A multi-objective sustainable optimization regulation method for wide-area heterogeneous data centers

Abstract

This invention belongs to the field of network resource scheduling technology and discloses a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers. It addresses the technical problem of traditional single-cluster scheduling lacking cross-domain global orchestration capabilities. The solution includes: collecting heterogeneous resource states, dynamic energy signals, and task constraints to construct a scheduling state space; performing dynamic statistical normalization to eliminate fluctuation interference, fusing multi-scale time-series components and node topology identities to generate a fused embedded state matrix; quantifying matching degree based on a multi-head cross-attention mechanism, extracting spatial aggregation features, and introducing residual direct-connection bypasses and deep concatenation with time-series features to generate a global spatiotemporal context vector; inputting a policy network to extract cross-domain resource allocation decision representations and outputting a probability distribution of heterogeneous task scheduling actions; constructing a dynamically normalized composite reward based on grid carbon emissions, electricity costs, and service quality constraints, extracting real-time state variance adaptive adjustment strategy pruning thresholds, and iteratively optimizing network parameters.

Description

Technical Field

[0001] This invention relates to the field of computer network technology, and specifically provides a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers. Background Technology

[0002] With the rapid development of the Internet of Things, 5G mobile communication, and artificial intelligence technologies, the global data volume is increasing exponentially, and large-scale data centers have become the core infrastructure for carrying massive computing tasks. However, the expansion of computing scale is accompanied by a sharp increase in data center energy consumption and carbon emissions. Under the "dual-carbon" strategy, the new power system is accelerating its evolution towards "source-grid-load-storage coordinated interaction." While the widespread access of distributed renewable energy provides abundant green electricity, it also introduces strong randomness and uncertainty, leading to complex high-frequency fluctuations in the carbon emission intensity on the grid side and electricity market prices. To address these challenges, building a cloud-edge collaborative architecture based on widely distributed heterogeneous data centers, fully exploring the periodic tidal fluctuations of computing load in the time dimension, and coordinating the use of spatial differences in environment and energy supply to carry out flexible spatiotemporal cross-domain scheduling of computing load has become a key path to achieve green and low-carbon computing.

[0003] To achieve effective management of data center computing load, existing technologies have proposed a series of resource scheduling methods. Early resource scheduling methods mainly relied on static rules and heuristic algorithms. These methods monitor the CPU and memory utilization of servers and allocate computing tasks to less loaded computing nodes in order to shorten the overall task completion time. With the continuous expansion of computing scale, researchers have introduced mathematical models such as mixed-integer linear programming to solve for the theoretically optimal solution of task allocation under specific physical resource constraints. For the complex resource allocation problem in large-scale clusters, metaheuristic search strategies such as genetic algorithms and particle swarm optimization have been widely used, achieving a preliminary trade-off between energy consumption and execution latency within the cluster through approximate solution mechanisms. In recent years, deep reinforcement learning technology, with its advantages of online adaptation and end-to-end decision-making, has been gradually applied to resource scheduling in dynamic environments. When dealing with discrete server selection and task offloading problems, value function-based algorithms and their improved variants (such as deep Q-networks and dual deep Q-networks) have been widely used. These methods utilize neural networks to fit action value functions, achieving a preliminary trade-off between system energy consumption and task waiting time within a single data center. To address the continuous demands for computing power and bandwidth resource allocation, algorithms such as deep deterministic policy gradient and flexible actor-commentator architectures based on the actor-commentator framework have been introduced. These algorithms combine experience replay with continuous policy gradient mechanisms to enhance the exploration capability and robustness of resource control strategies in complex continuous action spaces. Furthermore, to improve the training stability of models in strongly stochastic environments, algorithms such as proximal policy optimization, by limiting the policy update step size, have been applied to handle power flow constraints and the dynamic prioritization of computational tasks. These deep reinforcement learning techniques adaptively adjust resource allocation actions through real-time interaction between the agent and the cluster environment. Their core objective is primarily to maximize the utilization efficiency of the underlying hardware resources of a single cluster or to ensure the quality of basic business services.

[0004] Although the above-mentioned resource scheduling technologies have improved the system's operating efficiency to some extent, they still have the following obvious technical shortcomings when facing highly dynamic and strongly coupled cross-domain green and low-carbon scheduling scenarios:

[0005] 1. Lack of global collaborative orchestration capabilities in cross-domain distributed data centers. Existing container orchestration systems and scheduling frameworks are mostly limited to resource management within a single cluster or isolated data center, lacking comprehensive utilization of spatial environmental heterogeneity. When local data centers face computing resource congestion or periods of high carbon intensity, existing systems cannot effectively assess the wide area network communication overhead of cross-regional data transmission, making it difficult to offload computing loads across domains to remote nodes with abundant green energy or low electricity costs. This leads to low utilization of wide-area computing resources, forming a structural carbon emission bottleneck that is difficult to overcome.

[0006] 2. Accurate characterization and dependency modeling of heterogeneous states in complex spatiotemporal coupled environments are challenging. Widely distributed data centers exhibit significant heterogeneity in hardware architecture, topology, and environment. Existing scheduling methods typically employ simple statistical indicators, statically slicing and directly concatenating state characteristics. These methods fail to effectively decouple the periodic tidal fluctuations of computing load over time and fail to establish accurate dependency models. Due to the lack of long-sequence time dependency extraction between business tidal patterns and time-varying carbon emission intensity, existing models cannot explicitly distinguish between homomorphic but heterogeneous physical nodes with the same resource utilization but different underlying hardware energy efficiency. This leads to severe biases in the scheduling model's perception of complex spatiotemporal evolution environments, making it difficult to accurately capture the nonlinear coupling relationship between task resource attributes and dynamic environmental signals.

[0007] 3. Lack of a green, low-carbon, multi-objective combined optimization mechanism for sustainable development scheduling. In cross-domain cloud-edge collaborative scenarios, traditional resource scheduling algorithms mostly focus on maximizing the utilization rate of resources within the cluster. These algorithms primarily focus on the physical occupancy of underlying hardware resources such as central processing units, graphics processing units, and memory, or solely on ensuring the quality of business services as a single optimization orientation. However, existing methods severely isolate the interaction between the computational physical system and the external energy environment, failing to incorporate the real-time carbon emission intensity of the power grid and the high-frequency fluctuations in electricity prices in the electricity market into the scheduling decision framework. Due to the lack of joint modeling between carbon emission economic costs and computing power benefits, existing scheduling strategies cannot proactively shift computing loads to nodes with low carbon emissions and low electricity prices while ensuring the quality of business services. This leads to data centers facing severe risks of exceeding carbon emission limits and high energy costs during long-term operation, failing to meet the core requirements of green and sustainable scheduling under the new power system.

[0008] In summary, existing single-cluster scheduling mechanisms and optimization strategies that only focus on the utilization of underlying resources are severely detached from the physical constraints of the external energy environment and cannot meet the green and low-carbon scheduling requirements of data centers under new power systems. Therefore, there is an urgent need to provide a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers. This method should solve the problem of cross-domain resource allocation in dynamic environments by modeling the correlation and dependency of high-dimensional temporal characteristics and combining multi-objectives collaborative optimization, so as to achieve a long-term balance between system business efficiency and green sustainability. Summary of the Invention

[0009] The purpose of this invention is to solve the technical problems of traditional single cluster scheduling lacking cross-domain global orchestration capabilities, and being unable to accurately characterize heterogeneous states in non-stationary energy environments while taking into account the policy update oscillations, feature distribution shifts, and difficulties in green and low-carbon collaborative scheduling caused by low-carbon multi-objective optimization.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention proposes a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers, comprising:

[0012] S1: Collect heterogeneous resource status, external dynamic energy signals, and scheduling task constraints from a wide-area, multi-data center. The heterogeneous status data includes node resource availability, network transmission status, and dynamic signals such as grid carbon emission intensity, electricity price, and temperature. The task constraints aggregate heterogeneous resource requirements and deadlines, and combine this with underlying hardware power to estimate the expected energy consumption of the task. Based on the above multi-source collected data, a multi-dimensional scheduling state space is constructed, including electricity costs, carbon emission intensity, and service quality assurance, establishing a global perception boundary for cross-domain collaborative scheduling for the intelligent agent.

[0013] S2: Online statistical algorithms are used to dynamically normalize and preprocess the collected heterogeneous state data and the feature data of the tasks to be scheduled, extracting real-time state variance to eliminate interference from non-stationary fluctuations in the external environment and obtain standardized system state features. Multi-scale time decomposition embedding and topological identity embedding are then applied to these system state features to decouple the tidal fluctuations of the workload from the inherent physical properties of heterogeneous nodes. Subsequently, nonlinear fusion of temporal evolution features and node topological attributes is used to construct a data center fusion-embedded state matrix to map the dynamic changes of heterogeneous resources, thereby overcoming the feature distribution shift problem of static feature extraction mechanisms in dynamic environments.

[0014] S3: Based on the aforementioned standardized task feature data and the data center fusion state matrix, a multi-head cross-attention mechanism is used to quantify the physical matching degree between task requirements and node capabilities, and to extract spatial attention aggregation vectors. Furthermore, residual direct-connection bypass and spatiotemporal global semantic information are introduced for fusion. While preserving the spatial supply-demand matching representation capability without loss, temporal evolution semantics are successfully injected, generating feature vectors containing global spatiotemporal context, thus avoiding gradient decay and semantic loss problems in the deep feature fusion process.

[0015] S4: For feature vectors containing global spatiotemporal context, a policy network of reinforcement learning agent is constructed to extract deep decision representations related to cross-domain resource allocation. Based on the deep decision representations, action log probabilities are generated, and the action probability distribution for task scheduling in heterogeneous data centers is output. This completes the end-to-end accurate mapping from the feature space of high-dimensional heterogeneous spatiotemporal environment to the action space of low-dimensional cross-domain scheduling.

[0016] S5: The sampling output scheduling actions are distributed to a multi-datacenter environment for iterative interaction to obtain multi-objective composite rewards and the observation results of the next time step. Based on the extracted real-time state variance, the adaptive pruning threshold is dynamically adjusted, and the parameter update boundary is combined with the composite reward constraint strategy optimization. This enables the agent to respond efficiently to external non-stationary disturbances, effectively avoid policy collapse in high-fluctuation environments, and ultimately achieve robust multi-objective collaborative optimization of carbon emissions, electricity costs, and service quality.

[0017] The detailed description of the heterogeneous state space construction in S1 of the above method is as follows:

[0018] S11: Based on historical regional data, linear interpolation is performed on the real-time monitored signals of power grid carbon emission intensity, electricity price, and ambient temperature in the region where the data center node is located. This is then superimposed with a normally distributed coherent noise simulation to construct physically realistic scheduling constraints and generate the current external environment data.

[0019]

[0020] In the formula, , and These represent data center nodes. At any moment Real-time grid carbon emission intensity, electricity price, and ambient temperature; For regional historical datasets; Indicates the current moment; This represents a linear interpolation function; Indicates that it follows the mean. variance is Normally distributed coherent noise.

[0021] S12: Due to the heterogeneous differences in computing power and configuration among widely distributed heterogeneous data center nodes, the total amount of various computing resources and current load of each data center node are extracted, and the availability of computing resources is normalized.

[0022]

[0023] In the formula, Indicates time Data center node The normalized resource availability vector; , , Representing nodes respectively Total processor core capacity, total graphics card capacity, and total memory capacity; , , Representing nodes respectively At any moment The processor core load, graphics card load, and memory load.

[0024] S13: Since a single instantaneous resource utilization rate cannot fully reflect the queuing backlog of tasks, the remaining execution time of tasks in the running queue, the estimated total duration of tasks in the waiting queue, and the core demand of each task are normalized in combination with the total core processing scale of the cluster to calculate the load backlog time estimation characteristics. This quantifies the time overhead of each node in processing historical backlogged tasks and the dynamic accumulation of computational load.

[0025]

[0026] In the formula, Indicates the run queue; Indicates a waiting queue; Indicates tasks in the run queue The expected end time; Indicates the current moment; Indicates tasks in the run queue The core demand; Indicates tasks in the waiting queue The estimated total duration; Indicates tasks in the waiting queue The core demand; Represents a node The total core processing capacity of the cluster; This represents a system-preset minimum positive number to prevent the denominator from being zero.

[0027] S14: To further quantify the network communication overhead introduced by cross-domain resource allocation, calculate the network round-trip time between the source node and the target node, and retrieve the system's preset maximum delay constant. Calculation time... Data center node Normalized transmission delay characteristics To quantify the current level of network congestion and transmission overhead:

[0028]

[0029] In the formula, and These represent the source node and the destination node for data transmission, respectively. This represents the actual round-trip time between the source node and the target node; This represents the maximum delay constant preset by the system.

[0030] S15: The resource availability vector, load backlog time estimation features, and transmission delay features extracted above are concatenated and fused with real-time carbon intensity, electricity price, and ambient temperature to construct the data center nodes. At any moment Initial resource state vector :

[0031]

[0032] In the formula, Represents a normalized resource availability vector; Indicates the carbon emission intensity of the power grid; Indicates electricity price; Indicates ambient temperature; This indicates the characteristics of load backlog time estimation; This indicates transmission delay.

[0033] S16: Decouple the server's total power consumption into a base static power consumption independent of the load and a dynamic power consumption that is linearly positively correlated with processor utilization. Combine the estimated execution time of the scheduled task, the number of processor cores and GPUs requested, and the unit dynamic power consumption coefficient of the corresponding hardware unit with the server's base static power consumption to comprehensively calculate the scheduled task. Projected energy consumption :

[0034]

[0035] In the formula, Indicates tasks to be scheduled The estimated execution time; and These represent the tasks to be scheduled. The requested number of processor cores and graphics cards; and These represent the unit dynamic power consumption coefficients for the processor unit and the graphics card unit, respectively. This indicates the server's base static power consumption.

[0036] S17: Collect the heterogeneous resource requirements, estimated execution time, and remaining deadline of the tasks to be scheduled, and concatenate them with the aforementioned estimated energy consumption to construct the time-sharing data. Tasks to be scheduled initial feature vector :

[0037]

[0038] In the formula, , , These represent the tasks to be scheduled. The requested number of processor cores, number of graphics cards, and memory size; Indicates the estimated execution time; Indicates tasks to be scheduled Service quality remaining deadline constraints; This indicates the projected energy consumption.

[0039] In the above method, the fine-grained embedding characterization in S2 is described in detail below:

[0040] S21: In cross-domain scheduling scenarios, state features exhibit significant dimensional differences and non-stationary distribution characteristics. Normalization methods using fixed statistics are prone to data distribution shifts when dealing with electricity price spikes or sudden load surges. Therefore, for new observations of input features at the current time step, incremental recursive calculations are performed using the feature mean and second-order central moments from the previous time step. This achieves numerically stable real-time statistical updates without requiring the storage of all historical data.

[0041]

[0042]

[0043] In the formula, and These represent the system at time steps. With time step The characteristic dynamic mean; Indicates time step New observations of the arriving input features; Indicates the current time step index; and These represent the system at time steps. With time step The characteristic second-order central moment.

[0044] S22: Abnormal observations in a dynamic scheduling environment can easily interfere with the network gradient, leading to extreme instability in the training convergence process of the agent. Therefore, based on the incrementally updated second-order central moments of the features in the aforementioned steps, the dynamic variance estimate is calculated. The new observations are then standardized using Z-score, and a threshold constraint is applied to the standardized features using a truncation function. This eliminates the interference of abnormal extreme values ​​on the gradient, transforming the initial resource state vector obtained in S1 into a stable standard system state feature. :

[0045]

[0046] In the formula, This represents a cutoff function used to limit the range of features; This represents the dynamic variance estimate calculated based on the characteristic second-order central moments, and ; This represents a very small positive number preset by the system to prevent the denominator from being zero; and These represent the upper and lower threshold values ​​set by the truncation function, respectively.

[0047] S23: The computing load of the data center physical cluster and the carbon emission intensity of the power grid in the region exhibit a complex periodic pattern of intraday fluctuations combined with weekly trends. A single linearly increasing time step cannot directly reflect this time series evolution cycle. Therefore, the global discrete time step of the system needs to be considered. Perform orthogonal decomposition to extract hourly components representing intraday cycles. With the weekday component representing the weekly cycle This decouples continuous linear timescales into multi-granularity periodic temporal index coordinates:

[0048]

[0049] S24: The one-hot encoded vectors corresponding to discrete time indices are mutually orthogonal and highly sparse in the feature space, making it difficult to express the temporal correlation similarity between different time nodes. Therefore, independent hour embedding matrices and day-of-week embedding matrices are constructed. The one-hot encoded vectors corresponding to the extracted hour and day-of-week components are used as input, and they are projected onto a high-dimensional continuous latent space through matrix multiplication to extract independent periodic features with dense expressive power. Subsequently, periodic features at different scales are concatenated and fused. This parameterized mapping process introduces a strong inductive bias in the feature space, forcing the scheduling model to share the same hourly feature subspace at the same time on different dates, thus extracting a unified temporal feature vector.

[0050]

[0051] In the formula, This represents the unified temporal feature vector generated after mapping and fusion; and These represent hourly components respectively. With weekday portion Perform an index query in the corresponding feature dictionary to extract the embedded feature representation; This represents a vector concatenation operation; and Representing time indexes respectively and The corresponding one-hot encoded vector after conversion; and These represent the continuously optimizable hour embedding matrix and weekday embedding matrix, respectively.

[0052] S25: Assign a unique learnable topology embedding vector as a conditional context to each data center node, and combine it with the aforementioned standardized real-time state vector. By jointly representing the state projection matrix and the nonlinear activation function GeLU, the inherent topological physical attributes of nodes are injected while encoding time-varying load information. This explicitly maps the hardware specifications and geographical differences of heterogeneous resources, thereby generating a data center converged state matrix with data physical origin awareness capabilities. This addresses the common phenomenon of homomorphic but heterogeneous cross-domain heterogeneous clusters having the same resource occupancy status but different hardware specifications or geographical locations.

[0053]

[0054] In the formula, Represents data center node High-dimensional feature representation; GeLU represents the nonlinear activation function. Represents the state projection matrix; Represents data center node At any moment The standardized real-time state vector; Represented as data center node The unique learnable topological embedding vector assigned.

[0055] The multi-head cross-attention matching mechanism in S3 of the above method is described in detail below:

[0056] S31: Single feature concatenation cannot capture the nonlinear coupling relationship between multi-dimensional trade-offs such as computing power adaptability and carbon emission economy. To accurately extract the scheduling matching degree between specific task requirements and heterogeneous data center supply capacity, the extracted features of the task to be scheduled are constructed into a query vector, and the data center fusion state representation matrix is ​​constructed into key and value vectors. An asymmetric cross-domain projection operation is performed using independent learnable projection parameter matrices to map them to multiple independent semantic subspaces, and the query vector, key vector, and value vector corresponding to each attention head are calculated respectively.

[0057]

[0058] In the formula, , , They represent the first The query vector, key vector, and value vector of each attention head; This represents the query vector obtained after linear projection of the task features; The matrix represents the converged state of the data center. , , These represent the learnable projection parameter matrices corresponding to the query, key, and value, respectively.

[0059] S32: Given that direct dot product of high-dimensional vectors in the semantic subspace easily leads to numerical extrema and gradient vanishing, and that hard matching mechanisms are difficult to flexibly adjust decisions according to dynamic environments, this invention introduces a scaling factor and a temperature adjustment coefficient. This coefficient is used to perform a scaling dot product operation on the query vector and key vector corresponding to each attention head, generating an attention score vector that measures the alignment between task requirements and the capabilities of each node. After normalization, the value vectors are weighted and aggregated to generate the output feature vector of each attention head.

[0060]

[0061] In the formula, Indicates the first The output feature vector of each attention head; This represents the normalized exponential function; The transpose matrix of the key vector; Represents the feature dimensions of each attention head; Indicates the scaling factor; This represents the temperature regulation coefficient, used to smooth the distribution of attention scores to adjust the exploratory nature of the scheduling strategy.

[0062] S33: Next, the feature vectors output by all attention heads are concatenated, and linear transformation and layer normalization are performed through the multi-head output projection matrix. This aggregates the task and node matching information scattered in different semantic subspaces, generating a spatial attention aggregation vector that provides a global perspective on scheduling supply and demand matching representation.

[0063]

[0064] In the formula, Represents the multi-head space cross-attention aggregation vector; Indicates the layer normalization processing function; This represents a vector concatenation operation; express Output feature vectors of different attention heads; This represents the multi-head output projection matrix.

[0065] S34: The spatial cross-attention aggregation vector generated above is fused with the multi-scale temporal embedding vector described in S2, and input into the feedforward projection network for nonlinear extraction. The residual of the spatial attention aggregation vector is connected to the network output. While retaining the original spatial matching expressive ability without loss, spatiotemporal global semantics are injected to generate an embedding representation containing a complete global spatiotemporal context.

[0066]

[0067] In the formula, Represents a multi-scale temporal embedding vector; This represents a vector concatenation operation; and This represents the weight parameter matrix of the feedforward projection network; and This represents the bias term of the feedforward projection network; Represents the nonlinear activation function ReLU.

[0068] The detailed description of the cross-domain resource allocation decision action generation in S4 of the above method is as follows:

[0069] S41: The previously generated feature vector containing the global spatiotemporal context Inputting a feature extraction layer of a policy network based on a multilayer perceptron architecture, and performing forward propagation processing using a nonlinear activation function, yields a deep policy decision representation related to cross-domain resource allocation. :

[0070]

[0071] In the formula, Represents the ReLU activation function; The policy network contains parameters. Multilayer perceptron feature extraction layer;

[0072] S42: A fully connected layer is used to map the extracted deep policy decision representation into action log probabilities. A normalized exponential function is introduced to transform these action log probabilities into a standardized stochastic policy distribution, generating a state distribution for the system across heterogeneous data centers. Take action below The probability distribution of task scheduling actions provides a decision space for the random sampling of subsequent scheduling actions:

[0073]

[0074] In the formula, This represents the normalized exponential function; This represents the weight parameter matrix corresponding to the output layer of the policy network; This represents the deep decision-making process of the strategy. This represents the bias term parameters corresponding to the output layer of the policy network.

[0075] S43: The gradient updates of the policy network alone have extremely high variance. To ensure stable convergence of model training, a value network is constructed to assist in evaluating the decision benchmark of the policy, and feature vectors containing the global spatiotemporal context are used. The feature extraction layer of a multi-layer perceptron within the synchronous input value network architecture is combined with a non-linear activation function for feature dimensionality reduction, generating a deep value representation for state value assessment.

[0076]

[0077] In the formula, Represents the ReLU activation function; This indicates that the value network contains parameters. Multilayer perceptron feature extraction layer;

[0078] S44: Deep Value Representation Based on the Foregoing Generation By performing an affine transformation on the linear output layer of the value network, the current system state for estimating multi-objective scheduling is obtained. Scalar value benchmark for expected cumulative return This provides a quantitative evaluation basis for calculating the advantage function and reducing the policy gradient variance:

[0079]

[0080] In the formula, This represents the weight parameter matrix corresponding to the output layer of the value network; This represents the deeper meaning of value; This represents the bias term parameters corresponding to the output layer of the value network.

[0081] In the above method, the detailed description of the composite reward construction and network parameter iterative optimization in S5 is as follows:

[0082] S51: The carbon emission intensity of the power grid and electricity market prices exhibit non-stationary distributions and high-frequency spikes. Directly calculating rewards using absolute values ​​can easily lead to reward distribution shifts and strategy update oscillations. Based on the projected energy consumption of the task at the current moment and the real-time carbon emission intensity and electricity price of the power grid, initial carbon emission penalties and initial energy cost penalties are calculated respectively. Subsequently, the pre-maintained rolling mean and standard deviation are extracted, and dynamic standardization is performed on the initial penalty terms to generate standardized carbon emission rewards and standardized energy cost rewards that eliminate environmental dynamic interference.

[0083]

[0084]

[0085] In the formula, Indicates time step The calculated initial penalty for carbon emissions; This represents the set of tasks currently being executed in the corresponding data center; Indicates the task index in the set; Indicates task The projected energy consumption; Indicates the corresponding data center at time step Real-time grid carbon emission intensity; Indicates standardized carbon emission rewards; and These represent the rolling mean and standard deviation of the initial carbon emission penalty, respectively. This represents a very small positive number preset by the system to prevent the denominator from being zero; Indicates time step The calculated initial penalty term for energy consumption cost; Indicates the corresponding data center at time step Real-time electricity prices; Indicates a standardized energy consumption cost incentive; and These represent the rolling mean and standard deviation of the initial penalty term for energy consumption costs, respectively.

[0086] S52: Extract the actual completion time of the task and the system-required deadline, calculate the hard constraint penalty term with respect to the time threshold using an indicator function, trigger a fixed physical penalty when the actual completion time of the task exceeds the deadline limit, and generate a service quality penalty reward for the guided strategy network to avoid breach of contract. :

[0087]

[0088] In the formula, This indicates the preset physical penalty coefficient for breach of contract. This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Indicates the actual completion time of the task; This indicates the deadline specified for the task.

[0089] S53: Combining the scheduling strategy with the differences in the degree and scale of preference for different objectives, a weighted summation operation is performed on the standardized carbon emission reward, standardized energy consumption cost reward, and service quality penalty reward obtained above, to construct a scalarized multi-objective composite reward, guiding the agent to approach the Pareto optimal frontier in a complex policy space with multiple conflict constraints:

[0090]

[0091] In the formula, This represents the multi-objective composite reward calculated at the current time step; , and These represent the carbon emission target weight coefficient, energy consumption cost target weight coefficient, and service quality target weight coefficient configured in the weight vector, respectively.

[0092] S54: Cross-domain resource scheduling tasks suffer from significant long-term temporal dependencies and reward sparsity; relying solely on single-step rewards cannot accurately assess the long-term physical impact of actions. Based on the aforementioned constructed multi-objective composite reward, the system's preset discount factor, and the current and next-time scalar value benchmarks output by the value network, single-step temporal difference errors are calculated. Using the exponentially weighted moving average of multi-step temporal difference errors, a generalized advantage estimate is calculated, thereby quantifying the long-term physical benefit of the current scheduling action relative to the average benchmark.

[0093]

[0094]

[0095] In the formula, Indicates time step Single-step timing difference error; Indicates multi-objective composite reward At time step Instant rewards; Indicates the discount factor; and These represent the value network in relation to the current state. State at the next moment Output scalar value benchmark prediction; This represents the estimate of generalized advantage; Indicates the error attenuation coefficient; This indicates the maximum time step of the round.

[0096] S55: Static policy update pruning thresholds are highly susceptible to policy update oscillations and system scheduling downtime risks in complex industrial environments facing sudden changes in grid carbon emission intensity or drastic fluctuations in node load. To address this, the feature dynamic variance estimate, updated in real-time by an online algorithm during the preprocessing stage described in S1, is extracted. A norm is then taken to characterize the total volatility of the global physical environment. Combined with a preset baseline pruning threshold and a sensitivity adjustment coefficient, an adaptively adjusting dynamic pruning threshold is generated through exponential decay mapping.

[0097]

[0098] In the formula, This represents the dynamic clipping threshold calculated at the current time step; This represents the system's preset baseline clipping threshold; The sensitivity adjustment coefficient represents the degree of aversion of the control strategy to the risk of fluctuations in the physical environment; This represents the dynamic variance estimate of the system environment features extracted in real time; It represents the norm of the dynamic variance estimate of the feature.

[0099] S56: The parameter updates of the value network require an accurate fitting target to guide the convergence of state values. The generalized advantage estimate obtained above is linearly superimposed with the predicted output of the old value network for the current state from the previous iteration to construct the fitting target value of the value network. This target value is used to characterize the true expected cumulative return benchmark of the value network under the current system state. :

[0100]

[0101] In the formula, This represents the estimate of generalized advantage; This indicates that the old value network is related to the current state. Output scalar value benchmark prediction.

[0102] S57: The gradient update process of multi-agent policies requires a multi-dimensional trade-off between maximizing cumulative reward, constraining update step size, and preventing getting trapped in local optima. Based on the ratio of the probability distributions of actions in the old and new policies, combined with the previously calculated dynamic pruning threshold, generalized advantage estimate, fitted target value, and policy distribution entropy, a joint loss function is constructed, including the pruning agent objective, value network loss, and entropy regularization term. An adaptive moment estimation optimizer is used to perform backpropagation on this joint loss function, iteratively updating the policy network parameters and value network parameters until the overall performance of the multi-objective optimization of the agents converges.

[0103]

[0104]

[0105] In the formula, Represents the joint loss function; Represents the empirical averaging operator; This represents the ratio of the probability distributions of actions generated by the old and new strategies. This represents the truncation function; Indicates the dynamic clipping threshold; Indicates the loss of value network, and ; The entropy regularization term represents the policy distribution; and These represent the weight coefficients of the value network loss and the entropy regularization term, respectively. This represents the adaptive moment estimation optimizer function; and These represent the parameters of the policy network and value network in the current round, respectively. Represents the gradient of the joint loss function; Indicates the learning rate; and This indicates the network parameters after the next round of updates.

[0106] Because the present invention employs the above-described technical means, it possesses the following beneficial effects:

[0107] 1) A fine-grained, accurate representation mechanism for heterogeneous states under complex spatiotemporal coupling environments was constructed. By introducing multi-scale time decomposition and topological identity embedding mechanisms, and combining online incremental statistical normalization processing based on dynamic mean and second-order central moments, this invention achieves fine-grained feature decoupling and continuous spatial mapping of high-dimensional heterogeneous states. Compared with the existing bottleneck of traditional static feature extraction methods, which are prone to feature distribution failure due to differences in data dimensions and sudden extreme values ​​when facing periodic tidal fluctuations and non-stationary environmental signals, this mechanism can provide numerically stable global spatiotemporal context input with underlying physical perception capabilities for cross-domain collaborative scheduling, thereby effectively eliminating feature distribution shifts caused by electricity price spikes and load surges.

[0108] 2) A multi-dimensional cross-domain collaborative supply and demand matching architecture adapted to wide-area heterogeneous distributed data centers was designed. By constructing a multi-head cross-attention network architecture with residual direct-connection bypasses, this invention maps task requirements and data center states to independent semantic subspaces to perform asymmetric cross-domain accurate retrieval, and deeply injects temporal evolution semantics while retaining the spatial matching expression without loss. In the cross-domain resource allocation of wide-area heterogeneous clusters, traditional feature splicing methods face the physical game challenge of being unable to accurately assess heterogeneous demand and the supply capacity of remote nodes, which easily leads to matching distortion. This architecture can accurately capture the nonlinear coupling relationship between multi-dimensional trade-off features such as computing power affinity, network transmission overhead, and carbon emission economy. This not only significantly improves the accuracy of resource matching under complex cross-domain constraints, but also effectively overcomes the gradient decay defect during deep network training.

[0109] 3) A dynamic normalized multi-objective collaborative optimization method for non-stationary energy environments is proposed. Based on real-time extracted rolling mean and standard deviation, this invention constructs a composite reward mechanism including dynamic standardized penalties and hard constraint indicator functions. This mechanism dynamically standardizes grid carbon emissions and energy consumption costs, and quantifies service quality breaches as physical penalties of fixed intensity. In sustainable green and low-carbon scheduling, multiple optimization objectives often have vastly different physical dimensions. Existing methods use static weighting, which easily leads to the technical defect of the policy gradient being dominated by a single high-amplitude objective. This invention, through the aforementioned mechanism, can better guide the scheduling agent to smoothly approach the Pareto optimal frontier under multiple conflict constraints, eliminating the reward distribution shift caused by drastic fluctuations in the macroeconomic energy environment, and effectively achieving joint modeling of the external energy environment and the performance of the internal computing system.

[0110] 4) An adaptive policy update mechanism based on physical environment fluctuation perception is introduced. By extracting the first norm of the system feature dynamic variance estimate as an indicator of total environmental volatility, and using exponential decay mapping to generate an adaptively adjusted dynamic pruning threshold, this invention enables the agent to reduce the pruning amplitude for conservative updates when the environment fluctuates drastically, and increase the pruning amplitude to accelerate convergence when it tends to stabilize. Traditional near-end policy optimization algorithms use static pruning thresholds, which are prone to policy update oscillations or even scheduling crashes when faced with sudden changes in grid carbon emissions or drastic fluctuations in node loads. In contrast, this mechanism greatly improves the training robustness and online scheduling execution security of reinforcement learning models in strongly stochastic industrial environments.

[0111] 5) This invention achieves efficient collaborative scheduling of global computing power and green energy in cross-domain heterogeneous data centers. By deeply integrating the latency tolerance elasticity of wide-area computing loads in the time dimension with the spatial differences in new energy supply among heterogeneous data centers, this invention innovatively constructs a new spatiotemporal collaborative scheduling paradigm. This invention breaks through the bottleneck limitations of traditional single-cluster systems that only focus on the physical occupancy of underlying hardware resources and are severely detached from external energy and environmental constraints. Under the premise of strictly guaranteeing service quality such as task deadlines, the system achieves adaptive and precise migration of cross-domain computing resources to low-carbon emission, low-electricity-price nodes, resulting in a substantial reduction in overall carbon footprint emissions and grid power procurement and operating costs, ultimately achieving a long-term balance between data center business efficiency and green, low-carbon sustainability.

[0112] In summary, this invention breaks through the technical bottleneck of traditional single data center resource scheduling, fully integrating advanced technologies such as temporal dependency modeling, multi-head cross-attention mechanisms, environmental fluctuation perception-based adaptive updates, and dynamic multi-objective combinatorial optimization. Through accurate representation of high-dimensional spatiotemporal features and improvements to the underlying physical mechanisms of reinforcement learning algorithms, this invention not only solves the problems of reward shift and policy oscillation in non-stationary energy environments, but also achieves adaptive migration of cross-domain computing resources to areas with a high proportion of green energy. This provides a system-level, sustainable control solution for the green, low-carbon, safe, stable, and cost-effective operation of large-scale heterogeneous data centers. Attached Figure Description

[0113] Figure 1 A schematic diagram of the state space construction of a wide-area cluster in a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers, provided for the implementation of the present invention;

[0114] Figure 2 This is a diagram illustrating the cross-domain resource embedding representation structure in an example of the present invention.

[0115] Figure 3This is a diagram of the architecture of the deep reinforcement learning collaborative scheduling model based on adaptive spatiotemporal embedding constructed in the example of this invention. Detailed Implementation

[0116] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0117] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0118] The purpose of this invention is to address the technical problems of existing single-cluster scheduling mechanisms, such as the lack of global collaborative orchestration capabilities across data centers, the difficulty in accurately representing heterogeneous states and modeling temporal dependencies in complex spatiotemporal coupling environments, the serious detachment of traditional scheduling algorithms from external energy and environmental physical constraints, and the lack of green, low-carbon, multi-objective combined optimization mechanisms for sustainable development scheduling. This invention provides a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers. This method solves the problem of cross-domain resource allocation in non-stationary dynamic environments through high-dimensional temporal dependency modeling and multi-dimensional joint collaborative optimization, thereby achieving a long-term balance between the business efficiency and green sustainability of the computational physical system. This method effectively overcomes the challenges of policy update oscillations and feature distribution shifts caused by non-stationary energy environments, enabling adaptive and precise migration of cross-domain computing loads to areas with abundant green energy. This significantly improves the balanced reuse rate of global computing resources, the high-concurrency throughput stability of massive tasks, and the green, low-carbon sustainability of digital infrastructure.

[0119] This invention proposes a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers, including:

[0120] S1: Obtain the heterogeneous states of multiple data centers across a wide area and the initial resource vectors of tasks to be scheduled, constructing a scheduling state space with electricity cost, carbon emission intensity, and service quality assurance as optimization objectives. The heterogeneous state data includes the resource availability of data center nodes, network transmission status, and dynamic signals of the external environment such as grid carbon emission intensity, electricity prices, and ambient temperature. The feature data of the tasks to be scheduled aggregates the heterogeneous resource requirements and deadline constraints of the tasks, and combines this with the underlying hardware power to calculate the expected energy consumption of the tasks.

[0121] S2: Online statistical algorithms are used to dynamically normalize and preprocess the collected heterogeneous state data and the feature data of the tasks to be scheduled, extracting real-time state variance to eliminate interference from non-stationary fluctuations in the external environment and obtain standardized system state features. Multi-scale time decomposition embedding and topological identity embedding are then performed on these system state features to extract temporal evolution features and node topological attributes. The extracted temporal evolution features and node topological attributes are then nonlinearly fused to map the dynamic changes of heterogeneous resources, constructing a data center fusion embedded state matrix.

[0122] S3: Based on the task feature data and data center fusion state matrix of the aforementioned standardized processing, the matching score of task requirements and node capabilities is calculated through multi-head cross attention mechanism, spatial attention aggregation vector is extracted, and residual direct connection bypass and spatiotemporal global semantic information are introduced for fusion to generate feature vector containing global spatiotemporal context.

[0123] S4: For feature vectors containing global spatiotemporal context, construct a policy network for reinforcement learning agents to extract deep decision representations related to cross-domain resource allocation, and generate action log probabilities based on deep decision representations, outputting the task scheduling action probability distribution for heterogeneous data centers.

[0124] S5: Iteratively interacts the sampled scheduling actions with the multi-datacenter environment to calculate multi-objective composite rewards and the observation results of the next time step. Based on the extracted real-time state variance, the adaptive pruning threshold is dynamically adjusted. The composite reward and adaptive pruning threshold are combined for value assessment and policy network parameter updates, achieving multi-objective collaborative optimization of carbon emissions, electricity costs, and service quality.

[0125] The detailed description of the heterogeneous state space construction in S1 of the above method is as follows:

[0126] S11: Based on historical regional data, linear interpolation is performed on the real-time monitored signals of power grid carbon emission intensity, electricity price, and ambient temperature in the region where the data center node is located. This is then superimposed with a normally distributed coherent noise simulation to construct physically realistic scheduling constraints and generate the current external environment data.

[0127]

[0128] In the formula, , and These represent data center nodes. At any moment Real-time grid carbon emission intensity, electricity price, and ambient temperature; For regional historical datasets; Indicates the current moment; This represents a linear interpolation function; Indicates that it follows the mean. variance is Normally distributed coherent noise.

[0129] S12: Due to the heterogeneous differences in computing power and configuration among widely distributed heterogeneous data center nodes, the total amount of various computing resources and current load of each data center node are extracted, and the availability of computing resources is normalized.

[0130]

[0131] In the formula, Indicates time Data center node The normalized resource availability vector; , , Representing nodes respectively Total processor core capacity, total graphics card capacity, and total memory capacity; , , Representing nodes respectively At any moment The processor core load, graphics card load, and memory load.

[0132] S13: Since a single instantaneous resource utilization rate cannot fully reflect the queuing backlog of tasks, the remaining execution time of tasks in the running queue, the estimated total duration of tasks in the waiting queue, and the core demand of each task are normalized in combination with the total core processing scale of the cluster to calculate the load backlog time estimation characteristics. This quantifies the time overhead of each node in processing historical backlogged tasks and the dynamic accumulation of computational load.

[0133]

[0134] In the formula, Indicates the run queue; Indicates a waiting queue; Indicates tasks in the run queue The expected end time; Indicates the current moment; Indicates tasks in the run queue The core demand; Indicates tasks in the waiting queue The estimated total duration; Indicates tasks in the waiting queue The core demand; Represents a node The total core processing capacity of the cluster; This represents a system-preset minimum positive number to prevent the denominator from being zero.

[0135] S14: To further quantify the network communication overhead introduced by cross-domain resource allocation, calculate the network round-trip time between the source node and the target node, and retrieve the system's preset maximum delay constant. Calculation time... Data center node Normalized transmission delay characteristics To quantify the current level of network congestion and transmission overhead:

[0136]

[0137] In the formula, and These represent the source node and the destination node for data transmission, respectively. This represents the actual round-trip time between the source node and the target node; This represents the maximum delay constant preset by the system.

[0138] S15: The resource availability vector, load backlog time estimation features, and transmission delay features extracted above are concatenated and fused with real-time carbon intensity, electricity price, and ambient temperature to construct the data center nodes. At any moment Initial resource state vector :

[0139]

[0140] In the formula, Represents a normalized resource availability vector; Indicates the carbon emission intensity of the power grid; Indicates electricity price; Indicates ambient temperature; This indicates the characteristics of load backlog time estimation; This indicates transmission delay.

[0141] S16: Decouple the server's total power consumption into a base static power consumption independent of the load and a dynamic power consumption that is linearly positively correlated with processor utilization. Combine the estimated execution time of the scheduled task, the number of processor cores and GPUs requested, and the unit dynamic power consumption coefficient of the corresponding hardware unit with the server's base static power consumption to comprehensively calculate the scheduled task. Projected energy consumption :

[0142]

[0143] In the formula, Indicates tasks to be scheduled The estimated execution time; and These represent the tasks to be scheduled. The requested number of processor cores and graphics cards; and These represent the unit dynamic power consumption coefficients for the processor unit and the graphics card unit, respectively. This indicates the server's base static power consumption.

[0144] S17: Collect the heterogeneous resource requirements, estimated execution time, and remaining deadline of the tasks to be scheduled, and concatenate them with the aforementioned estimated energy consumption to construct the time-sharing data. Tasks to be scheduled initial feature vector :

[0145]

[0146] In the formula, , , These represent the tasks to be scheduled. The requested number of processor cores, number of graphics cards, and memory size; Indicates the estimated execution time; Indicates tasks to be scheduled Service quality remaining deadline constraints; This indicates the projected energy consumption.

[0147] In the above method, the fine-grained embedding characterization in S2 is described in detail below:

[0148] S21: In cross-domain scheduling scenarios, state features exhibit significant dimensional differences and non-stationary distribution characteristics. Normalization methods using fixed statistics are prone to data distribution shifts when dealing with electricity price spikes or sudden load surges. Therefore, for new observations of input features at the current time step, incremental recursive calculations are performed using the feature mean and second-order central moments from the previous time step. This achieves numerically stable real-time statistical updates without requiring the storage of all historical data.

[0149]

[0150]

[0151] In the formula, and These represent the system at time steps. With time step The characteristic dynamic mean; Indicates time step New observations of the arriving input features; Indicates the current time step index; and These represent the system at time steps. With time step The characteristic second-order central moment.

[0152] S22: Abnormal observations in a dynamic scheduling environment can easily interfere with the network gradient, leading to extreme instability in the training convergence process of the agent. Therefore, based on the incrementally updated second-order central moments of the features in the aforementioned steps, the dynamic variance estimate is calculated. The new observations are then standardized using Z-score, and a threshold constraint is applied to the standardized features using a truncation function. This eliminates the interference of abnormal extreme values ​​on the gradient, transforming the initial resource state vector obtained in S1 into a stable standard system state feature. :

[0153]

[0154] In the formula, This represents a cutoff function used to limit the range of features; This represents the dynamic variance estimate calculated based on the characteristic second-order central moments, and ; This represents a very small positive number preset by the system to prevent the denominator from being zero; and These represent the upper and lower threshold values ​​set by the truncation function, respectively.

[0155] S23: The computing load of the data center physical cluster and the carbon emission intensity of the power grid in the region exhibit a complex periodic pattern of intraday fluctuations combined with weekly trends. A single linearly increasing time step cannot directly reflect this time series evolution cycle. Therefore, the global discrete time step of the system needs to be considered. Perform orthogonal decomposition to extract hourly components representing intraday cycles. With the weekday component representing the weekly cycle This decouples continuous linear timescales into multi-granularity periodic temporal index coordinates:

[0156]

[0157] S24: The one-hot encoded vectors corresponding to discrete time indices are mutually orthogonal and highly sparse in the feature space, making it difficult to express the temporal correlation similarity between different time nodes. Therefore, independent hour embedding matrices and day-of-week embedding matrices are constructed. The one-hot encoded vectors corresponding to the extracted hour and day-of-week components are used as input, and they are projected onto a high-dimensional continuous latent space through matrix multiplication to extract independent periodic features with dense expressive power. Subsequently, periodic features at different scales are concatenated and fused. This parameterized mapping process introduces a strong inductive bias in the feature space, forcing the scheduling model to share the same hourly feature subspace at the same time on different dates, thus extracting a unified temporal feature vector.

[0158]

[0159] In the formula, This represents the unified temporal feature vector generated after mapping and fusion; and These represent hourly components respectively. With weekday portion Perform an index query in the corresponding feature dictionary to extract the embedded feature representation; This represents a vector concatenation operation; and Representing time indexes respectively and The corresponding one-hot encoded vector after conversion; and These represent the continuously optimizable hour embedding matrix and weekday embedding matrix, respectively.

[0160] S25: Assign a unique learnable topology embedding vector as a conditional context to each data center node, and combine it with the aforementioned standardized real-time state vector. By jointly representing the state projection matrix and the nonlinear activation function GeLU, the inherent topological physical attributes of nodes are injected while encoding time-varying load information. This explicitly maps the hardware specifications and geographical differences of heterogeneous resources, thereby generating a data center converged state matrix with data physical origin awareness capabilities. This addresses the common phenomenon of homomorphic but heterogeneous cross-domain heterogeneous clusters having the same resource occupancy status but different hardware specifications or geographical locations.

[0161]

[0162] In the formula, Represents data center node High-dimensional feature representation; GeLU represents the nonlinear activation function. Represents the state projection matrix; Represents data center node At any moment The standardized real-time state vector; Represented as data center node The unique learnable topological embedding vector assigned.

[0163] The multi-head cross-attention matching mechanism in S3 of the above method is described in detail below:

[0164] S31: Single feature concatenation cannot capture the nonlinear coupling relationship between multi-dimensional trade-offs such as computing power adaptability and carbon emission economy. To accurately extract the scheduling matching degree between specific task requirements and heterogeneous data center supply capacity, the extracted features of the task to be scheduled are constructed into a query vector, and the data center fusion state representation matrix is ​​constructed into key and value vectors. An asymmetric cross-domain projection operation is performed using independent learnable projection parameter matrices to map them to multiple independent semantic subspaces, and the query vector, key vector, and value vector corresponding to each attention head are calculated respectively.

[0165]

[0166] In the formula, , , They represent the first The query vector, key vector, and value vector of each attention head; This represents the query vector obtained after linear projection of the task features; The matrix represents the converged state of the data center. , , These represent the learnable projection parameter matrices corresponding to the query, key, and value, respectively.

[0167] S32: Given that direct dot product of high-dimensional vectors in the semantic subspace easily leads to numerical extrema and gradient vanishing, and that hard matching mechanisms are difficult to flexibly adjust decisions according to dynamic environments, this invention introduces a scaling factor and a temperature adjustment coefficient. This coefficient is used to perform a scaling dot product operation on the query vector and key vector corresponding to each attention head, generating an attention score vector that measures the alignment between task requirements and the capabilities of each node. After normalization, the value vectors are weighted and aggregated to generate the output feature vector of each attention head.

[0168]

[0169] In the formula, Indicates the first The output feature vector of each attention head; This represents the normalized exponential function; The transpose matrix of the key vector; Represents the feature dimensions of each attention head; Indicates the scaling factor; This represents the temperature regulation coefficient, used to smooth the distribution of attention scores to adjust the exploratory nature of the scheduling strategy.

[0170] S33: Next, the feature vectors output by all attention heads are concatenated, and linear transformation and layer normalization are performed through the multi-head output projection matrix. This aggregates the task and node matching information scattered in different semantic subspaces, generating a spatial attention aggregation vector that provides a global perspective on scheduling supply and demand matching representation.

[0171]

[0172] In the formula, Represents the multi-head space cross-attention aggregation vector; Indicates the layer normalization processing function; This represents a vector concatenation operation; express Output feature vectors of different attention heads; This represents the multi-head output projection matrix.

[0173] S34: The spatial cross-attention aggregation vector generated above is fused with the multi-scale temporal embedding vector described in S2, and input into the feedforward projection network for nonlinear extraction. The residual of the spatial attention aggregation vector is connected to the network output. While retaining the original spatial matching expressive ability without loss, spatiotemporal global semantics are injected to generate an embedding representation containing a complete global spatiotemporal context.

[0174]

[0175] In the formula, Represents a multi-scale temporal embedding vector; This represents a vector concatenation operation; and This represents the weight parameter matrix of the feedforward projection network; and This represents the bias term of the feedforward projection network; Represents the nonlinear activation function ReLU.

[0176] The detailed description of the cross-domain resource allocation decision action generation in S4 of the above method is as follows:

[0177] S41: The previously generated feature vector containing the global spatiotemporal context Inputting a feature extraction layer of a policy network based on a multilayer perceptron architecture, and performing forward propagation processing using a nonlinear activation function, yields a deep policy decision representation related to cross-domain resource allocation. :

[0178]

[0179] In the formula, Represents the ReLU activation function; The policy network contains parameters. Multilayer perceptron feature extraction layer;

[0180] S42: A fully connected layer is used to map the extracted deep policy decision representation into action log probabilities. A normalized exponential function is introduced to transform these action log probabilities into a standardized stochastic policy distribution, generating a state distribution for the system across heterogeneous data centers. Take action below The probability distribution of task scheduling actions provides a decision space for the random sampling of subsequent scheduling actions:

[0181]

[0182] In the formula, This represents the normalized exponential function; This represents the weight parameter matrix corresponding to the output layer of the policy network; This represents the deep decision-making process of the strategy. This represents the bias term parameters corresponding to the output layer of the policy network.

[0183] S43: The gradient updates of the policy network alone have extremely high variance. To ensure stable convergence of model training, a value network is constructed to assist in evaluating the decision benchmark of the policy, and feature vectors containing the global spatiotemporal context are used. The feature extraction layer of a multi-layer perceptron within the synchronous input value network architecture is combined with a non-linear activation function for feature dimensionality reduction, generating a deep value representation for state value assessment.

[0184]

[0185] In the formula, Represents the ReLU activation function; This indicates that the value network contains parameters. Multilayer perceptron feature extraction layer;

[0186] S44: Deep Value Representation Based on the Foregoing Generation By performing an affine transformation on the linear output layer of the value network, the current system state for estimating multi-objective scheduling is obtained. Scalar value benchmark for expected cumulative return This provides a quantitative evaluation basis for calculating the advantage function and reducing the policy gradient variance:

[0187]

[0188] In the formula, This represents the weight parameter matrix corresponding to the output layer of the value network; This represents the deeper meaning of value; This represents the bias term parameters corresponding to the output layer of the value network.

[0189] In the above method, the detailed description of the composite reward construction and network parameter iterative optimization in S5 is as follows:

[0190] S51: The carbon emission intensity of the power grid and electricity market prices exhibit non-stationary distributions and high-frequency spikes. Directly calculating rewards using absolute values ​​can easily lead to reward distribution shifts and strategy update oscillations. Based on the projected energy consumption of the task at the current moment and the real-time carbon emission intensity and electricity price of the power grid, initial carbon emission penalties and initial energy cost penalties are calculated respectively. Subsequently, the pre-maintained rolling mean and standard deviation are extracted, and dynamic standardization is performed on the initial penalty terms to generate standardized carbon emission rewards and standardized energy cost rewards that eliminate environmental dynamic interference.

[0191]

[0192]

[0193] In the formula, Indicates time step The calculated initial penalty for carbon emissions; This represents the set of tasks currently being executed in the corresponding data center; Indicates the task index in the set; Indicates task The projected energy consumption; Indicates the corresponding data center at time step Real-time grid carbon emission intensity; Indicates standardized carbon emission rewards; and These represent the rolling mean and standard deviation of the initial carbon emission penalty, respectively. This represents a very small positive number preset by the system to prevent the denominator from being zero; Indicates time step The calculated initial penalty term for energy consumption cost; Indicates the corresponding data center at time step Real-time electricity prices; Indicates a standardized energy consumption cost incentive; and These represent the rolling mean and standard deviation of the initial penalty term for energy consumption costs, respectively.

[0194] S52: Extract the actual completion time of the task and the system-required deadline, calculate the hard constraint penalty term with respect to the time threshold using an indicator function, trigger a fixed physical penalty when the actual completion time of the task exceeds the deadline limit, and generate a service quality penalty reward for the guided strategy network to avoid breach of contract. :

[0195]

[0196] In the formula, This indicates the preset physical penalty coefficient for breach of contract. This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Indicates the actual completion time of the task; This indicates the deadline specified for the task.

[0197] S53: Combining the scheduling strategy with the differences in the degree and scale of preference for different objectives, a weighted summation operation is performed on the standardized carbon emission reward, standardized energy consumption cost reward, and service quality penalty reward obtained above, to construct a scalarized multi-objective composite reward, guiding the agent to approach the Pareto optimal frontier in a complex policy space with multiple conflict constraints:

[0198]

[0199] In the formula, This represents the multi-objective composite reward calculated at the current time step; , and These represent the carbon emission target weight coefficient, energy consumption cost target weight coefficient, and service quality target weight coefficient configured in the weight vector, respectively.

[0200] S54: Cross-domain resource scheduling tasks suffer from significant long-term temporal dependencies and reward sparsity; relying solely on single-step rewards cannot accurately assess the long-term physical impact of actions. Based on the aforementioned constructed multi-objective composite reward, the system's preset discount factor, and the current and next-time scalar value benchmarks output by the value network, single-step temporal difference errors are calculated. Using the exponentially weighted moving average of multi-step temporal difference errors, a generalized advantage estimate is calculated, thereby quantifying the long-term physical benefit of the current scheduling action relative to the average benchmark.

[0201]

[0202]

[0203] In the formula, Indicates time step Single-step timing difference error; Indicates multi-objective composite reward At time step Instant rewards; Indicates the discount factor; and These represent the value network in relation to the current state. State at the next moment Output scalar value benchmark prediction; This represents the estimate of generalized advantage; Indicates the error attenuation coefficient; This indicates the maximum time step of the round.

[0204] S55: Static policy update pruning thresholds are highly susceptible to policy update oscillations and system scheduling downtime risks in complex industrial environments facing sudden changes in grid carbon emission intensity or drastic fluctuations in node load. To address this, the feature dynamic variance estimate, updated in real-time by an online algorithm during the preprocessing stage described in S1, is extracted. A norm is then taken to characterize the total volatility of the global physical environment. Combined with a preset baseline pruning threshold and a sensitivity adjustment coefficient, an adaptively adjusting dynamic pruning threshold is generated through exponential decay mapping.

[0205]

[0206] In the formula, This represents the dynamic clipping threshold calculated at the current time step; This represents the system's preset baseline clipping threshold; The sensitivity adjustment coefficient represents the degree of aversion of the control strategy to the risk of fluctuations in the physical environment; This represents the dynamic variance estimate of the system environment features extracted in real time; It represents the norm of the dynamic variance estimate of the feature.

[0207] S56: The parameter updates of the value network require an accurate fitting target to guide the convergence of state values. The generalized advantage estimate obtained above is linearly superimposed with the predicted output of the old value network for the current state from the previous iteration to construct the fitting target value of the value network. This target value is used to characterize the true expected cumulative return benchmark of the value network under the current system state. :

[0208]

[0209] In the formula, This represents the estimate of generalized advantage; This indicates that the old value network is related to the current state. Output scalar value benchmark prediction.

[0210] S57: The gradient update process of multi-agent policies requires a multi-dimensional trade-off between maximizing cumulative reward, constraining update step size, and preventing getting trapped in local optima. Based on the ratio of the probability distributions of actions in the old and new policies, combined with the previously calculated dynamic pruning threshold, generalized advantage estimate, fitted target value, and policy distribution entropy, a joint loss function is constructed, including the pruning agent objective, value network loss, and entropy regularization term. An adaptive moment estimation optimizer is used to perform backpropagation on this joint loss function, iteratively updating the policy network parameters and value network parameters until the overall performance of the multi-objective optimization of the agents converges.

[0211]

[0212]

[0213] In the formula, Represents the joint loss function; Represents the empirical averaging operator; This represents the ratio of the probability distributions of actions generated by the old and new strategies. This represents the truncation function; Indicates the dynamic clipping threshold; Indicates the loss of value network, and ; The entropy regularization term represents the policy distribution; and These represent the weight coefficients of the value network loss and the entropy regularization term, respectively. This represents the adaptive moment estimation optimizer function; and These represent the parameters of the policy network and value network in the current round, respectively. Represents the gradient of the joint loss function; Indicates the learning rate; and This indicates the network parameters after the next round of updates.

[0214] Example

[0215] This invention introduces the DCcluster-Opt, an open benchmark platform for high-fidelity distributed clusters in the industry, to construct a test scenario. This platform was primarily developed by HP Labs. It is deeply integrated into the ExaDigiT consortium's open framework for supercomputer digital twins. The platform's underlying physical engine directly targets the complex operational characteristics of national-level large-scale computing infrastructures such as the U.S. Department of Energy's U.S. Scientific Cloud. Based on this benchmark, this embodiment can rigorously verify the engineering application value of the proposed multi-objective sustainable optimization and control method within the next-generation federated scientific computing ecosystem. The specific experimental configuration and evaluation system are as follows:

[0216] 1. Experimental Dataset Processing Instructions

[0217] The underlying environment of this embodiment is jointly driven by multiple real-world industrial-grade datasets to reproduce the non-stationary fluctuation characteristics of complex heterogeneous environments.

[0218] 1) Workload Data: Import the Alibaba Cluster Trace 2020 production cluster trajectory as the raw task flow. This trajectory realistically records the concurrent call status of over 6500 graphics processing units (GPUs). Short tasks with a runtime of less than 15 minutes were filtered out, retaining the core business load. Subsequently, daily and weekly fluctuation patterns were extracted from the raw trajectory, and random noise following a normal distribution was superimposed. The two-month periodic data was extended to the whole year to simulate the tidal fluctuations and uncertainties of tasks in a real industrial environment.

[0219] 2) Physical Environment Data: The simulation engine is driven by access to real physical environment data. For carbon emission intensity, the 2023 global power grid carbon intensity sequence provided by Electricity Maps is used, with a time resolution set to 15 minutes, to accurately map the spatiotemporal differences in the energy structure of different regional power grids. For electricity prices, wholesale electricity market prices from multiple regional power grids are integrated to reproduce the price spikes in the real market. For meteorological data, historical temperature data from Open-Meteo is introduced to drive the built-in cooling system model of the benchmark environment, dynamically calculate power usage efficiency (PUE), and construct physical energy consumption constraints based on thermodynamics.

[0220] 2. Test Environment Configuration Instructions

[0221] This embodiment constructs a global heterogeneous cluster topology containing five geographically distributed data centers, as shown in Table 1. This topology explicitly creates an asymmetric mismatch in resources, environment, and cost, forcing the scheduling system to face an explicit physical conflict between computing architecture affinity, green and low-carbon practices, and operating costs.

[0222] Table 1 Detailed Configuration of Distributed Data Center Topology (2023)

[0223]

[0224] Among them, Data Center DC1 (Canada-Ontario) is a green, low-cost node with the lowest carbon emission intensity and electricity price in the entire network. It has ample graphics card resources but extremely limited processor resources. Data Center DC2 (USA-Texas) is a general-purpose computing node with ample processor resources, but its carbon emission costs and electricity prices are at a moderate level. Data Center DC3 (India-West) is a high-capacity disaster recovery node equipped with massive processors and memory, but it has the highest carbon emission intensity in the entire network and is dedicated to providing physical backup for handling extreme concurrent requests. Data Center DC4 (Australia-Victoria) is a balanced node with moderate electricity prices, but its carbon emissions remain high due to the dominance of coal-fired power. Data Center DC5 (Singapore) is a high-performance, high-cost node with ample computing resources but extremely high average electricity prices.

[0225] 3. Evaluation Indicators Explanation

[0226] To comprehensively evaluate the multi-dimensional trade-off capability of scheduling strategies under complex constraints, this embodiment divides the evaluation indicators into the following two categories:

[0227] In terms of green, low-carbon, and economic costs, this includes total electricity cost (USD), total energy consumption (kWh), and total carbon emissions (kg). These indicators quantify the actual impact of cross-domain scheduling strategies on the external physical environment during execution, directly reflecting the system's sustainable optimization efficiency in energy conservation, emission reduction, and reduced operation and maintenance costs.

[0228] In terms of service quality assurance, this includes task completion rate (%), task compliance rate (i.e., the proportion of tasks that meet service quality constraints on time, %), average processor (CPU) utilization rate (%), and average graphics processing unit (GPU) utilization rate (%). These metrics measure the business delivery capability of the scheduling strategy in the face of sudden surges in heterogeneous tasks, as well as the global balancing and dynamic reuse level of the underlying hardware physical computing resources.

[0229] 4. Test Instructions

[0230] This invention discloses a multi-objective sustainable optimization and control method for wide-area heterogeneous data centers. The core of this invention lies in constructing a fine-grained spatiotemporal state representation and a multi-dimensional cross-domain collaborative supply-demand matching architecture. By introducing an adaptive reinforcement learning decision-making mechanism that is aware of environmental fluctuations, this invention achieves efficient collaborative scheduling of cross-domain computing resources and wide-area green energy. Compared with traditional single-cluster scheduling methods, this invention effectively solves the feature distribution offset problem under complex spatiotemporal coupling environments and overcomes technical bottlenecks such as policy gradient dominance and update oscillations caused by non-stationary energy environments. This verification test is mainly used in practical industrial application scenarios of collaborative scheduling of globally distributed data centers. The experiment aims to verify whether this invention can still achieve robust coordination between business service quality and energy conservation and carbon reduction goals under stringent constraints such as drastic changes in grid carbon emission intensity, electricity price peaks, and sudden tidal surges in computing load, thereby achieving better overall performance than existing methods, that is, substantially reducing the overall carbon footprint and operating costs of the system while ensuring a low service default rate.

[0231] I. Performance Testing and Result Analysis

[0232] This experiment utilizes an industrial-grade distributed cluster simulation benchmark platform. The simulation environment integrates the workload trajectory of a real production cluster and historical sequences of carbon intensity and dynamic electricity prices from multiple global regional power grids. It also constructs a multi-datacenter network topology with heterogeneous characteristics such as green and low-cost operation, ultra-large capacity disaster recovery, and high performance at high cost. Model training employs a centralized training and distributed execution paradigm, advancing the simulation with a fixed time step and collecting interactive samples to iterate network parameters. To comprehensively evaluate the scheduling performance of AST-Embed (a multi-objective reinforcement learning algorithm based on adaptive spatiotemporal embedding), this experiment selects the following baseline methods for comparative testing:

[0233] 1) Traditional heuristic scheduling rules, including local priority and round-robin scheduling algorithms. These baselines do not incorporate complex intelligent learning mechanisms, aiming to quantify the contribution of the cross-domain collaborative scheduling mechanism in this invention to basic quality service assurance by reducing energy costs and improving resource utilization.

[0234] 2) Model-free reinforcement learning algorithms, covering the value function-based RAINBOW algorithm, the maximum entropy-based flexible actor-critic algorithm SAC, the dominant actor-critic architecture-based A3C algorithm, and the APPO and IMPALA algorithms for distributed high-concurrency environments. The aim is to compare the exploration stability and sample utilization efficiency of different reinforcement learning networks when handling multi-objective conflict constraints.

[0235] 3) Model-based reinforcement learning algorithm. This test selected the DreamerV3 algorithm, based on a world model architecture. This algorithm can predict potential environmental dynamics in high-dimensional state spaces. Using it as a baseline, the aim is to verify the advantages of the AST-Embed algorithm in the present invention in terms of feature representation accuracy and multi-objective optimization capability when facing non-stationary spatiotemporal environments, relying on prior physical-spatiotemporal decoupling and residual cross-attention mechanisms.

[0236] Based on the statistical data from multiple rounds of simulation tests, the comprehensive performance comparison and analysis of each algorithm is shown in Tables 2 and 3:

[0237] Table 2. Scheduling optimization results for each baseline in terms of green, low-carbon, and economic costs (full year 2023)

[0238]

[0239] Table 2 presents the annual simulation scheduling results of each algorithm in terms of green and low-carbon aspects and economic costs. As shown in Table 2, the AST-Embedde algorithm achieves the best performance in both the electricity cost and carbon emission, two core sustainability indicators. Specifically, when dealing with multi-dimensional conflict constraints, traditional heuristic rules and conventional model-free reinforcement learning algorithms (such as APPO, A3C, and SAC) have significantly higher long-term carbon emissions and electricity costs. Table 2 shows that, taking the weaker APPO algorithm as an example, its total electricity cost and total carbon emissions are 7.95% and 9.98% higher than the AST-Embedde algorithm, respectively. This indicates that conventional algorithms cannot effectively perceive the high-frequency fluctuations in global grid carbon emission intensity and electricity prices. When the network's computing power is limited or there is a sudden load surge, these algorithms tend to offload computational tasks directly to nodes with ample computing power but high costs and high carbon emissions. While this avoids default risks, it significantly increases the overall carbon footprint and operating costs. Furthermore, Table 2 shows that the DreamerV3 algorithm based on the world model achieves a slight reduction in total power consumption of 0.20% compared to the AST-Embed algorithm. However, despite lower power consumption, DreamerV3's total electricity cost and total carbon emissions are 1.10% and 1.04% higher than the AST-Embed algorithm, respectively. This objectively demonstrates that in heterogeneous power grid environments, simply minimizing power consumption is not equivalent to achieving green and low-carbon development. The AST-Embed algorithm, relying on multi-scale time decomposition embedding and dynamic normalization reward mechanisms, extracts the spatiotemporal coupling characteristics of time-varying carbon emission intensity and fluctuating electricity prices. This mechanism guides the scheduling agent to proactively transfer load in the time and space dimensions, allocating computing tasks to environmental windows with low carbon emissions and low electricity prices. Ultimately, this method substantially reduces the data center's grid power procurement costs and total carbon emissions while maintaining similar power consumption.

[0240] Table 3. Scheduling optimization results for each baseline in terms of service quality assurance (full year 2023)

[0241]

[0242] Table 3 presents the scheduling results of each algorithm in terms of business efficiency and resource guarantee throughout the simulation period. The data shows that the AST-Embed algorithm achieves an effective balance between high task throughput and global computing power reuse in multi-objective conflict scenarios. Specifically, in terms of task completion rate, the traditional heuristic rules Local Only and Round Robin perform significantly lower. In wide-area clusters with massive task concurrency, the lack of a flexible cross-domain offloading mechanism easily exhausts local node resources, leading to large-scale task dropout. Conversely, AST-Embed and other reinforcement learning algorithms, relying on cross-domain collaborative scheduling, maintain a stable task completion rate above 99.97%, effectively ensuring high-throughput operation of the system. Combined with the average CPU utilization index, it can be seen that other reinforcement learning baselines exhibit severe resource reuse imbalances under multi-dimensional constraints. Among them, APPO and DreamerV3 have average CPU utilization rates of only 4.28% and 18.50%, respectively. This reflects that existing intelligent agents adopt an extremely conservative concession strategy to avoid default penalties, excessively concentrating tasks on local redundant nodes with no latency risk, resulting in a large area of ​​idle remote nodes. AST-Embed, on the other hand, steadily increases the average CPU usage to 28.75% while maintaining a 99.98% task completion rate. This effectively confirms that our method, relying on spatiotemporal feature decoupling and multi-head cross-attention mechanisms, accurately perceives the remaining computing power of global nodes, prompting a reasonable distribution of computational load in the spatial dimension, thus effectively overcoming the local optimum defect of idle computing power. Further cross-comparison of Tables 2 and 3 reveals the physical trade-offs of our method in multi-objective combinatorial optimization. AST-Embed's task achievement rate is 27.96%, lower than Round Robin. However, as shown in Table 2, Round Robin's higher achievement rate is the result of blindly distributing tasks at the cost of a high task failure rate, high carbon footprint, and high electricity costs. In contrast, AST-Embed actively transfers a large amount of computational load to remote, low-carbon, and low-cost nodes. This cross-domain wide area network transmission inevitably introduces additional communication delays, thereby lowering the task completion rate under objective physical constraints.

[0243] In summary, the AST-Embed algorithm does not achieve single-dimensional performance targets by excessive computing power or by compromising with a high-carbon power grid. This invention, by tolerating some cross-domain transmission overhead, achieves substantial reductions in grid power procurement costs and overall carbon footprint while ensuring high throughput for global tasks and healthy reuse of computing power. Ultimately, this control method approaches the Pareto optimal frontier under complex conflict constraints, effectively meeting the long-term operational requirements of large-scale heterogeneous data centers for green, low-carbon, safe, and stable operation.

[0244] III. Test Summary

[0245] This section presents a comprehensive scheduling performance test of the invention based on an industrial-grade distributed cluster simulation benchmark platform. Through comparative analysis with traditional heuristic rules and various mainstream reinforcement learning baseline algorithms, the engineering rationality and advancement of the proposed multi-objective reinforcement learning (AST-Embed) method based on adaptive spatiotemporal embedding are verified. Experimental results show that the invention achieves Pareto optimal balanced scheduling in both business efficiency assurance and green low-carbon cost control, bringing practical environmental and economic benefits such as efficient reuse of global computing resources and green migration of cross-domain computing power. While strictly ensuring a high task completion rate of 99.98%, it achieves substantial reductions in the overall carbon footprint of the data center and grid power procurement costs. This invention achieves cross-domain spatiotemporal feature adaptive learning through a multi-scale fine-grained spatiotemporal feature embedding mechanism, and solves the problems of state perception bias and feature distribution failure that are easily caused by traditional static feature extraction when facing massive heterogeneous nodes and complex environmental fluctuations. It also solves the carbon emission overflow problem caused by traditional single cluster scheduling being severely decoupled from external energy physical constraints, and effectively overcomes the local computing power idle trap and policy update oscillation problem that conventional reinforcement learning algorithms are prone to fall into when dealing with non-stationary environments and multi-dimensional conflict constraints.

Claims

1. A multi-objective sustainable optimization regulation method for wide-area heterogeneous data centers, characterized in that, Includes the following steps: S1: Acquire heterogeneous state data and task feature data from a wide-area multi-data center to construct a scheduling state space with electricity cost, carbon emission intensity, and service quality assurance as optimization objectives; S2: Perform dynamic statistical normalization on the collected data in the heterogeneous scheduling state space to eliminate environmental fluctuation interference, extract the real-time state variance, fuse multi-scale time-series components and node topology identity, and generate a fused embedded state matrix. S3: Based on the task feature data and data center fusion state matrix of the aforementioned standardized processing, the matching score of task requirements and node capabilities is calculated through multi-head cross attention mechanism, spatial attention aggregation features are extracted, and residual direct connection bypass and spatiotemporal global semantic information are introduced for fusion to generate feature vectors containing global spatiotemporal context. S4: Input the feature vector containing the global spatiotemporal context into the policy network of the reinforcement learning agent to extract deep decision representations related to cross-domain resource allocation, and generate action log probabilities based on the deep decision representations, outputting the task scheduling action probability distribution for heterogeneous data centers. S5: Iteratively interact the sampled scheduling actions with the multi-data center environment, calculate the multi-objective composite reward and the observation results of the next time step, dynamically adjust the adaptive pruning threshold based on the real-time state variance extracted in step S2, and perform value evaluation and policy network parameter update by combining the composite reward and the adaptive pruning threshold to achieve multi-objective collaborative optimization of carbon emissions, electricity costs and service quality.

2. The method of claim 1, wherein, The detailed description of the construction of the heterogeneous state space in S1 is as follows: S11: Based on historical regional data, linear interpolation is performed on the real-time monitored signals of power grid carbon emission intensity, electricity price, and ambient temperature in the region where the data center node is located. This is then superimposed with a normally distributed coherent noise simulation to construct physically realistic scheduling constraints and generate the current external environment data. wherein, , and represent data center nodes at time a real-time grid carbon intensity, electricity price and ambient temperature; is a regional historical dataset; represents a current time; represents a linear interpolation processing function; represents coherent noise subject to a normal distribution with a mean and a variance . S12: Extract the total amount and current load of various computing resources of data center nodes, and normalize the computing resource availability. wherein, denotes the time instant data center node a normalized resource availability vector of the data center node , , denote the total capacity of the processor cores, the total capacity of the graphic cards and the total capacity of the memory of the node , , , denote the load occupancy of the processor cores, the load occupancy of the graphic cards and the load occupancy of the memory of the node at the time instant . S13: In combination with the total core processing scale of the cluster, the remaining execution time of the tasks in the running queue, the estimated total duration of the tasks in the waiting queue, and the core demand of each task are normalized to calculate the load backlog time estimation feature In this way, the time cost of processing the historical backlog tasks of each node and the dynamic accumulation degree of the computing load are quantified: wherein, denotes the running queue; denotes the waiting queue; denotes the estimated end time of a task in the running queue; denotes the current time; denotes the core demand of a task in the running queue; denotes the estimated total duration of a task in the waiting queue; denotes the core demand of a task in the waiting queue; denotes the core demand of a task in the waiting queue; denotes the core demand of a task in the waiting queue; denotes the core demand of a task in the waiting queue; denotes the total core processing capacity of a cluster of nodes denotes the total core processing capacity of a cluster of nodes denotes the total core processing capacity of a cluster of nodes denotes the preset minimum normal number of the system to prevent the denominator from being zero; S14: calculating the network round-trip time between the source node and the target node and the system preset maximum delay constant , the time of the data center node normalized transmission delay feature to quantify the current network congestion and transmission overhead: In the formula, with respectively represent the source node and the target node of data transmission; S15: splice the extracted resource availability vector , load backlog time estimation feature and transmission delay feature with the real grid carbon intensity , electricity price and ambient temperature to construct the initial resource state vector of each data center node at time : S16: Decouple the server's total power consumption into a base static power consumption independent of the load and a dynamic power consumption that is linearly positively correlated with processor utilization, and combine this with the analyzed tasks to be scheduled. Estimated execution time Requested number of processor cores With the number of graphics cards And the unit dynamic power consumption coefficient of the corresponding hardware unit and the basic static power consumption of the server. The tasks to be scheduled are calculated comprehensively. Projected energy consumption : In the formula, and These represent the unit dynamic power consumption coefficients for the processor unit and the graphics card unit, respectively. S17: Collect the heterogeneous resource requirements, estimated execution time, and remaining deadline of the tasks to be scheduled, and concatenate them with the aforementioned estimated energy consumption to construct the time-sharing data. Tasks to be scheduled initial feature vector : In the formula, These represent the tasks to be scheduled. The requested memory size; Indicates the estimated execution time; Indicates tasks to be scheduled Service quality remaining deadline constraints; Indicates projected energy consumption; S18: The initial resource state vector set of each data center node at time t is jointly represented with the initial feature vector set of all tasks to be scheduled to construct the scheduling state space.

3. The method according to claim 1, characterized in that, In S2, the fine-grained embedding representation is described in detail below: S21: Based on the characteristic mean and second-order central moments of the previous time step, perform incremental recursive calculations on the new observations of the current time step to update the real-time statistics: In the formula, and These represent the system at time steps. With time step The characteristic dynamic mean; Indicates time step New observations of the arriving input features; Indicates the current time step index; and These represent the time steps. With time step The characteristic second-order central moment; S22: Calculate the dynamic variance estimate based on the updated characteristic second-order central moments. The new observations are Z-score standardized, and a threshold constraint is applied to the standardized features using a truncation function. This results in the initial resource state vector obtained in S1 being... Transform into the state characteristics of a standard system with a stationary distribution : In the formula, This represents a cutoff function used to limit the range of features; This represents a very small positive number preset by the system to prevent the denominator from being zero; and These represent the upper and lower threshold values ​​of the feature set by the truncation function; S23: For the global discrete time step Perform orthogonal decomposition to extract hourly components representing intraday cycles. With the weekday component representing the weekly cycle This decouples continuous linear timescales into multi-granularity periodic temporal index coordinates: S24: Construct mutually independent hour embedding matrices and day-of-week embedding matrices, and combine the previously extracted hour components. With weekday portion The corresponding one-hot encoding is used as input, and its projection is mapped to a high-dimensional continuous latent space through matrix multiplication to extract independent periodic features with dense expressive power. Then, periodic features at different scales are concatenated and fused to extract a unified temporal feature vector. In the formula, This represents the unified temporal feature vector generated after mapping and fusion; and These represent components based on hours. With weekday portion Perform an index query in the corresponding feature dictionary to extract the embedded feature representation; This represents a vector concatenation operation; and Representing time indexes respectively and The corresponding one-hot encoded vector after conversion; and These represent the continuously optimizable hour embedding matrix and weekday embedding matrix, respectively. S25: Assign a unique learnable topology embedding vector as a conditional context to each data center node, and combine it with the normalized real-time state vector. By jointly representing the state projection matrix and the nonlinear activation function GeLU, the inherent topological physical attributes of nodes are injected while encoding time-varying load information. This explicitly maps the hardware specifications and geographical differences of heterogeneous resources, generating a data center fusion embedded state matrix with data physical origin awareness capabilities. In the formula, Represents data center node High-dimensional feature representation; GeLU represents the nonlinear activation function; Represents the state projection matrix; Represents data center node At any moment The standardized real-time state vector; Represented as data center node The unique learnable topological embedding vector assigned.

4. The method according to claim 1, characterized in that, The multi-head cross-attention matching mechanism in S3 is described in detail below: S31: Construct the features of the task to be scheduled into a query vector, construct the data center fusion state representation matrix into key and value vectors, and perform an asymmetric cross-domain projection operation using independent learnable projection parameter matrices. Map the projection parameter matrix to multiple independent semantic subspaces, and calculate the query vector, key vector, and value vector corresponding to each attention head respectively. In the formula, , , They represent the first The query vector, key vector, and value vector of each attention head; This represents the query vector obtained after linear projection of the task features; The matrix represents the converged state of the data center. , , These represent the learnable projection parameter matrices corresponding to the query, key, and value, respectively. S32: Introducing scaling factors and temperature adjustment coefficients, scaling dot product operations are performed on the query vector and key vector corresponding to each attention head to generate an attention score vector that measures the alignment between task requirements and the capabilities of each node. After normalization, the value vectors are weighted and aggregated to generate the output feature vector of each attention head. In the formula, Indicates the first The output feature vector of each attention head; This represents the normalized exponential function; The transpose matrix of the key vector; Represents the feature dimensions of each attention head; Indicates the scaling factor; This represents the temperature regulation coefficient, used to smooth the distribution of attention scores to adjust the exploratory nature of the scheduling strategy; S33: Concatenate the feature vectors output by all attention heads, and perform linear transformation and layer normalization through the multi-head output projection matrix to aggregate task and node matching information scattered in different semantic subspaces, generating a spatial attention aggregation vector to provide a global perspective on scheduling supply and demand matching representation: In the formula, Represents the multi-head space cross-attention aggregation vector; Indicates the layer normalization processing function; This represents a vector concatenation operation; express Output feature vectors of different attention heads; This represents the multi-head output projection matrix; S34: The spatial cross-attention aggregation vector is fused with the unified temporal feature vector obtained in S2, and input into the feedforward projection network for nonlinear extraction. The residual of the spatial attention aggregation vector is connected to the network output. While retaining the original spatial matching expressive ability without loss, spatiotemporal global semantics are injected to generate a feature vector containing a complete global spatiotemporal context. In the formula, A unified temporal feature vector is represented; This represents a vector concatenation operation; and This represents the weight parameter matrix of the feedforward projection network; and This represents the bias term of the feedforward projection network; Represents the nonlinear activation function ReLU.

5. The method according to claim 1, characterized in that, The detailed description of cross-domain resource allocation decision action generation in S4 is as follows: S41: The previously generated feature vector containing the global spatiotemporal context Inputting a feature extraction layer of a policy network based on a multilayer perceptron architecture, and performing forward propagation processing using a nonlinear activation function, yields a deep policy decision representation related to cross-domain resource allocation. : In the formula, Represents the ReLU activation function; The policy network contains parameters. Multilayer perceptron feature extraction layer; S42: Representing deep policy decisions using fully connected layers The probability of an action is mapped to its logarithmic probability, and a normalized exponential function is introduced to transform this logarithmic probability into a standardized stochastic policy distribution, generating a state distribution for systems in heterogeneous data centers. Take action below The probability distribution of task scheduling actions provides a decision space for the random sampling of subsequent scheduling actions: In the formula, This represents the normalized exponential function; This represents the weight parameter matrix corresponding to the output layer of the policy network; This represents the deep decision-making process of the strategy. This represents the bias term parameters corresponding to the output layer of the policy network; S43: Incorporate feature vectors containing the global spatiotemporal context The feature extraction layer of a multi-layer perceptron within the synchronous input value network architecture is combined with a non-linear activation function for feature dimensionality reduction, generating a deep value representation for state value assessment. In the formula, Represents the ReLU activation function; This indicates that the value network contains parameters. Multilayer perceptron feature extraction layer; S44: Deep Value Representation Based on the Foregoing Generation By performing an affine transformation on the linear output layer of the value network, the current system state for estimating multi-objective scheduling is obtained. Scalar value benchmark for expected cumulative return This provides a quantitative evaluation basis for calculating the advantage function and reducing the policy gradient variance: In the formula, This represents the weight parameter matrix corresponding to the output layer of the value network; This represents the deeper meaning of value; This represents the bias term parameters corresponding to the output layer of the value network.

6. The method according to claim 1, characterized in that, In S5, the construction of composite rewards and iterative optimization of network parameters are described in detail below: S51: Based on the estimated energy consumption of the task at the current moment, the real-time grid carbon emission intensity, and the electricity price, calculate the initial carbon emission penalty and the initial energy cost penalty respectively; then extract the pre-maintained rolling mean and standard deviation, and perform dynamic standardization processing on the initial penalty to generate standardized carbon emission rewards and standardized energy cost rewards that eliminate environmental dynamic interference. In the formula, Indicates time step The calculated initial penalty for carbon emissions; This represents the set of tasks currently being executed in the corresponding data center; Indicates the task index in the set; Indicates task The projected energy consumption; Indicates the corresponding data center at time step Real-time grid carbon emission intensity; Indicates standardized carbon emission rewards; and These represent the rolling mean and standard deviation of the initial carbon emission penalty, respectively. This represents a very small positive number preset by the system to prevent the denominator from being zero; Indicates time step The calculated initial penalty term for energy consumption cost; Indicates the corresponding data center at time step Real-time electricity prices; Indicates a standardized energy consumption cost incentive; and These represent the rolling mean and standard deviation of the initial penalty term for energy consumption costs, respectively. S52: Extract the actual completion time of the task and the system-required deadline, calculate the hard constraint penalty term with respect to the time threshold using an indicator function, trigger a fixed physical penalty when the actual completion time of the task exceeds the deadline limit, and generate a service quality penalty reward for the guided strategy network to avoid breach of contract. : In the formula, This indicates the preset physical penalty coefficient for breach of contract. This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Indicates the actual completion time of the task; Indicates the deadline specified for the task; S53: Combining the scheduling strategy with the differences in the degree and scale of preference for different objectives, a weighted summation operation is performed on the standardized carbon emission reward, standardized energy consumption cost reward, and service quality penalty reward obtained above, to construct a scalarized multi-objective composite reward, guiding the agent to approach the Pareto optimal frontier in a complex policy space with multiple conflict constraints: In the formula, This represents the multi-objective composite reward calculated at the current time step; , and These represent the carbon emission target weight coefficient, energy cost target weight coefficient, and service quality target weight coefficient configured in the weight vector, respectively. S54: Based on multi-objective composite rewards, system-preset discount factors, and the current and next time-series scalar value benchmarks output by the value network, calculate the single-step temporal difference error. Using the exponentially weighted moving average of the multi-step temporal difference error, calculate the generalized advantage estimate, thereby quantitatively evaluating the long-term physical benefit of the current scheduling action relative to the average benchmark. In the formula, Indicates time step Single-step timing difference error; Indicates multi-objective composite reward At time step Instant rewards; Indicates the discount factor; and These represent the value network in relation to the current state. State at the next moment Output scalar value benchmark prediction; This represents the estimate of generalized advantage; Indicates the error attenuation coefficient; Indicates the maximum time step of a round; S55: Feature Extraction and Dynamic Variance Estimation A norm is taken to characterize the total volatility of the global physical environment, and a dynamic clipping threshold that adapts to environmental fluctuations is generated by combining a preset baseline clipping threshold and a sensitivity adjustment coefficient through exponential decay mapping. In the formula, This represents the dynamic clipping threshold calculated at the current time step; This represents the system's preset baseline clipping threshold; The sensitivity adjustment coefficient represents the degree of aversion of the control strategy to the risk of fluctuations in the physical environment; This represents the dynamic variance estimate of the system environment features extracted in real time; The norm representing the estimate of the dynamic variance of the feature; S56: Generalized dominance estimate By linearly superimposing the predicted value of the old value network from the previous iteration for the current state, a fitted target value for the value network is constructed, which is used to characterize the true expected cumulative return benchmark of the value network under the current system state. : In the formula, This represents the estimate of generalized advantage; This indicates that the old value network is related to the current state. Output scalar value benchmark prediction; S57: Based on the ratio of the probability distributions of actions under the new and old policies, combined with the previously calculated dynamic pruning threshold, generalized advantage estimate, fitted target value, and entropy of the policy distribution, a joint loss function is constructed, which includes the pruning agent target, value network loss, and entropy regularization term. The adaptive moment estimation optimizer is then used to perform backpropagation on this joint loss function, iteratively updating the policy network parameters and value network parameters until the overall performance of the agent's multi-objective optimization converges. In the formula, Represents the joint loss function; Represents the empirical averaging operator; This represents the ratio of the probability distributions of actions generated by the old and new strategies. This represents the truncation function; Indicates the dynamic clipping threshold; Indicates the loss of value network, and ; The entropy regularization term represents the policy distribution; and These represent the weight coefficients of the value network loss and the entropy regularization term, respectively. This represents the adaptive moment estimation optimizer function; and These represent the parameters of the policy network and value network in the current round, respectively. Represents the gradient of the joint loss function; Indicates the learning rate; and This indicates the network parameters after the next round of updates.

Images