Electrical and mechanical collaborative scheduling method and system based on entropy flow manifold mapping and energy topology

Abstract

The present application relates to the technical field of smart grid, specifically to a computing power and electricity collaborative scheduling method and system based on entropy manifold mapping and energy topology. The present application includes obtaining computing power node data, converting high-dimensional computing power feature tensor into continuous information entropy manifold, constructing space-time power demand potential field and feature extraction, obtaining initial transient power compensation benchmark; converting into energy state variable in Hamilton system and constructing global energy evaluation functional; constructing multi-agent Nash game framework, reaching consensus at dynamic Nash equilibrium point, outputting synchronous adaptation instruction tensor and reducing dimension and decoupling, calculating node target voltage command and digital quantization coding; deriving the lead time of target current, calculating microsecond-level feedforward pulse width modulation duty cycle, synchronously issuing clock frequency constraint value and hardware execution structure body. The present application realizes the accurate prediction of microsecond-level computing power burst and the optimal transportation of energy space, effectively solving the prediction lag and energy scheduling problem in computing power and electricity collaboration.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, specifically to a computational-electricity collaborative scheduling method and system based on entropy manifold mapping and energy topology. Background Technology

[0002] Against the backdrop of large-scale AI model training and the "East-to-West Computing" strategy, data centers, as core computing infrastructure, are experiencing a non-linear surge in computing density and energy consumption. Achieving accurate prediction and real-time matching of electricity demand with computing power needs—a process known as "computing-power synergy"—has become crucial for the development of green intelligent computing centers. Currently, the industry primarily addresses this challenge through smart microgrids and high-frequency power semiconductor technologies. At the scheduling level, existing technologies generally rely on accumulated server operation data and power grid monitoring data, employing time-series models to statistically predict computing load, and then estimating electricity demand using static conversion coefficients, forming a sequential scheduling logic of "predictive calculation first, then power supply response."

[0003] However, existing technologies still have shortcomings in terms of "ultra-fast response" and "precise spatial routing": Computing tasks (especially GPU / NPU clusters) generate drastic current surges at the microsecond level, while traditional power grid response and compensation mechanisms operate at the millisecond to second level, creating a severe "semantic gap." This spatiotemporal lag in serial logic leads to transient voltage drops and computing frequency reduction during computing power peaks, while causing power redundancy and energy waste during troughs. There is a "modal heterogeneity" between the multidimensional characteristics of computing tasks (such as tensor dimensions and communication latency) and the physical state of the power grid (such as topological impedance and bus voltage). Existing systems lack a unified multiphysics joint modeling framework. When spatial computing power load is uneven within the data center, the rigid power network cannot achieve dynamic topology optimization and energy adaptive routing, resulting in the coexistence of local "power hotspots" and nearby "power outages," which limits overall energy efficiency and reliability.

[0004] Therefore, a computational-electricity collaborative scheduling method and system based on entropy manifold mapping and energy topology is needed. Summary of the Invention

[0005] This invention provides a computing-electricity collaborative scheduling method and system based on entropy manifold mapping and energy topology. The aim is to reconstruct the characteristics of massive discrete computing power tasks into a continuous high-dimensional manifold space by introducing the concept of "entropy manifold" from information thermodynamics, thereby solving the prediction lag problem of microsecond-level computing power bursts. At the same time, an energy topology routing network based on Hamiltonian system is constructed to realize the optimal spatial transport of electrical energy from the bus to the micro-computing unit under the multi-physics coupling framework.

[0006] The technical solution of this invention is as follows: The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology includes the following steps: Step S1. Obtain relevant data information of computing power nodes, extract multidimensional heterogeneous features of computing power instructions, and obtain high-dimensional computing power feature tensors through attention processing; transform the high-dimensional computing power feature tensors into continuous information entropy manifolds, and construct continuous spatiotemporal power demand potential fields; extract features from the spatiotemporal power demand potential fields to obtain the initial transient power compensation benchmark. Step S2. The initial transient power compensation reference state is transformed into the energy state variable in the Hamiltonian system to obtain the global phase space state matrix. A global energy assessment functional is constructed, and the physical damping term simulating the actual heat loss of the cable is reconstructed to obtain the global optimal power routing and allocation tensor. Step S3. Using the global optimal power routing and allocation tensor as the trump card of the power supply agent, construct a multi-agent Nash game framework. By constructing mutually constraining utility functions and using the alternating gradient ascent algorithm, a consensus is finally reached at a dynamic Nash equilibrium point, and the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility is output. Step S4. Reduce the dimensionality of the synchronization adaptation instruction tensor and decouple it into independent physical current references for each power supply phase to obtain the specific target current of each phase inductor. Calculate the target voltage command of the node and its corresponding digital quantization code, and write the obtained digital quantization code into the power management main control chip. Calculate the microsecond-level feedforward pulse width modulation duty cycle that is forcibly applied to each phase power transistor by differentiating the lead time of the target current. Establish an optimal function based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and construct a hardware execution structure for synchronous distribution, triggering the hardware fence power calculation synchronization to take effect.

[0007] Furthermore, the process of obtaining the global phase space state matrix in step S2 is as follows: Based on the electrical connection relationship of the underlying real power distribution network, a microgrid impedance topology diagram is constructed. For any two physically connected computing nodes in the diagram, a topological connectivity conductance weighting coefficient for computing nodes is defined. The expected power distribution landing point of the computing node is defined as the generalized position, and the transient driving force of power change is defined as the generalized momentum. The generalized position and the generalized momentum are tightly bound together to construct the fused phase space state tensor of the node, thereby obtaining the global phase space state matrix containing the states of all nodes.

[0008] Furthermore, in step S2, the details are as follows: Each sub-tensor in the global phase space state matrix is ​​automatically unpacked, and its generalized momentum and generalized position are separated to participate in the calculation. Finally, through dynamic evolution, the global phase space state matrix is ​​driven to slide towards the steady-state direction that minimizes the scalar value of the global energy assessment functional. By introducing a physical damping term to simulate the actual heat loss of cables, the state tensor is evolved by forward integration using an ordinary differential equation solver, and finally the globally optimal power routing and allocation tensor is reconstructed.

[0009] Furthermore, in step S3, within the policy space of the multi-agent Nash game framework, the specific process is as follows: The multi-agent system includes a power supply agent and a computation agent. The Nash equilibrium gap is introduced as a convergence measure. When the power supply agent and the computation agent cannot obtain substantial utility improvement even if they modify their strategies, the system is considered to have reached the Nash equilibrium point.

[0010] Furthermore, after the game converges, the extracted new strategy tensor generated by the power supply side and the new strategy tensor generated by the computing side represent the maximum concession and optimal choice made by the power supply side and the computing side under the current stringent physical constraints, respectively. The game result is transformed into control code that can be directly recognized by the underlying driver, and Pareto optimal fusion mapping is performed on the bilateral steady-state tensor to generate the final synchronization adaptation instruction tensor.

[0011] Furthermore, during the decoupling process in step S4, real-time thermal health weights for each phase are introduced for dynamic load balancing, as detailed below: First, by using a linear projection matrix and a bias activation function, the absolute value of the total physical current demand of the nodes in the next very short time is extracted from the high-dimensional tensor. Under the premise of meeting the absolute value of the total physical current demand, the system will automatically allocate the current load to the phase with low temperature and sufficient thermal margin to obtain an independent expected current reference scalar. Based on the output independent expected current reference scalar, the specific target current of each phase inductor is determined.

[0012] Furthermore, step S1 constructs a continuous spatiotemporal potential field for electricity demand, as detailed below: Using impulse functions and thermal diffusion kernels from physics, discrete computation timestamps are continuously mapped onto the real physical time axis and anchored in three-dimensional physical space.

[0013] Furthermore, the calculation process for the microsecond-level feedforward pulse width modulation duty cycle in step S4 is as follows: in, For nodes The Phase circuit in The per-unit value of the microsecond-level feedforward pulse width modulation duty cycle that should be output at any given moment; This is the input voltage to the upstream main bus. The steady-state target voltage; It is a feedforward forcing term; This is the nominal value of the output filter inductance for this phase; This is the first-order time derivative obtained for the feedforward predicted current; For inductor DC resistance voltage drop compensation; The proportional-integral-differential operator represents the error between the actual sampled voltage and the target voltage. Feedback adjustment.

[0014] A computational-electricity collaborative scheduling system based on entropy manifold mapping and energy topology includes the following: The unified representation module for computing power tasks acquires relevant data information of computing power nodes, extracts multidimensional heterogeneous features of computing power instructions, and obtains a high-dimensional computing power feature tensor through attention processing. The high-dimensional computing power feature tensor is transformed into a continuous information entropy manifold, and a continuous spatiotemporal power demand potential field is constructed. Features are extracted from the spatiotemporal power demand potential field to obtain the initial transient power compensation benchmark. The Hamiltonian potential energy mapping and energy optimization module transforms the initial transient power compensation reference state into the energy state variables of the Hamiltonian system, obtains the global phase space state matrix, constructs a global energy evaluation functional, reconstructs the physical damping term simulating the actual heat loss of the cable, and obtains the global optimal power routing and allocation tensor. The game-theoretic decision-making adaptation module uses the global optimal power routing and allocation tensor as the trump card of the power supply agent, constructs a multi-agent Nash game framework, and finally reaches a consensus at a dynamic Nash equilibrium point by constructing mutually constrained utility functions and using the alternating gradient ascent algorithm, outputting the final synchronous adaptation instruction tensor that takes into account both computing power performance and grid flexibility. The microsecond-level feedforward control module reduces and decouples the synchronization adaptation instruction tensor into independent physical current references for each power supply phase, obtains the specific target current of each phase inductor, calculates the target voltage command of the node and the corresponding digital quantization code; it differentiates the lead time of the target current to calculate the microsecond-level feedforward pulse width modulation duty cycle forcibly applied to each phase power transistor; it establishes an optimal function based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and constructs a hardware execution structure for synchronous distribution, triggering the hardware fence power calculation synchronization to take effect.

[0015] Furthermore, a global phase space state matrix module is constructed within the Hamiltonian potential energy mapping and energy optimization module. Based on the electrical connection relationship of the underlying real distribution network, a microgrid impedance topology diagram is constructed. For any two physically connected computing nodes in the diagram, a topological connectivity conductance weighting coefficient for the computing nodes is defined. The expected power allocation landing point of the computing node is defined as the generalized position, and the transient driving force of power change is defined as the generalized momentum. The generalized position and the generalized momentum are tightly bound together to construct the fused phase space state tensor of the node, thereby obtaining a global phase space state matrix containing the states of all nodes.

[0016] The beneficial effects of this invention are as follows: 1. First, this invention overcomes the physical hysteresis limitations of traditional passive power supply responses, reducing the risk of transient voltage drops during concurrent computing power. Existing data center power management modules (PMICs) primarily rely on closed-loop feedback mechanisms (such as PID or droop control), which only begin increasing inductor current after a substantial voltage drop has occurred. This "millisecond-level" hysteresis is ineffective against the "microsecond-level" current surges during AI large-scale model training, easily leading to GPU crashes. This invention, through a unified computing task representation module, directly intercepts code micro-tasks in the computing power scheduling queue, transforming discrete computing attributes into a continuous physical power demand potential field. Experiments show that this mechanism achieves precise microsecond-level feedforward compensation, reducing the maximum transient voltage drop of the core voltage by approximately 63.5%, significantly improving the power supply stability of the underlying hardware.

[0017] 2. This invention solves the problem of rigid spatial scheduling in physical microgrids, significantly revitalizing "stranded power" within data centers. In traditional high-density racks, power distribution is typically static and rigid. When local nodes are overloaded, causing power supply overheating, nearby idle nodes cannot cross the bus to provide support. This invention, through Hamiltonian potential mapping and energy optimization modules, abstracts the power supply network into a dynamic system, utilizing physical impedance and spatial attenuation to construct energy functional optimization. This adaptive routing mechanism based on the global situational field breaks down the physical islands of power supply links, improving rack-level effective power utilization by approximately 18.2%, effectively supporting higher-density computing power deployments without increasing the total grid capacity.

[0018] 3. This invention cleverly resolves the physical conflict between "ultimate computing power performance" and "grid thermal safety boundaries," achieving Pareto optimality under multiple objectives. Faced with stringent thermal design power limits, existing technologies often employ a drastic, one-size-fits-all approach of frequency throttling, leading to a precipitous drop in throughput for computing clusters performing critical training tasks. This invention, through a game-theoretic decision-making adaptation module, assigns independent utility functions to the power supply and computing ends for double-blind game theory. While ensuring the grid never experiences thermal breakdown, the system can dynamically extract every bit of power margin to supply core computing power. Test data shows that when the same stringent power capping event is triggered, this method retains approximately 24.7% more effective floating-point throughput than traditional hard-cutoff strategies.

[0019] 4. This invention achieves true cross-modal hardware-level "computing and power integration" synchronous control, eliminating the time lag between software and hardware execution. Traditionally, there is a system bus time lag between the CPU / GPU's downclocking instructions and the VRM's (voltage regulation module's) power regulation instructions. This blind spot is a major source of OOM (Out of Memory) and NaN (NaN) errors. This invention innovatively designs a hardware-level "synchronization gate" through a microsecond-level feedforward control module, packaging and issuing PWM duty cycle, adaptive VID, and clock frequency constraints to ensure physical activation within the same nanosecond clock cycle. This low-level locking mechanism reduces the response latency of computing and power integration from the traditional hundreds of microseconds to less than 15 microseconds, providing a highly robust low-level control protocol for next-generation green intelligent computing centers. Attached Figure Description

[0020] Figure 1 This is a comprehensive comparison chart of maximum voltage drop and stranded power under the high-density intelligent computing dataset; Figure 2 A comparison chart of the effective computing power throughput retention rate of each model under different computing power scenarios; Figure 3 This is a comparison chart of the synchronization delay of computing and power coordination instructions for various models under the high-density intelligent computing dataset. Detailed Implementation

[0021] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. It should also be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of the invention.

[0022] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0023] This embodiment provides a computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology, including the following steps: S1. Obtain relevant data information / node spatiotemporal state vectors of computing power nodes, extract multidimensional heterogeneous features of computing power instructions, and obtain high-dimensional computing power feature tensors through attention processing; transform the high-dimensional computing power feature tensors into continuous information entropy manifolds, and construct continuous spatiotemporal power demand potential fields; perform feature extraction on the spatiotemporal power demand potential fields to obtain the initial transient power compensation benchmark.

[0024] Specifically, this step, serving as the feedforward sensing end of the entire computing and power collaboration system, aims to break the lag response mechanism of traditional data centers—where "computing load occurs first, and power is passively adjusted later"—and solve the problem of instantaneous voltage drops caused by microsecond-level concurrent computing power. Its core logic is as follows: the system directly connects to the underlying computing task scheduler, intercepting and parsing these micro-task sequences before the computing instructions are actually issued to the hardware for execution; subsequently, it uses an attention mechanism to extract the heterogeneous impact of different computing characteristics on power consumption, transforming it into a continuous "information entropy manifold"; finally, through a cross-medium time smoothing kernel and spatial anchoring function, it maps the discrete computing logic space onto a continuous physical spatiotemporal coordinate axis, generating a high-resolution "power demand potential field."

[0025] S101. Since different types of computing power tasks (such as matrix multiplication and other computationally intensive tasks that mainly cause high-frequency switching of kernel logic units, while memory-intensive tasks such as large model weight loading will cause a surge in memory bandwidth and power consumption) have fundamentally different impacts on the current ramp-up rate of the underlying power supply network, fine-grained semantic feature extraction must be performed first.

[0026] For any underlying physical computing node within the data center power supply topology In the future microsecond-level forward timestamps planned by the task scheduler The system extracts its multidimensional computational features and constructs a high-dimensional computational power feature tensor by combining an attention mechanism. The calculation formula is as follows: in, A globally unique index representing the underlying physical computing power node; For microsecond-level discrete timestamps extracted from the task scheduling queue; This represents the expected floating-point operation rate at that moment, used to characterize the transient power consumption caused by pure logic computation; This is an estimate of the expected memory bandwidth throughput, representing the surge in memory power consumption caused by high-frequency memory read and write operations. This represents the effective payload size for inter-node communication, reflecting the interconnect power consumption under distributed training. This represents a cascading and concatenating operation of features along the channel dimension. and These are the learnable weight matrix and bias vector of the self-attention perception layer, respectively, used to adaptively amplify the feature dimension that has the most significant impact on transient current; It is a non-linear activation function; This is the static physical topology location encoding vector of the computing node in the three-dimensional space of the data center rack; For the tensor Kronecker product; the output is It is a high-dimensional comprehensive computing power feature tensor that deeply integrates task computing attributes and physical spatial location.

[0027] S102. After obtaining the high-dimensional computing power feature tensor, since traditional computing power feature values ​​are discrete and isolated, they cannot be directly used as a reference benchmark for continuous control of the power system. This step innovatively introduces the concept of information thermodynamics, transforming the feature tensor into a continuous "information entropy manifold." The level of information entropy intuitively reflects the severity of computing power fluctuations: the higher the entropy value, the more unpredictable the concurrent mutations in computing power, and the higher the requirements for the flexible compensation capability of the power management chip. Define this node in... Entropy manifold tensor at time step The calculation formula is as follows: in, For the comprehensive computing power feature tensor The total number of dimensions of the expanded one-dimensional feature vector; For the eigenvector of the th The actual value of each element; The goal is to obtain the "feature probability distribution" after normalizing the feature values ​​using the Softmax function, so that it satisfies the calculation conditions of Shannon information entropy. It is the natural logarithm; This is a pre-trained "computational-electrical cross-Jacobi sensitivity matrix." The core physical meaning of this matrix lies in describing the partial derivative mapping relationship between small state changes in the digital computation domain and fluctuations in the analog power domain, thus acting as a bridge for cross-modal conversion; the output... It is a continuous manifold tensor that incorporates the intensity of high-frequency fluctuations in microscopic computing power and the direction of physical sensitivity.

[0028] S103. The issuance of computing power commands is highly discrete within a computer clock cycle, manifesting as a series of pulses, while power transmission and distribution in the physical world (such as capacitor charging and discharging, and inductor current hysteresis) is absolutely continuous. Directly using discrete entropy manifolds for control would lead to high-frequency oscillations in the control signal. Therefore, by utilizing impulse functions and thermal diffusion kernels from physics, discrete computation timestamps are... Continuous mapping to the real physical timeline Above, and anchored in three-dimensional physical space In this process, a globally continuous "spatiotemporal demand potential field" is constructed. The calculation formula is as follows: in, For the generated three-dimensional physical space coordinates and real continuous physical time The power demand potential field tensor at the location; This represents the total number of computing nodes participating in this coordinated scheduling. and The start and end points of the predicted time window for the task scheduling queue; This is a three-dimensional Dirac impulse function, its function is to accurately locate the nodes. The fluctuations in computing power are strictly anchored to their corresponding physical absolute space coordinates. Place; It is a Gaussian time-smoothing kernel function used to simulate the inherent physical charging and discharging inertia of inductive and capacitive components in power supply networks; The intrinsic time damping constant of the underlying power supply network is determined by the equivalent series resistance and inductance characteristics of the actual hardware; the output of this step is... The discrete software code instructions were successfully transformed into a continuous "energy demand topography map" covering the entire physical space of the rack. The areas where the potential energy rises sharply in the map represent physical nodes where computing power surges and high-intensity current extractions are about to occur.

[0029] S104. After obtaining the continuous spatiotemporal demand potential field, in order to achieve microsecond-level advance prediction, the system needs not only to know "how much electricity is needed at a certain moment," but also "how fast the electricity demand increases." By taking the first-order time derivative of the potential field, its dynamic gradient characteristics are extracted to generate the initial transient power compensation reference state. The formula is as follows: in, For nodes The extracted basic power feedforward compensation tensor contains the demand base and the demand ramp-up rate; This represents the static potential field amplitude at the node location calculated via S103. For the actual physical time The partial derivative operator is obtained. In a physical sense, it directly maps the slope of the current surge in the next instant; Indicates feature fusion; The feature-reduced projection matrix; This is a layer normalization operation to prevent gradient explosion; This is the Hadamard product (element-level multiplication). This is the physical safety threshold mask vector for the underlying power supply device, used to hard-truncate the prediction results at the algorithm level, ensuring that the initial output state never exceeds the maximum power envelope allowed by the PMIC hardware; the final output... This will serve as the foundation for energy optimization and will be directly input into the next Hamiltonian potential mapping and energy optimization module.

[0030] S2. The initial transient power compensation reference state is transformed into the energy state variable in the Hamiltonian system to obtain the global phase space state matrix. A global energy assessment functional is constructed, and the physical damping term simulating the actual heat loss of the cable is reconstructed to obtain the global optimal power routing and allocation tensor.

[0031] Although step S1 accurately predicts the power demand of a single computing node in the very short term, in actual high-density intelligent computing center microgrids, each server node is physically coupled to the bus and distribution network. If power is supplied locally and isolated based solely on the surge in demand of a single node, without considering the impedance and voltage drop of the physical cables globally, it is easy to cause local power supply links to overload, while the power capacity of adjacent links remains idle. Therefore, this paper innovatively abstracts the underlying physical distribution network as a continuously evolving "Hamiltonian energy dynamic system," uses the feedforward prediction result of step S1 as an external excitation force field, and packages the power allocation state of each node as a Hamiltonian phase space tensor. By simulating the natural optimal flow of energy in the physical topology (seeking the state with the lowest total potential energy and the highest work efficiency), the globally optimal power redirection routing strategy is calculated.

[0032] S201. To perform energy routing from a global perspective, the physical power supply network of the data center first needs to be transformed into a mathematical graph structure that can be processed by a computer. The system constructs a microgrid impedance topology diagram based on the electrical connections of the underlying real power distribution network. For any two physically connected computing nodes in the diagram... and Define its topological connectivity conductance weighting coefficient. The calculation formula is as follows: in, Represents a node With nodes The directional conduction weights (characterizing equivalent conductivity) between them during energy routing. For actual physical bus connections at nodes and The equivalent AC / DC mixed impedance scalar value (including cable internal resistance and parasitic reactance) between them; the larger this value is, the more difficult it is for electrical energy to flow along this path. To prevent the use of a minimal constant compensation term with a denominator of zero; This represents the squared Euclidean distance between the two nodes calculated using the physical space coordinates that have been anchored in step S1. This is a preset space electromagnetic energy radiation attenuation control factor; It is a physical topology indicator function if and only if the node and The value is 1 if there is a direct physical copper bus or cable link in the underlying PMIC power management network, and 0 otherwise. This step generates the weighting coefficient. This forms a global impedance adjacency matrix, which serves as a physical spatial channel constraint for "energy transport," rigidly defining the effective path for subsequent route optimization.

[0033] S202. After establishing the physical transmission channel, the system needs to convert the power compensation benchmark output in step S1 into a valid data set. This is transformed into energy state variables within a Hamiltonian system. In classical analytical mechanics, a Hamiltonian system must be jointly described by "generalized coordinates (position)" and "generalized momentum"; the absence of either one prevents subsequent system state evolution. Here, the "expected power allocation landing point of the computing node" is defined as the generalized position. The "transient driving force of power change" is defined as generalized momentum. And bind the two tightly to construct a node. Fusion phase space state tensor The calculation formula is as follows: in, For nodes The initial generalized position tensor in the dynamical system; and The linear transformation weights and biases that map the feedforward reference to the Hamiltonian phase space position dimension; It is a hyperbolic tangent function used to constrain power demand within a standard physical per-unit range. ; Let be the initial generalized momentum tensor of this node, representing the dynamic potential energy reserve for power adjustment; and We assign weights and biases to the momentum mapping; LeakyReLU is used to introduce a slight nonlinearity into the projected position vector and prevent the complete loss of negative value information. This represents a concatenation operation along the feature channel dimension. The resulting... It is a complete phase space state tensor that strongly binds position to momentum.

[0034] Furthermore, the definition includes all The global phase space state matrix of the states of each node. This matrix will serve as the sole independent variable for subsequent global functional calculations and differential equation solutions.

[0035] S203. Defines the global phase space state matrix. Then, the system needs to construct a global energy assessment functional (total Hamiltonian). This is used to assess the "physical rationality" of the current power distribution state. An efficient power routing strategy without local hotspots should satisfy the following: maximizing the utility of each node (approaching computing power requirements), minimizing the power distribution difference between adjacent nodes (preventing local bus over-limit breakdown), and ensuring stable overall system kinetic energy. A total Hamiltonian pseudo-energy functional is constructed. The calculation formula is as follows: in, The scalar value representing the total system energy functional of the entire data center power supply network at the current moment has its only input variable being the global state matrix generated in step S202. The system will automatically unpack the formula. Each subtensor in And separate its generalized momentum. and generalized position Participate in specific calculations; the first term of the formula "System pseudo-kinetic energy" This is the transpose of generalized momentum. The preset inverse mass matrix is ​​mapped to the inverse inertia coefficient of the underlying power supply filter capacitor and energy storage inductor. This is a multilayer perceptron network with "node self-potential" used to evaluate the current allocation scheme. The degree of approximation to the computing power requirement predicted in step S201; the last term of the formula is "topological coupling potential energy", which is obtained using the topological weights in step S201. Squared power state difference between nodes To impose penalties. The ultimate goal of the system is to drive [the system] through dynamic evolution. To the envoy Minimal steady-state directional sliding.

[0036] S204. In order to find the minimum point of the above functional (i.e., the globally optimal power allocation scheme), the system uses... Starting from the initial point, on the virtual algorithm integration timeline The spontaneous flow of energy is simulated. To ensure the system converges quickly to a physical steady state, this step introduces a physical damping term to simulate the actual heat loss of the cable. The state tensor is then solved using an ordinary differential equation (ODE) solver. By performing forward integral evolution, the globally optimal power routing and allocation tensor is finally reconstructed and output to the next module. The formula is as follows: in, This is the virtual integral time variable used for optimization iteration within the Hamiltonian system, with the evolution starting point being... And the initial state Equal to the product generated in step S202 ; This represents the derivative of the state tensor with respect to virtual time. It is a standard Hamiltonian symmetric matrix; The partial derivative of the total energy functional with respect to the current node state tensor in step S203 (i.e., the Hamiltonian gradient field) guides the state to slide towards the lowest energy valley. An artificially introduced dissipative damping diagonal matrix is ​​used to simulate the real Joule heat loss in the distribution network, breaking the absolute conservation of energy to promote network convergence. The feedforward demand in step S1 is injected into the differential equation as an external, continuous forcing force to prevent the system from decaying to zero; after a preset virtual termination step size... After integral evolution, the system reaches physical steady state, at which point the nodes are extracted. final state tensor Subsequently, learnable decoding weights are utilized. With bias as well as Smooth activation function, for Perform inverse reduction of physical dimensions; finally, use the Hadamard product... and To achieve secure integration.

[0037] The final output of this step It is a "globally optimal power routing and allocation tensor" that has undergone global impedance trade-offs and spatial topology energy flow optimization. It completely resolves the local overload problem and will be directly used as the basic strategy input into the next game decision adaptation module.

[0038] S3. Using the global optimal power routing and allocation tensor as the trump card of the power supply agent, by constructing mutually constrained utility functions and applying the alternating gradient ascent algorithm, a consensus is finally reached at a dynamic Nash equilibrium point, and the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility is output.

[0039] Specifically, while step S2 achieves optimal energy transport at the macroscopic level based on physical impedance, it is essentially "power supply-oriented," potentially suppressing the instantaneous bursts of demand from certain high-priority computing tasks, thus causing performance degradation. To resolve the deep physical conflict between the "upper limit of power supply thermal safety" and the "lower limit of extreme computing performance," a multi-agent Nash game framework is innovatively constructed. Within a microsecond-level decision window, the system virtually creates a "power supply agent" representing the interests of the power grid side and a "computing agent" representing the interests of the server side. As the trump card of the power supply intelligent agent, by constructing mutually constraining utility functions and using the alternating gradient ascent algorithm, the bargaining process between the two parties is simulated, and a consensus is finally reached at a dynamic Nash equilibrium point, outputting the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility.

[0040] S301. The first step in the game is to establish the "action strategy space" and "bargaining chips (i.e., utility functions)" for both sides. The "power supply agent" wants the actual output power to be as close as possible to the globally safe routing solution given in step S2 to prevent bus overload; while the "computation agent" wants to obtain sufficient power to support its ideal full-load performance. For a specific node... In the initial iteration step of the game At that time, the initial action policy tensor of the power supply agent is defined as The initial action policy tensor of a computational agent is defined as follows: Based on this, a conflict utility function for both parties is constructed. and The calculation formula is as follows: in, The number of discrete virtual iterations in the game process; For the power supply agent in the first The proposed "planned power supply quota" tensor; For computational agents in the first... The proposed "expected load computing power" tensor; This is the absolute ideal computing power performance expectation tensor of the node extracted through the task priority queue (i.e., the full-power state without considering any power limitations). Representing the scalar utility gain of the power supply side under the current strategies of both parties, its first term utilizes the L2 norm ( ) and weighting coefficients Forced penalty power supply strategy deviates from Hamiltonian potential energy mapping and energy optimization module global optimal solution The behavior of the second term introduces a smooth approximation function. Temperature mapping matrix The Hadamard product of the power supply and load exceeds the hardware thermal design limits. Partial application Severe hot crash penalty for weighting; The scalar utility gain representing the computational side is represented by its first term, which is derived from the energy conversion efficiency matrix. Multiply the actual power supply by the calculated load, and supplement with The weight represents the positive "computing throughput gain," while the second term utilizes... Severe penalties for computational agents, such as reduced frequency or load (i.e., deviation from ideal expectations), result in the agent being penalized by weights. The compromise behavior of the two utility functions forms a mutual constraint in mathematics: the power supplier dares not supply power arbitrarily (for fear of thermal collapse), and the computing party dares not arbitrarily reduce the frequency (for fear of performance loss), which constitutes a solid foundation for non-cooperative game theory.

[0041] S302. Once the utility function is defined, both parties must dynamically adjust their strategies based on the other's previous bid in order to maximize their respective interests (i.e., maximize the utility function value). Due to the physical limitations of the underlying hardware system (e.g., PMIC cannot output negative voltage, GPU frequency cannot be infinitely high), any strategy update must be strictly limited to the effective action space that conforms to physical laws. This step uses the multi-agent alternating gradient ascent rule with momentum acceleration to iteratively evolve the bilateral policy tensor. The update formula is as follows: in, and The power supply side and the computing side are respectively in the first The new strategy tensor generated by round-robin game; and These are the game learning rates (representing the step size of negotiation compromise) for both sides updating their strategies; the partial derivative terms. This means that the power supply agent modifies the power supply plan along the gradient direction that maximizes its own utility, while the computational agent has already "sneaked in" on the power supplier's new offer during the update process. And take the partial derivative accordingly. This reflects the following characteristic of sequential games; and This is the momentum damping coefficient. The momentum term is introduced to eliminate the high-frequency oscillations of the gradient in the canyon region and accelerate the negotiation process, representing the change in strategy in the previous round. For the power supply physical feasible region projection operator, the updated power supply tensor is forcibly truncated within the space of the maximum output current and voltage regulation accuracy allowed by the power module hardware; similarly, To compute the projection operator of the physically feasible region, it is necessary to ensure that the expected load strategy does not exceed the chip's performance state or dynamic voltage / frequency adjustment range. This step ensures that virtual negotiation always takes place within the safety boundaries allowed by the underlying hardware.

[0042] S303. Dynamic Nash Equilibrium Gap Measure and Steady-State Strategy Convergence Extraction The goal of game theory is to reach consensus; therefore, the system must possess criteria for determining when negotiations end. This step introduces the "Nash equilibrium gap" as a convergence measure. When both parties cannot achieve a substantial increase in utility even after modifying their strategies, the system is considered to have reached a Nash equilibrium. The scalar value of the Nash equilibrium gap is then calculated. And convergence extraction is performed, and the discrimination formula is as follows: in, This indicates that the Frobenius norm is calculated for the change in the policy tensor, which is used to quantify the overall fluctuation of the entire tensor matrix in this iteration; That is, the current number The sum of Nash equilibrium gaps in round-robin games; the smaller this value, the more resolute and consistent the attitudes of both sides are. The preset limit of convergence tolerance threshold is used to determine whether a consensus has been reached; A hard upper limit on the maximum number of iterations is set to prevent the algorithm from getting stuck in an infinite loop. Since actual scheduling must be completed in microseconds, this value is usually constrained to a small range. When the condition is met, the iteration loop is broken, the system freezes the last updated policy tensor, and extracts it as the optimal power supply policy in the Nash equilibrium state. With the optimal computing load as the bottom line .

[0043] S304. Extracted after game convergence and These represent the maximum concessions and optimal choices that the power supply and computing sides can make under the current stringent physical constraints, respectively. To transform the game result into control code that the underlying driver can directly recognize, this step performs Pareto-optimal fusion mapping on the bilateral steady-state tensors, generating a synchronization adaptation instruction tensor output to the final execution module. The synthesis formula is as follows: in, This is the final product of this step, namely, after compromises involving feedforward computing power requirements, network-wide impedance routing, and double-blind game theory, achieving a perfect computing-power collaborative scheduling and control tensor in both physical and logical spaces; the formula calls... and The steady-state output of the game, following the steps; It is a bilinear pooling outer product operator. Its mathematical significance lies in not only preserving the independent decision-making characteristics of power supply and computation, but also deeply capturing the second-order product coupling relationship between voltage and frequency. and These are the final fusion projection matrix and bias vector used for dimensional restoration and instruction encoding formatting, respectively.

[0044] The output of this step It completely eliminates the modal gap between computing power scheduling and power transmission and distribution, and will serve as the ultimate guiding principle, seamlessly fed into the last microsecond-level feedforward control module to directly drive the underlying PMIC and GPU registers.

[0045] S4. The synchronization adaptation instruction tensor is reduced in dimension and decoupled into independent physical current references for each power supply phase to obtain the specific target current of each phase inductor. The target voltage command of the node and the corresponding digital quantization code are calculated, and the obtained digital quantization code is written into the power management main control chip. The derivative of the lead time of the target current is calculated to obtain the microsecond-level feedforward pulse width modulation duty cycle that is forcibly applied to each phase power transistor. The optimal function is established based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and a hardware execution structure is constructed for synchronous distribution to trigger the hardware fence power calculation synchronization to take effect.

[0046] Specifically, this step closely follows the final synchronization adaptation instruction tensor of computing and power coordination output in step S3. Although step S3 derives a Pareto optimal solution that balances computing performance and power grid safety through Nash game theory, this solution remains a high-dimensional floating-point matrix stored in computer memory. The underlying server's motherboard hardware (such as the intelligent PMIC chip, the multiphase voltage regulation module VRM, and the CPU / GPU dynamic voltage frequency regulation DVFS register) cannot directly read such tensors. Therefore, this step, acting as the "analog-to-digital conversion and execution center" of the entire computing-power collaborative system, aims to... The system reduces and decouples the components into specific hardware control primitives. Through multiphase current reference decoupling, adaptive voltage positioning encoding, and feedforward pulse width modulation duty cycle generation containing calculus control laws, the system raises the inductor current and locks the safe frequency a few microseconds before the computing power commands are actually transmitted by the microarchitecture pipeline, thus completely eliminating the risk of downtime caused by transient voltage drops.

[0047] S401. Modern high-performance AI chips typically employ multi-phase voltage regulator modules (VRMs) for power supply to distribute the extremely large operating current (often reaching hundreds or even thousands of amperes) and balance heat generation. This step first receives the comprehensive adaptation tensor output from step S3. The system reduces its dimensionality and decouples it into independent physical current references for each power supply phase. To prevent thermal breakdown of a single phase due to prolonged high load, the system introduces real-time thermal health weights for each phase during decoupling for dynamic load balancing. (For nodes...) The Phase power supply circuit ( Its independent desired current reference scalar The calculation formula is as follows: in, For nodes The total effective power supply phases of the underlying PMIC configuration; the formula is first expressed through a linear projection matrix. With bias Cooperate Activation function, from high-dimensional tensors Extracting nodes Absolute value of total physical current demand in the very short term ; For the first time specifically used for feature decoupling Phase channel mapping vector; For the real-time feedback of the underlying sensors The physical junction temperature health index of the phase power stage (the higher the temperature, the larger this value). This is the thermal penalty amplification factor; the fractional part utilizes the term with thermal penalty. The mechanism performs normalized weight allocation, the core physical meaning of which lies in: satisfying the total current Under the given demand, the system will automatically allocate more current load to the phase with lower current temperature and more sufficient thermal margin, thereby achieving optimal dynamic thermal balance at the physical hardware level. The output of this step... The specific target current for each phase inductor was determined.

[0048] S402. After determining the target current, to reserve sufficient voltage sag margin during peak computing power periods, the system must perform Adaptive Voltage Positioning (AVP). Simultaneously, the target voltage must be converted into a digital voltage identification code (VID) before being written to the CPU / GPU power supply registers. Compute Node Target voltage command and its corresponding digital quantization code The formula is as follows: in, The node after joint correction by AVP droop control and artificial intelligence Desired analog input voltage value; The nominal no-load reference voltage for this operating condition, as specified by the chip manufacturer; The equivalent load line impedance designed for hardware circuits serves to control the current flow. When the current increases, the target voltage is actively reduced, and the parasitic capacitance is used to release the charge instantaneously, thus alleviating the voltage drop at the initial stage of the current jump. This is a nonlinear dynamic voltage compensation term derived from game theory tensors in an AI algorithm, used to break away from traditional fixed voltage compensation methods. Rigid constraints; This is a rounding function used to convert analog continuous voltages into discrete digital codes; The lowest effective base voltage supported by the underlying PMIC; This is the minimum voltage regulation step resolution of the PMIC; For Hadamah accumulation; This is a security register mask to prevent the generated VID from exceeding the limit and burning out the chip.

[0049] The final generated number The power management controller chip can be directly written via I2C or PMBus bus.

[0050] S403. Although the first two steps determine the output voltage and current, the circuit contains a large filter inductor ( According to the laws of physics The current ramp-up has a significant inherent time lag. If adjustments are made only when the computing power peaks, it will be too late. This step is the core of "feedforward," using the derivative of the target current's lead time to directly calculate the microsecond-level pulse width modulation duty cycle that should be forcibly applied to each phase power transistor. This forces the inductor to store energy in advance. The calculation formula is as follows: in, For nodes The Phase circuit in The per-unit value of the PWM duty cycle that should be output at any given time (range 0 to 1); This refers to the input voltage of the upstream main bus of the cabinet; the first term within the parentheses in the formula. The steady-state target voltage; the second term To overturn the core of traditional feedback control, the "feedforward forcing term". This is the nominal value of the output filter inductance for this phase. The first-order time derivative (i.e., the predicted rate of change of current) for the feedforward predicted current increases sharply when the system predicts a dramatic increase in current within microseconds, directly maximizing the duty cycle and forcing the inductor to charge ahead of time; the third term... This is for inductor DC resistance voltage drop compensation; in addition to large-signal feedforward prediction, the formula retains the classic small-signal closed-loop residual fallback mechanism at the end. The proportional-integral-derivative (PID) operator represents the error between the actual sampled voltage and the target voltage. The classic feedback adjustment ensures steady-state accuracy. This step generates... This will be converted into an extremely high-frequency switching drive signal and applied to the transistor gate.

[0051] S404. Within the same extremely short clock cycle generated by the feedforward power command, the system must also synchronously send the adapted computing power limit to the processor to ensure that the computing power does not exceed the limit before the power is fully ready. This step generates the final upper limit of the computing clock frequency constraint. And construct an execution structure containing all underlying primitives. To distribute the data synchronously, use the following formula: in, For underlying computing nodes In future physical time Upper limit of dynamic clock frequency under power supply constraints; This is a scalar function that takes the minimum value and is used for safe instruction truncation. This is a hard ceiling on the maximum turbo frequency physically supported by the processor chip at the factory. This is a linear projection weight matrix used to reduce the dimensionality of abstract high-dimensional features and map them to specific clock frequency values; It is a sigmoid nonlinear activation function used to smoothly scale the input features to... The proportional range is designed to facilitate frequency adjustment; To strictly accept the computational game synchronization adaptation instruction comprehensive tensor output from step S3; This is the atomic-level hardware execution structure that the system ultimately generates, which encapsulates all the underlying control primitives for computing power and electricity. The notation for a set represents a collection of items from the first set to the last set. to the first The set of microsecond-level feedforward pulse width modulation duty cycle scalars for all physical circuits of the phase (total effective power supply phases) calculated in step S403; To strictly adhere to the adaptive digital voltage identification code calculated in step S402; The "SyncBarrier" operation primitive, representing the underlying hardware driver interface layer, forces the representation of the power supply state via a high-speed hardware bus. and Control signals, and clock adjustment signals representing the status of power-consuming components (computing power). It must be latched and physically activated simultaneously by the chip's motherboard registers within the same nanosecond-level hardware clock tick.

[0052] This hardware-level binding mechanism completely seals off the instantaneous voltage breakdown vulnerability caused by the asynchronous transmission delay of digital control signals, which may result in "voltage drop but computing power not drop". This achieves closed-loop microsecond-level integrated collaboration of computing and power at the lowest physical level.

[0053] This invention provides a computational-powered collaborative scheduling system based on entropy manifold mapping and energy topology, including the following: The unified representation module for computing power tasks acquires relevant data information of computing power nodes, extracts multidimensional heterogeneous features of computing power instructions, and obtains a high-dimensional computing power feature tensor through attention processing. The high-dimensional computing power feature tensor is transformed into a continuous information entropy manifold, and a continuous spatiotemporal power demand potential field is constructed. Features are extracted from the spatiotemporal power demand potential field to obtain the initial transient power compensation benchmark. The Hamiltonian potential energy mapping and energy optimization module transforms the initial transient power compensation reference state into energy state variables in the Hamiltonian system, obtains the global phase space state matrix, constructs a global energy evaluation functional, reconstructs the physical damping term simulating the actual heat loss of the cable, and obtains the global optimal power routing and allocation tensor. The game-theoretic decision-making adaptation module uses the global optimal power routing and allocation tensor as the trump card of the power supply agent to construct a multi-agent Nash game framework. By constructing mutually constraining utility functions and using the alternating gradient ascent algorithm, a consensus is finally reached at a dynamic Nash equilibrium point, and the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility is output. The microsecond-level feedforward control module reduces and decouples the synchronization adaptation instruction tensor into independent physical current references for each power supply phase, obtains the specific target current of each phase inductor, calculates the target voltage command of the node and the corresponding digital quantization code; it differentiates the lead time of the target current to calculate the microsecond-level feedforward pulse width modulation duty cycle forcibly applied to each phase power transistor; it establishes an optimal function based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and constructs a hardware execution structure for synchronous issuance, triggering the hardware fence power calculation synchronization to take effect; In the Hamiltonian potential energy mapping and energy optimization module, a global phase space state matrix module is constructed. Based on the electrical connection relationship of the underlying real distribution network, a microgrid impedance topology diagram is constructed. For any two physically connected computing nodes in the diagram, a topological connectivity conductance weighting coefficient for computing nodes is defined. The expected power allocation landing point of the computing node is defined as the generalized position, and the transient driving force of power change is defined as the generalized momentum. The generalized position and the generalized momentum are tightly bound together to construct the fused phase space state tensor of the node, thereby obtaining a global phase space state matrix containing the states of all nodes.

[0054] In this embodiment of the invention, in order to systematically verify the method of the invention, "computing and power co-scheduling method based on entropy manifold mapping and energy topology", we constructed and used two datasets with different characteristics and simulation testbeds to conduct comparative experiments on the real co-scheduling performance in general data center scenarios and extreme high-density AI computing power clusters.

[0055] 1. Description of Dataset and Test Environment Open-Compute-Power-2025 (Public Benchmark Dataset): This dataset is sourced from operational logs of large-scale public cloud data centers in international open-source environments. It primarily includes standard web services, database queries, and routine machine learning inference tasks. Its characteristic is relatively smooth fluctuations in computing load, and it is mainly used to validate the model's benchmark capabilities in handling basic load forecasting and static power allocation under typical business scenarios.

[0056] Jereh-AI-Grid-2026 (Self-built High-Density Intelligent Computing Double-Blind Dataset): This is an extremely challenging dataset that was self-built using a real hardware testbed (including a 64-node high-power NPU array and a corresponding multi-phase PMIC power supply microgrid). This dataset fully records the entire process of distributed tensor parallel training of a large language model (LLM) with hundreds of billions of parameters over 30 days. The data includes microsecond-level task launch queues, cross-node communication payloads, and synchronously acquired physical telemetry data such as bus voltage, phase current, and inductor junction temperature. The aim is to verify the core advantages of this invention in handling extreme current surges, power supply hotspot routing, and microsecond-level feedforward game theory.

[0057] Experimental Setup: To comprehensively evaluate performance, we selected three representative methods currently mainstream in industry and academia for comparison with the method of this invention. PID / Droop Control (Traditional Hardware Droop Control): The most widely used pure analog / digital closed-loop feedback control scheme in industry, representing a purely passive power response benchmark without computational feedforward. LSTM-Power (Time-Series Deep Learning Prediction): Utilizes LSTM to collect historical CPU / GPU utilization for power consumption prediction, and adds a simple fixed-ratio power limiting strategy, representing the first generation of AI-based passive power prediction technology. MADDPG (Multi-Agent Deep Reinforcement Learning): Employs mainstream reinforcement learning algorithms to dynamically allocate rack power, representing a strong pure data-driven AI competitor in the current scheduling field, but lacks the underlying Hamiltonian physics mechanism and nanosecond-level hardware synchronization constraints.

[0058] Evaluation indicators: See attached document Figure 1 Maximum transient voltage drop: measures the voltage regulation capability of the power supply network when computing power surges. The lower the better (usually, chip safety requirements stipulate that the voltage drop should not exceed 10% of the nominal voltage). Stranded power ratio: This measures the percentage of idle power in a server rack that cannot be utilized due to routing instability, reflecting space allocation capabilities; the lower the better. See attached document Figure 2 Computing throughput retention rate: This measures the amount of effective computing power retained by the system when the network power cap is triggered, compared to the ideal unlimited state. The higher the better. See attached document Figure 3 Synchronization delay of computing power coordination instructions: the physical time difference between when computing power instructions are detected and when the power duty cycle changes substantially, the lower the better; Consensus Iteration Count: The average number of algorithm iterations required to reach a safe solution for multi-objective scheduling, which measures the convergence computational cost of the game optimization module.

[0059] Table 1. Performance comparison of different methods on the Open-Compute-Power-2025 and Jereh-AI-Grid-2026 datasets. The experimental results are shown in Table 1. Figure 2 , Figure 3 and Figure 1 As shown, by comparing and analyzing the actual test data of each method in Table 1 on the dual dataset, the technical advantages of this invention in complex microgrid scheduling are verified: First, in scenarios with extremely high computing power surges, the feedforward voltage regulation capability of this invention demonstrates a significant advantage. As shown in Table 1, when dealing with step currents of up to several hundred amperes generated during large model training (Jereh-AI-Grid-2026 dataset), traditional PID / Droop control, due to its physical lag, experiences a maximum voltage drop of up to 185mV, severely threatening the safe operating threshold of the NPU core. While LSTM-Power introduces prediction, it is limited by the long inference latency (285μs) of the timing model, often resulting in "predicting but not being able to execute in time." This invention, through information entropy manifold mapping in step S1 and PWM duty cycle feedforward generation in step S4, compresses the synchronization delay to 12μs, forcing the inductor to store energy before the current actually surges, successfully controlling the maximum voltage drop to 62mV, and improving the voltage regulation performance by approximately 66.5% compared to the traditional hardware baseline.

[0060] Secondly, Hamiltonian routing demonstrates extremely high resource utilization efficiency when dealing with global spatial power allocation. While the purely data-driven MADDPG model optimizes the "stranded power ratio" to 15.6% for large-scale clusters, its lack of inherent awareness of physical cable impedance still makes it prone to localized power deadlocks. This invention, based on the Hamiltonian dynamics equations constructed in step S2, deeply couples the physical conductance matrix with node states, allowing power to automatically bypass high-impedance hotspots and flow to power-starved computing nodes, much like water flowing. This rigorous physical topology optimization reduces the stranded power ratio to 5.2% in high-density scenarios, utilizing almost every inch of power redundancy within the server rack.

[0061] Finally, the game theory mechanism greatly preserves effective computing power under extreme constraints and demonstrates a generational architectural advantage in terms of extremely fast convergence. The phenomenon in Table 1 where some benchmark models show N / A for the "Number of Consensus Iterations" metric reflects the differences in the underlying scheduling architecture. Traditional PID / Droop control is a purely hardware-level analog / digital passive response, lacking the concept of multi-agent negotiation at the software layer; therefore, this metric is inapplicable (N / A). Similarly, LSTM-Power, as a unidirectional timing prediction model, only issues instructions in one direction (predicting computing power and allocating power), lacking a "bargaining" process between computing and power, and thus also lacking game theory iteration characteristics. When rack-level power capping is triggered, both methods can only crudely reduce the frequency of all chips, causing the effective computing power throughput retention rate to drop to 65.0%-72.4%.

[0062] In contrast, while the MADDPG model, which incorporates intelligent allocation, attempts to optimize policies through multiple agents, its action space in reinforcement learning is too large, making convergence extremely difficult within microseconds (iterations > 800), thus failing to meet the nanosecond-level synchronization requirements of real-world hardware. Step S3 of this invention employs a unique Nash double-blind game architecture. Due to the strict physical feasible region truncation applied during partial derivative optimization (i.e., searching for solutions only within the actual hardware dead zones of the power supply and chip), this method converges to a safe consensus in approximately 32 rapid internal iterations. Under the premise of absolutely avoiding thermal safety red lines, the computing power throughput retention rate under extreme conditions is maintained at a high level of 91.8%. This fully demonstrates that the "multi-objective flexible adaptation" proposed in this invention possesses high robustness and microsecond-level feasibility in industrial applications of computing-power collaboration.

Claims

1. A computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology, characterized in that, Includes the following steps: Step S1. Obtain relevant data information of computing power nodes, extract multidimensional heterogeneous features of computing power instructions, and obtain high-dimensional computing power feature tensors through attention processing; transform the high-dimensional computing power feature tensors into continuous information entropy manifolds, and construct continuous spatiotemporal power demand potential fields; extract features from the spatiotemporal power demand potential fields to obtain the initial transient power compensation benchmark. Step S2. The initial transient power compensation reference state is transformed into the energy state variable in the Hamiltonian system to obtain the global phase space state matrix. A global energy assessment functional is constructed, and the physical damping term simulating the actual heat loss of the cable is reconstructed to obtain the global optimal power routing and allocation tensor. Step S3. Using the global optimal power routing and allocation tensor as the trump card of the power supply agent, construct a multi-agent Nash game framework. By constructing mutually constraining utility functions and using the alternating gradient ascent algorithm, a consensus is finally reached at a dynamic Nash equilibrium point, and the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility is output. Step S4. Reduce the dimension of the synchronization adaptation instruction tensor and decouple it into an independent physical current reference for each power supply phase to obtain the specific target current of each phase inductor, calculate the target voltage command of the node and the corresponding digital quantization code, and write the obtained digital quantization code into the power management main control chip; differentiate the lead time of the target current to calculate the microsecond-level feedforward pulse width modulation duty cycle that is forcibly applied to each phase power transistor. The optimal function is established based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and a hardware execution structure is constructed for synchronous distribution to trigger the hardware fence to achieve synchronous power calculation.

2. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, The process of obtaining the global phase space state matrix in step S2 is as follows: Based on the electrical connection relationship of the underlying real power distribution network, a microgrid impedance topology diagram is constructed. For any two physically connected computing nodes in the diagram, a topological connectivity conductance weighting coefficient for computing nodes is defined. The expected power distribution landing point of the computing node is defined as the generalized position, and the transient driving force of power change is defined as the generalized momentum. The generalized position and the generalized momentum are tightly bound together to construct the fused phase space state tensor of the node, thereby obtaining the global phase space state matrix containing the states of all nodes.

3. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, In step S2, the specific details are as follows: Each sub-tensor in the global phase space state matrix is ​​automatically unpacked, and its generalized momentum and generalized position are separated to participate in the calculation. Finally, through dynamic evolution, the global phase space state matrix is ​​driven to slide towards the steady-state direction that minimizes the scalar value of the global energy assessment functional. By introducing a physical damping term to simulate the actual heat loss of cables, the state tensor is evolved by forward integration using an ordinary differential equation solver, and finally the globally optimal power routing and allocation tensor is reconstructed.

4. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, In step S3, the specific process within the policy space of the multi-agent Nash game framework is as follows: The multi-agent system includes a power supply agent and a computation agent. The Nash equilibrium gap is introduced as a convergence measure. When the power supply agent and the computation agent cannot obtain substantial utility improvement even if they modify their strategies, the system is considered to have reached the Nash equilibrium point.

5. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, After the game converges, the extracted new strategy tensor generated by the power supply side and the new strategy tensor generated by the computing side represent the maximum concession and optimal choice made by the power supply side and the computing side under the current stringent physical constraints, respectively. The game result is transformed into control code that can be directly recognized by the underlying driver. Pareto optimal fusion mapping is performed on the bilateral steady-state tensor to generate the final synchronization adaptation instruction tensor.

6. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, In the decoupling process described in step S4, real-time thermal health weights for each phase are introduced for dynamic load balancing. The specific process is as follows: First, by using a linear projection matrix and a bias activation function, the absolute value of the total physical current demand of the nodes in the next very short time is extracted from the high-dimensional tensor. Under the premise of meeting the absolute value of the total physical current demand, the system will automatically allocate the current load to the phase with low temperature and sufficient thermal margin to obtain an independent expected current reference scalar. Based on the output independent expected current reference scalar, the specific target current of each phase inductor is determined.

7. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, Step S1 involves constructing a continuous spatiotemporal potential field for electricity demand, as detailed below: Using impulse functions and thermal diffusion kernels from physics, discrete computation timestamps are continuously mapped onto the real physical time axis and anchored in three-dimensional physical space.

8. The computational-electricity collaborative scheduling method based on entropy manifold mapping and energy topology according to claim 1, characterized in that, The calculation process for the microsecond-level feedforward pulse width modulation duty cycle in step S4 is as follows: ， in, For nodes The Phase circuit in The per-unit value of the microsecond-level feedforward pulse width modulation duty cycle that should be output at any given moment; This is the input voltage to the upstream main bus. The steady-state target voltage; It is a feedforward forcing term; This is the nominal value of the output filter inductance for this phase; This is the first-order time derivative obtained for the feedforward predicted current; For inductor DC resistance voltage drop compensation; The proportional-integral-differential operator represents the error between the actual sampled voltage and the target voltage. Feedback adjustment.

9. A computational-power collaborative scheduling system based on entropy manifold mapping and energy topology, applied to the computational-power collaborative scheduling method based on entropy manifold mapping and energy topology as described in claim 1, characterized in that, include: The unified representation module for computing power tasks acquires relevant data information of computing power nodes, extracts multi-dimensional heterogeneous features of computing power instructions, and obtains high-dimensional computing power feature tensors through attention processing. The high-dimensional computing power feature tensor is transformed into a continuous information entropy manifold, and a continuous spatiotemporal power demand potential field is constructed. Features are extracted from the spatiotemporal power demand potential field to obtain the initial transient power compensation benchmark. The Hamiltonian potential energy mapping and energy optimization module transforms the initial transient power compensation reference state into energy state variables in the Hamiltonian system, obtains the global phase space state matrix, constructs a global energy evaluation functional, reconstructs the physical damping term simulating the actual heat loss of the cable, and obtains the global optimal power routing and allocation tensor. The game-theoretic decision-making adaptation module uses the global optimal power routing and allocation tensor as the trump card of the power supply agent to construct a multi-agent Nash game framework. By constructing mutually constraining utility functions and using the alternating gradient ascent algorithm, a consensus is finally reached at a dynamic Nash equilibrium point, and the final synchronization adaptation instruction tensor that takes into account both computing power performance and grid flexibility is output. The microsecond-level feedforward control module reduces and decouples the synchronization adaptation instruction tensor into independent physical current references for each power supply phase, obtains the specific target current of each phase inductor, and calculates the target voltage command and corresponding digital quantization code of the node. By differentiating the lead time of the target current, the microsecond-level feedforward pulse width modulation duty cycle forcibly applied to each phase power transistor is calculated. The optimal function is established based on the microsecond-level feedforward pulse width modulation duty cycle to obtain the clock frequency constraint value, and a hardware execution structure is constructed for synchronous distribution to trigger the hardware fence to achieve synchronous power calculation.

10. The computational-electricity collaborative scheduling system based on entropy manifold mapping and energy topology according to claim 9, characterized in that, In the Hamiltonian potential energy mapping and energy optimization module, a global phase space state matrix module is constructed. Based on the electrical connection relationship of the underlying real distribution network, a microgrid impedance topology diagram is constructed. For any two physically connected computing nodes in the diagram, a topological connectivity conductance weighting coefficient for computing nodes is defined. The expected power allocation landing point of the computing node is defined as the generalized position, and the transient driving force of power change is defined as the generalized momentum. The generalized position and the generalized momentum are tightly bound together to construct the fused phase space state tensor of the node, thereby obtaining a global phase space state matrix containing the states of all nodes.

Images