily integrated with load prediction and energy scheduling optimization algorithms for use, thereby realizing more efficient energy management and optimization. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art and provide a physically driven digital native model energy prediction method and system for an algorithm center.

[0007] To achieve the above purpose, the following technical solutions are adopted: a physically driven digital native model energy prediction method for an algorithm center, comprising: Obtain time series mass flow, temperature data and geometric physical parameters required for the operation of the cold storage tank, and set initial to-be-identified parameter sets and prior information such as flow direction to obtain input data and parameters; Based on energy conservation, construct a hierarchical ordinary differential equation model describing the temperature change of each layer using the input data and parameters; Use an adaptive step ODE solver to predict the temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model according to the input data and parameters to obtain the predicted temperature; Construct a loss function by comparing the predicted temperature with the actual measured value; Calculate the gradient based on the loss function, and adjust the parameters through an optimization algorithm to minimize the loss function.

[0008] The further technical solution is as follows: the temperature data includes the inlet water temperature, the ambient temperature, and the observed temperature of each layer inside the cold storage tank at each time; the geometric and physical parameters include the volume, height, cross-sectional area, density, and specific heat of each layer.

[0009] The further technical solution is as follows: the initial set of parameters to be identified includes layer-specific response factor, thermal inertia factor, interlayer thermal conductivity, and tank environment heat transfer coefficient.

[0010] Its further technical solution is: the layer-specific response factor satisfy The thermal inertia factor satisfy Interlayer thermal conductivity satisfy Tank environment heat transfer coefficient satisfy .

[0011] The further technical solution is as follows: Based on energy conservation, a layered ordinary differential equation model describing the temperature changes of each layer is constructed using the input data and parameters, including: The upstream temperature is determined based on the time series mass flow rate and its direction in order to determine the temperature effect. The contribution of interlayer heat conduction to temperature change is assessed by the second-order difference of the temperature difference between adjacent layers and the interlayer thermal conductivity. The influence of environmental heat transfer on the layer temperature is calculated by using the side wall area of ​​each layer, the temperature difference of the environment, and the heat transfer coefficient of the tank environment. By integrating the effects of temperature, the contribution of interlayer heat conduction to temperature change, and the influence of environmental heat transfer on layer temperature, an ordinary differential equation describing the rate of temperature change over time is formed. The corresponding ordinary differential equations are established for each layer of water in the cold storage tank, and then combined into a whole layered ODE model to obtain the layered ordinary differential equation model.

[0012] The further technical solution is as follows: the ordinary differential equation is expressed as: ,in, Let be the upstream temperature of the i-th layer; ; This refers to the interlayer thermal conductivity. This represents the area of ​​the sidewalls on each floor. Given the ambient temperature difference; assume the cold storage tank is divided along its height into... The first water layer, the... Layer temperature is The layer volume is The effective heat capacity is ;in This is the thermal inertia factor of the layer. For the density of water, It is a specific heat at constant pressure.

[0013] The further technical solution is as follows: the adaptive step-size ODE solver predicts the temperature distribution of each layer at each time point based on the input data and parameters according to the hierarchical ordinary differential equation model, so as to obtain the predicted temperature, including: The adaptive step-size ODE solver is invoked, and the layered ODE is integrated from the initial time to each observation time based on the input data and parameters according to the layered ordinary differential equation model, so as to obtain the predicted temperature of each layer at each observation time.

[0014] The further technical solution is as follows: the loss function includes ,in, The number of observation times. For the number of layers, in Overlay regularization terms that affect parameter range or smoothness. The set of physical parameters to be identified.

[0015] A further technical solution is as follows: the step of calculating the gradient based on the loss function and adjusting the parameters through an optimization algorithm to minimize the loss function includes: Based on the loss function, the hierarchical ordinary differential equation model is backpropagated using automatic differentiation or an adjoint extension system to calculate the gradient of the loss with respect to each physical parameter, and the parameters are updated according to a preset optimization algorithm.

[0016] This invention also provides a digital-native model energy prediction system for a physically driven computing center, comprising: The acquisition unit is used to acquire time-series mass flow rate, temperature data and geometric physical parameters required for the operation of the cold storage tank, and to set the initial set of parameters to be identified and prior information such as flow direction in order to obtain input data and parameters. The model building unit is used to construct a layered ordinary differential equation model describing the temperature changes of each layer based on the energy conservation principle and using the input data and parameters. The prediction unit is used to predict the temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model using an adaptive step-size ODE solver, according to the input data and parameters, so as to obtain the predicted temperature. The function building unit is used to construct a loss function by comparing the predicted temperature with the actual measured value; The tuning unit is used to calculate the gradient based on the loss function and adjust the parameters through an optimization algorithm to minimize the loss function.

[0017] The advantages of this invention compared to existing technologies are as follows: By acquiring operational data and geometric physical parameters, this invention constructs a low-dimensional ordinary differential equation model along spatial discreteness, requiring only numerical integration along the time dimension, significantly reducing computational complexity. It uses an adaptive step-size ODE solver for efficient temperature prediction, and simultaneously utilizes the difference between actual and predicted values ​​to construct a loss function, achieving accurate calibration by adjusting model parameters through optimization algorithms. This method, while preserving the main physical mechanisms of the cold storage tank's charging and discharging process, significantly reduces model complexity and simplifies the computational process. This allows the model to not only quickly and accurately simulate the behavior of the cold storage tank but also easily integrate with load forecasting and energy dispatch optimization algorithms, thereby effectively supporting efficient energy management and optimization strategies. This method overcomes the problems of high computational overhead and difficulty in coupling with upper-level algorithms in traditional high-dimensional PDE models, providing a more flexible and accurate solution.

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the digital native model energy prediction method for a physical-driven computing center provided in this embodiment of the invention; Figure 2 A schematic diagram of a sub-process of the digital native model energy prediction method for a physical-driven computing center provided in an embodiment of the present invention; Figure 3 A schematic block diagram of a digital native model energy prediction system for a physically driven computing center provided in an embodiment of the present invention; Figure 4 A schematic block diagram of a computer device provided in an embodiment of the present invention; Figure 5 A schematic diagram of the ODE model for the digital native model energy prediction method of the physical-driven computing center provided in this embodiment of the invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] Please see Figure 1 , Figure 1 This is a flowchart illustrating the digital native model energy prediction method for a physically driven computing center provided in this embodiment of the invention. Additionally, please refer to... Figure 5 This physics-driven computing center's digital-native model energy prediction method is applied to servers. By integrating time-series mass flow rate, temperature data, and geometric parameters, and utilizing the law of energy conservation to construct an ODE model describing temperature changes at each layer, it employs an adaptive step-size ODE (Ordinary Differential Equation) solver to predict temperature distribution. Simultaneously, it incorporates actual measurements to construct a loss function to optimize parameters. This method, while preserving the core physical mechanisms, reduces the number and range of parameters to be identified and simplifies computational complexity by introducing a thermal inertia factor, thereby reducing model dimensionality. As a result, the model can efficiently simulate the charging and discharging process of cold storage tanks and is easily integrated with load forecasting and energy dispatch optimization algorithms to achieve more efficient energy management and optimization, providing strong technical support for intelligent buildings and district cooling systems.

[0026] Specifically, under the premise of ensuring energy conservation and strictly satisfying the main physical mechanisms (convection, inlet mixing, interlayer heat transfer, and environmental heat exchange), a continuous-time, differentiable, low-dimensional, efficient, and parameter-interpretable hierarchical ordinary differential equation model suitable for cold storage tanks in computing centers is constructed. This model can be integrated with load forecasting and energy dispatch optimization algorithms for high-precision prediction of cold storage behavior and digital-native simulation. This model can be widely applied to the design, selection, operation optimization, and energy consumption assessment of cold storage systems on the cold source side of large-scale computing centers such as artificial intelligence data centers, providing support for peak shaving and valley filling, improving the efficiency of cooling systems, and ensuring the safe and stable operation of computing infrastructure.

[0027] Figure 1 This is a flowchart illustrating the digital native model energy prediction method for a physically driven computing center provided in this embodiment of the invention. Figure 1 As shown, the method includes the following steps S110 to S150.

[0028] S110. Obtain the time series mass flow rate, temperature data and geometric physical parameters required for the operation of the cold storage tank, and set the initial set of parameters to be identified and prior information such as flow direction to obtain input data and parameters.

[0029] In this embodiment, input data refers to a series of actual operational data that need to be collected and used when implementing the method of this embodiment.

[0030] Parameters refer to the fundamental numerical values ​​used to build and calibrate models. They can be divided into two categories: geometric and physical parameters and the initial set of parameters to be identified.

[0031] Time series quality flow rate The temperature data includes the inlet water temperature. Ambient temperature and the observed stratified temperatures inside the tank at various times. The geometric and physical parameters include the volume of each layer. Floor height Cross-sectional area The density of water Specific heat .

[0032] The initial set of parameters to be identified includes layer-specific response factors, thermal inertia factors, interlayer thermal conductivity, and tank environment heat transfer coefficients. The initial set of parameters to be identified is then set. And prior information such as flow direction.

[0033] The layer-specific response factor satisfy The thermal inertia factor satisfy Interlayer thermal conductivity satisfy Tank environment heat transfer coefficient satisfy .

[0034] In this embodiment, a layer temperature vector is set. Constructing initial values The convection term switches upstream temperature according to the flow direction; the remaining terms act according to the interlayer temperature difference and the ambient temperature difference; inlet mixing occurs only at the front. Once the layer takes effect, the weight calculation formula is as follows: ;in For the first Distance from floor to entrance (in terms of floor number) For mixed strength, This is the attenuation coefficient.

[0035] Parameters and constraints: Set layer-specific response factors for each layer. Used to correct effective mass flow; thermal inertia factor This uniform effect, acting on the denominators of the convection, heat conduction, and ambient heat transfer terms, is equivalent to correcting the effective heat capacity of this layer to... This allows for a uniform adjustment of the layer's response speed to all heat flux terms; and introduces the interlayer thermal conductivity. Heat transfer coefficient with environment Describes interlayer heat transfer and heat dissipation to the environment. Parameter values ​​are limited by the following physical constraints: layer-specific response factor. satisfy Used to limit the amplification factor of the effective mass flow in each layer; thermal inertia factor. satisfy This is used to limit the scaling factor of the effective heat capacity of each layer; interlayer thermal conductivity. satisfy Tank environment heat transfer coefficient satisfy .

[0036] The above value range is derived from engineering experience with cold storage tanks to ensure numerical stability and physical consistency of the model.

[0037] S120. Based on energy conservation, construct a layered ordinary differential equation model describing the temperature changes of each layer using the input data and parameters.

[0038] In this embodiment, the layered ordinary differential equation model refers to establishing a mathematical model that can accurately describe the temperature change of different layers of water in the cold storage tank over time, based on the actual operating data and geometric physical parameters of the cold storage tank.

[0039] In one embodiment, please refer to Figure 2 The above-mentioned step S120 may include steps S121 to S125.

[0040] S121. Determine the upstream temperature based on the time series mass flow rate and its direction to determine the temperature effect.

[0041] In this embodiment, the mass flow rate through the cold storage tank is identified. And its flow direction, and based on this, select the upstream temperature of the i-th layer. This determines the current temperature. A key factor.

[0042] By analyzing real-time recorded mass flow rate data, the direction of water flow can be determined, thereby selecting the correct upstream temperature as the influencing factor on the flow direction.

[0043] S122. The contribution of interlayer heat conduction to temperature change is evaluated by the second-order difference of the temperature difference between adjacent layers and the interlayer thermal conductivity.

[0044] In this embodiment, the heat transfer effect caused by temperature difference between layers is quantized; a discrete second-order difference is constructed using the temperature difference between adjacent layers. ,use Describe the effective heat conduction between layers and calculate the heat conduction contribution of each layer.

[0045] S123. Calculate the impact of environmental heat transfer on the layer temperature using the side wall area of ​​each layer, the temperature difference of the environment, and the heat transfer coefficient of the tank environment.

[0046] In this embodiment, the influence of the external environment on the internal temperature of the cold storage tank is considered.

[0047] Based on the side wall area of ​​each floor Ambient temperature difference and the heat transfer coefficient of the tank environment To calculate the heat exchange between each floor and the environment.

[0048] S124. Integrating the effects of temperature, the contribution of interlayer heat conduction to temperature change, and the influence of environmental heat transfer on layer temperature, an ordinary differential equation describing the rate of temperature change over time is formed.

[0049] By combining all the above influencing factors, a complete mathematical expression is formed to describe the rate of change of temperature in each water layer.

[0050] Integrating the contributions from convection, inlet mixing, interlayer heat transfer, and ambient heat transfer, the final ordinary differential equation is as follows: ,in, Let be the upstream temperature of the i-th layer; ; This refers to the interlayer thermal conductivity. This represents the area of ​​the sidewalls on each floor. This is due to the difference in ambient temperature.

[0051] S125. Establish the corresponding ordinary differential equations for all water layers in the cold storage tank, and combine them into a whole layered ODE model to obtain the layered ordinary differential equation model.

[0052] In this embodiment, the cold storage tank is divided along the height direction into... The first water layer, the... Layer temperature is The layer volume is The effective heat capacity is ;in This is the thermal inertia factor of the layer. For the density of water, It is a specific heat at constant pressure.

[0053] Specifically, for each layer inside the cold storage tank, corresponding ordinary differential equations are established, and these equations are combined to form a unified overall model.

[0054] In this embodiment, the actual water temperature inside the cold storage tank exhibits stratification, and the modeling in this embodiment closely reflects reality. For the N water layers divided along the height of the cold storage tank, ordinary differential equations as shown above are established for each layer, and these equations are combined into a system, thus obtaining the stratified ordinary differential equation model of the entire cold storage tank. The right-hand side of this model is continuously differentiable with respect to the state and parameters, making it naturally suitable for automatic differentiation and gradient-based training, prediction, and control.

[0055] Through the above steps, this embodiment successfully constructed a hierarchical ordinary differential equation model that can efficiently simulate the charging and discharging process of a cold storage tank and is easy to integrate with other prediction and optimization algorithms. This not only significantly reduces the model's dimensionality and computational complexity but also improves the model's fitting accuracy and robustness to actual operating conditions, making it an indispensable part of energy management in computing centers.

[0056] S130. Using an adaptive step-size ODE solver, the temperature distribution of each layer at each time point is predicted based on the input data and parameters according to the hierarchical ordinary differential equation model, so as to obtain the predicted temperature.

[0057] In this embodiment, temperature prediction refers to obtaining the temperature distribution of each layer of the cold storage tank at each observation time by calling the adaptive step-size ODE solver to integrate the hierarchical ordinary differential equation model based on the input data and parameters.

[0058] In one embodiment, an adaptive step-size ODE solver is invoked to integrate the layered ODE from the initial time to each observation time based on the input data and parameters according to the layered ordinary differential equation model, so as to obtain the predicted temperature of each layer at each observation time.

[0059] Given an initial temperature distribution and input sequence In this case, the adaptive step-size ODE solver is invoked to solve the layered ODE from the initial time. Integrating to each observation time This allows us to obtain the predicted temperature for each observation time and each stratum. This step achieves the effect of the current parameters. The physical simulation of the charging and discharging process of the cold storage tank is used to form a differentiable forward model for subsequent gradient calculation and optimization scheduling.

[0060] S140. Construct a loss function by comparing the predicted temperature with the actual measured value.

[0061] In this embodiment, the loss function includes ,in, The number of observation times. For the number of layers, in The regularization terms for parameter range or smoothness are superimposed.

[0062] Specifically, at observation points with measured data The predicted temperature obtained from the forward solution is compared with the corresponding measured temperature to construct a loss function. The mean absolute error (MAE) is preferably used, in the form: ,in, The number of observation times. For the number of floors, it can also be... The superimposed regularization terms on the parameter range or smoothness provide constraints for the automatic calibration process. The set of physical parameters to be identified.

[0063] S150. Calculate the gradient based on the loss function, and adjust the parameters using an optimization algorithm to minimize the loss function.

[0064] In this embodiment, based on the loss function, the hierarchical ordinary differential equation model is backpropagated using automatic differentiation or an adjoint extension system to calculate the gradient of the loss with respect to each physical parameter, and the parameters are updated according to a preset optimization algorithm.

[0065] In the loss function Based on this, backpropagation calculations are performed on the hierarchical ODE using automatic differentiation or an adjoint extension system to calculate the gradient of the loss with respect to each physical parameter. Then, the parameters are updated according to the preset optimization algorithm (such as gradient descent or adaptive learning rate algorithm): ;in, The learning rate is used. If necessary, the updated parameters are projected or clipped to keep them within a pre-defined, physically reasonable range.

[0066] This invention can be implemented by using JAX for automatic differentiation of the discrete-time step equations and Adam for parameter optimization. The resulting parameter set is obtained after model convergence. Can be used in the prediction phase By iteratively executing the forward prediction process (steps S110-S130) and the parameter calibration process (steps S130-S150) within the same differentiable hierarchical model framework, this invention creates a continuous closed loop of "prediction-calibration-re-prediction-re-calibration" in the time dimension. On the one hand, given control inputs such as unit output, inlet and outlet water temperatures, and environmental conditions, steps S110-S130 can perform high-precision dynamic predictions of temperature changes, available cooling capacity, and overall cooling state of each layer of the cold storage tank under future operating conditions, providing a complete state trajectory for subsequent optimization and control. On the other hand, steps S130-S150 construct a loss function based on measurement data and automatically adjust physical parameters, including layer-specific response factors, thermal inertia factors, interlayer thermal conductivity, and tank environment heat transfer coefficients, so that the state evolution equation and parameter update rules form a unified iterative mechanism. This allows the model to gradually approximate the dynamic characteristics of the real cold storage device while continuously absorbing new data, and obtains good generalization ability and adaptive convergence characteristics while maintaining energy conservation and the main physical mechanisms. Based on this unified differentiable structure, the response of the cold storage tank to any control input sequence can be transformed into a differentiable objective function and constraints in a high-level optimization and control problem, and its gradient can be efficiently calculated through automatic differentiation or adjoint methods. This enables key performance indicators such as peak shaving and valley filling, unit load smoothness, cooling redundancy, and cooling capacity utilization efficiency to be collaboratively implemented with algorithms such as cooling load forecasting, energy optimization scheduling, and model predictive control in the computing center within the same solution framework, achieving deep integration and end-to-end joint optimization of "physical model + data-driven".

[0067] In this embodiment, a loss function (e.g., layer-by-layer, time-weighted mean absolute error (MAE)) is introduced based on the differentiable ordinary differential equation (ODE) forward operator for the stratified temperature observation data of the cold storage tank. The gradient of this loss function with respect to each physical parameter is calculated through automatic differentiation or adjoint sensitivity methods. This method has a constant memory complexity of O(1) and does not require storing the entire forward trajectory. Combined with gradient descent or adaptive learning rate optimization algorithms, this invention can periodically or online update key physical parameters, including layer-specific response factors, thermal inertia factors, interlayer thermal conductivity, and tank environment heat transfer coefficients, achieving automatic model calibration within preset physical constraints and ensuring consistency with the real device during long-term operation and changes in operating conditions.

[0068] The cold storage tank is modeled as a set of continuous-time ordinary differential equations, which simultaneously include convection, inlet mixing, interlayer heat transfer, and ambient heat transfer terms within the same framework. By integrating this model using an adaptive step-size ODE solver, the charging and discharging process of the cold storage tank can be regarded as a differentiable forward operator. This allows for the direct insertion of loss functions at observation points with unequal time intervals, providing fundamental support for subsequent gradient-based optimization and control. The differentiable cold storage tank ODE model, as a cold source-side operator, is integrated and coupled with higher-level algorithms in AIDC such as IT load forecasting, electricity / carbon price forecasting, and energy dispatch optimization. By utilizing the model's forward integration to output the stratified temperature distribution, available cooling capacity, or cold storage status indicators for future periods, it provides accurate and differentiable physical constraints and predictive capabilities for peak shaving and valley filling strategies, emergency cooling supply assurance, and overall energy consumption / carbon emission optimization.

[0069] This embodiment, while preserving the main physical mechanisms of the cold storage tank, transforms the one-dimensional convection-diffusion PDE model into a lower-dimensional system of ordinary differential equations through spatial discretization. Under an automatic differentiation framework, parameters with physical meaning are uniformly calibrated, enabling the resulting model to efficiently simulate the charging and discharging process of the cold storage tank and facilitate integration with load forecasting and energy dispatch optimization algorithms. This invention significantly reduces model dimensionality and computational complexity, supporting rapid simulation and batch scenario calculations. By rationally designing an interpretable, physically constrained parameter structure and combining historical operating data for parameter identification and continuous optimization under an automatic differentiation framework, the model's fitting accuracy and robustness to actual operating conditions are improved. Simultaneously, the constructed ODE model is deeply integrated with algorithms for cold load forecasting, electricity pricing strategies and dispatch optimization, and model predictive control, achieving precise prediction and collaborative optimization of the cold storage tank's charging and discharging behavior. This better supports peak shaving and valley filling, cold supply security, and energy conservation and carbon reduction goals for computing centers.

[0070] For the cold storage tank in the cold source system of a computing center, a hierarchical ordinary differential equation (ODE) model is constructed. This model divides the cold storage tank into several layers along its height, and describes the temperature changes of the water in each layer in continuous state variables at each time. It comprehensively considers physical phenomena such as convection and transfer caused by inlet flow rate and temperature, hot and cold water mixing, interlayer heat conduction, and heat exchange between the tank wall and the environment. By uniformly expressing the above physical mechanisms as a set of continuously differentiable ODE equations for both state and parameters, and combining automatic differentiation and optimization algorithms, key parameters can be automatically calibrated while ensuring physical rationality, thereby achieving high-precision prediction of the cold storage process. The parameter set... It has a clear physical meaning and is constrained by a preset interval to ensure numerical stability and energy conservation. For observation data with unequal intervals and asynchronous timing, the model output is aligned with the actual timestamp and layer during training. An adaptive step-size ODE solver combined with the adjoint sensitivity method is used to achieve end-to-end differentiability optimization.

[0071] Existing PDE schemes require spatial subdivision to form high-dimensional states and large-scale equation systems, resulting in high overhead for a single simulation. This invention uses a layered ODE to reduce the order of the cold storage tank model, using a finite number of layer temperature states to simultaneously characterize the main mechanisms such as convection, inlet mixing, interlayer heat transfer, and environmental heat exchange, significantly reducing computational complexity. It is more suitable for long-term, multi-condition, high-frequency AIDC simulations and online optimization, and its engineering implementation and deployment are also simpler.

[0072] The discrete solution process of PDEs is difficult to efficiently calculate gradients within an automatic differentiation framework, limiting its coupling with upper-level algorithms. This invention models the cold storage tank as a continuous-time hierarchical ODE, with the right-hand side function being continuously differentiable with respect to the state and parameters. This function can be directly embedded as a "differentiable forward operator" into the automatic differentiation framework, facilitating the establishment of end-to-end gradient propagation links with algorithms such as cold load forecasting, electricity / carbon price forecasting, model predictive control, joint scheduling optimization, and reinforcement learning.

[0073] PDE schemes typically contain a large number of distributed local parameters, which are difficult to identify holistically under limited measurement points, often relying on one-time empirical tuning. This invention designs finite-dimensional, learnable parameters with clear physical meanings (such as layer-specific response factors, thermal inertia factors, interlayer thermal conductivity, and tank environment heat transfer coefficients) for convection, inlet mixing, interlayer heat transfer, and environmental heat transfer. It also applies physical constraints such as nonnegativity, upper and lower bounds, and monotonicity, and combines automatic differentiation or adjoint methods for periodic or online parameter updates to achieve automatic calibration based on measured data. This allows the model to maintain long-term fit with the real device despite equipment aging and changes in operating conditions.

[0074] PDE models are mostly used for offline fine simulations and are difficult to integrate directly into the integrated business and energy decision-making of computing centers. The ODE model of this invention has a compact structure and high solution efficiency. It can be coordinated with business-side signals such as IT load forecasting, renewable energy output, and electricity / carbon prices to output indicators such as future stratified temperature and available cooling capacity (cold storage state). It provides real-time available and differentiable physical constraints for peak shaving and valley filling strategy formulation, emergency cooling supply guarantee, and energy consumption and carbon emission optimization, significantly improving the value of the model in the practical application of AIDC.

[0075] In one embodiment, the high-dimensional mechanism model based on PDE is continued and discretely solved using numerical difference; in addition, a completely black-box data-driven model (such as a deep neural network) can be used to directly fit the input-output relationship of the cold storage tank.

[0076] The aforementioned physical-driven computing center's digital-native model energy prediction method acquires operational data and geometric physical parameters, discretizes along space to construct a low-dimensional ordinary differential equation model, and only requires numerical integration along the time dimension, significantly reducing computational complexity. It uses an adaptive step-size ODE solver for efficient temperature prediction, and constructs a loss function based on the difference between actual and predicted values, achieving accurate calibration by adjusting model parameters through optimization algorithms. This method retains the main physical mechanisms of the cold storage tank's charging and discharging process while significantly reducing model complexity and simplifying the computational flow. This allows the model to not only quickly and accurately simulate the behavior of the cold storage tank but also easily integrate with load forecasting and energy dispatch optimization algorithms, effectively supporting efficient energy management and optimization strategies. This method overcomes the problems of high computational cost and difficulty in coupling with upper-level algorithms in traditional high-dimensional PDE models, providing a more flexible and accurate solution.

[0077] Figure 3 This is a schematic block diagram of a digital native model energy prediction system 300 with a physically driven computing center, provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above-described energy prediction method for digital native models of physically driven computing centers, this invention also provides an energy prediction system 300 for digital native models of physically driven computing centers. This energy prediction system 300 includes a unit for executing the above-described energy prediction method for digital native models of physically driven computing centers, and the system can be configured in a server. Specifically, please refer to... Figure 3 The digital native model energy prediction system 300 of the physical-driven computing center includes an acquisition unit 301, a model building unit 302, a prediction unit 303, a function building unit 304, and an optimization unit 305.

[0078] The acquisition unit 301 is used to acquire time-series mass flow rate, temperature data, and geometric physical parameters required for the operation of the cold storage tank, and to set an initial set of parameters to be identified and prior information such as flow direction to obtain input data and parameters; the model building unit 302 is used to construct a hierarchical ordinary differential equation model describing the temperature changes of each layer based on energy conservation and the input data and parameters; the prediction unit 303 is used to use an adaptive step-size ODE solver to predict the temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model according to the input data and parameters to obtain the predicted temperature; the function construction unit 304 is used to construct a loss function by comparing the predicted temperature with the actual measured value; the optimization unit 305 is used to calculate the gradient based on the loss function and adjust the parameters through an optimization algorithm to minimize the loss function.

[0079] In one embodiment, the model building unit 302 includes: The temperature influence determination subunit is used to determine the upstream temperature based on the time series mass flow rate and its direction, thereby determining the temperature influence. The interlayer heat conduction influence subunit is used to evaluate the contribution of interlayer heat conduction to temperature change through the second difference of the temperature difference between adjacent layers and the interlayer thermal conductivity. The environmental heat transfer influence subunit is used to calculate the influence of environmental heat transfer on layer temperature using the sidewall area of ​​each layer, the environmental temperature difference, and the tank's environmental heat transfer coefficient. The integration subunit is used to integrate the temperature influence, the contribution of interlayer heat conduction to temperature change, and the influence of environmental heat transfer on layer temperature to form an ordinary differential equation describing the rate of temperature change over time. The combination subunit is used to establish the corresponding ordinary differential equations for all layers of water in the cold storage tank and combine them into a whole layered ODE model to obtain a layered ordinary differential equation model.

[0080] In one embodiment, the prediction unit 303 is used to call an adaptive step-size ODE solver to integrate the layered ODE from the initial time to each observation time based on the layered ordinary differential equation model according to the input data and parameters, so as to obtain the predicted temperature of each observation time and each layer.

[0081] In one embodiment, the tuning unit 305 is used to perform backpropagation calculation on the hierarchical ordinary differential equation model based on the loss function, through automatic differentiation or adjoint extension system, to calculate the gradient of the loss with respect to each physical parameter, and update the parameters according to a preset optimization algorithm.

[0082] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the aforementioned physical-driven computing center digital native model energy prediction system 300 and its various units can be found in the corresponding descriptions in the aforementioned method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0083] The aforementioned physical-driven computing center's digital-native model energy prediction system 300 can be implemented as a computer program, which can, for example... Figure 4 It runs on the computer device shown.

[0084] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0085] See Figure 4 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0086] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a digital native model energy prediction method for a physically driven computing center.

[0087] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0088] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a digital native model energy prediction method for a physically driven computing center.

[0089] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0090] The processor 502 is used to run a computer program 5032 stored in the memory to implement all steps of the digital native model energy prediction method of the physical-driven computing center.

[0091] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0092] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0093] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the digital-native model energy prediction method of the physically driven computing center.

[0094] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0096] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0097] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0098] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A digital native model energy prediction method for physical-driven computing centers, characterized in that, include: Acquire time-series mass flow rate, temperature data, and geometric physical parameters required for the operation of the cold storage tank, and set the initial set of parameters to be identified and prior information such as flow direction to obtain input data and parameters; Based on the law of conservation of energy, a layered ordinary differential equation model describing the temperature changes of each layer is constructed using the input data and parameters. The adaptive step-size ODE solver is used to predict the temperature distribution of each layer at each time point based on the input data and parameters according to the hierarchical ordinary differential equation model, so as to obtain the predicted temperature. A loss function is constructed by comparing the predicted temperature with the actual measured value. The gradient is calculated based on the loss function, and the parameters are adjusted by an optimization algorithm to minimize the loss function.

2. The energy prediction method for a digital native model of a physically driven computing center according to claim 1, characterized in that, The temperature data includes inlet water temperature, ambient temperature, and the observed temperature of each layer inside the cold storage tank at various times; the geometric and physical parameters include the volume, height, cross-sectional area, density, and specific heat of each layer.

3. The energy prediction method for a digital native model of a physically driven computing center according to claim 2, characterized in that, The initial set of parameters to be identified includes layer-specific response factor, thermal inertia factor, interlayer thermal conductivity, and tank environment heat transfer coefficient.

4. The energy prediction method for a digital native model of a physically driven computing center according to claim 3, characterized in that, The layer-specific response factor satisfy The thermal inertia factor satisfy Interlayer thermal conductivity satisfy Tank environment heat transfer coefficient satisfy .

5. The digital native model energy prediction method for a physically driven computing center according to claim 4, characterized in that, The method, based on energy conservation, utilizes the input data and parameters to construct a layered ordinary differential equation model describing the temperature changes of each layer, including: The upstream temperature is determined based on the time series mass flow rate and its direction in order to determine the temperature effect. The contribution of interlayer heat conduction to temperature change is assessed by the second-order difference of the temperature difference between adjacent layers and the interlayer thermal conductivity. The influence of environmental heat transfer on the layer temperature is calculated by using the side wall area of ​​each layer, the temperature difference of the environment, and the heat transfer coefficient of the tank environment. By integrating the effects of temperature, the contribution of interlayer heat conduction to temperature change, and the influence of environmental heat transfer on layer temperature, an ordinary differential equation describing the rate of temperature change over time is formed. The corresponding ordinary differential equations are established for each layer of water in the cold storage tank, and then combined into a whole layered ODE model to obtain the layered ordinary differential equation model.

6. The energy prediction method for a digital native model of a physically driven computing center according to claim 5, characterized in that, The ordinary differential equation is expressed as: ,in, For the first i upstream temperature of the layer; ; This refers to the interlayer thermal conductivity. The area of ​​the sidewalls of each floor; Given the ambient temperature difference; assume the cold storage tank is divided along its height into... The first water layer, the... Layer temperature is The layer volume is The effective heat capacity is ;in This is the thermal inertia factor of the layer. For the density of water, It is a specific heat at constant pressure.

7. The energy prediction method for a digital native model of a physically driven computing center according to claim 1, characterized in that, The process of using an adaptive step-size ODE solver to predict the temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model according to the input data and parameters, in order to obtain the predicted temperature, includes: The adaptive step-size ODE solver is invoked, and the layered ODE is integrated from the initial time to each observation time based on the input data and parameters according to the layered ordinary differential equation model, so as to obtain the predicted temperature of each layer at each observation time.

8. The energy prediction method for a digital native model of a physically driven computing center according to claim 1, characterized in that, The loss function includes ,in, The number of observation times. For the number of layers, in Overlay regularization terms that affect parameter range or smoothness. The set of physical parameters to be identified.

9. The energy prediction method for a digital native model of a physically driven computing center according to claim 8, characterized in that, The step of calculating the gradient based on the loss function and adjusting the parameters using an optimization algorithm to minimize the loss function includes: Based on the loss function, the hierarchical ordinary differential equation model is backpropagated using automatic differentiation or an adjoint extension system to calculate the gradient of the loss with respect to each physical parameter, and the parameters are updated according to a preset optimization algorithm.

10. A digital-native model energy prediction system with a physically driven computing center, characterized in that, include: The acquisition unit is used to acquire time-series mass flow rate, temperature data and geometric physical parameters required for the operation of the cold storage tank, and to set the initial set of parameters to be identified and prior information such as flow direction in order to obtain input data and parameters. The model building unit is used to construct a layered ordinary differential equation model describing the temperature changes of each layer based on the energy conservation principle and using the input data and parameters. The prediction unit is used to predict the temperature distribution of each layer at each time point based on the hierarchical ordinary differential equation model using an adaptive step-size ODE solver, according to the input data and parameters, so as to obtain the predicted temperature. The function building unit is used to construct a loss function by comparing the predicted temperature with the actual measured value; The tuning unit is used to calculate the gradient based on the loss function and adjust the parameters through an optimization algorithm to minimize the loss function.

Images