An application method of computer integration data based on whole life cycle quality guarantee

Abstract

The application relates to a computer fusion data application method based on a full-life-cycle quality guarantee. The method comprises the following steps: normalizing and integrating computer original data of a computing power network to obtain a computer fusion initial data set; performing standard system processing on a space-time three-dimensional field in the computer fusion initial data set to obtain a standardized data set; performing coupling representation learning on computer-related data in the standardized data set to obtain a computer coupling portrait set; performing collaborative optimization arrangement on the computing power network according to the standardized data set and the computer coupling portrait set to obtain integrated regulation and transaction plan information; and performing instruction conversion on strategy elements of the integrated regulation and transaction plan information to obtain a multi-scale execution instruction set. The method can guarantee that the utilization degree of computer fusion data is difficult to meet the current power grid requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to an electric computer fusion data application method based on full life cycle quality guarantee. BACKGROUND

[0002] In the traditional technology, the electric computer fusion data application follows the basic process of "device collection-centralized storage-offline / quasi real-time analysis-result delivery". That is, the first-line devices collect voltage, current, load and fault alarm data through SCADA / PMU / smart meters, and aggregate them to the dispatching / data platform for cleaning and format unification; on this basis, load and distributed power prediction, state estimation, reactive power and voltage optimization, fault location and other analysis and calculation are carried out, and the analysis results are used for operation and maintenance inspection and dispatching decision. In the traditional technology, there are problems such as non-uniform data standards, insufficient model stability caused by data quality fluctuations, high cost of cross-system data fusion, and limited response and explainability for strong real-time scenarios, which makes it difficult to utilize the electric computer fusion data to meet the current requirements of power grid. SUMMARY

[0003] Therefore, it is necessary to provide an electric computer fusion data application method, device and computer equipment based on full life cycle quality guarantee, which can guarantee the utilization degree of electric computer fusion data to meet the current requirements of power grid.

[0004] In a first aspect, the present application provides an electric computer fusion data application method based on full life cycle quality guarantee, comprising:

[0005] The electric computer raw data of the computing power network is integrated to obtain an electric computer fusion initial data set;

[0006] The space-time three-dimensional field in the electric computer fusion initial data set is processed according to a standard system to obtain a standardized data set;

[0007] The electric computer related data in the standardized data set is coupled and characterized to obtain an electric computer coupling portrait set;

[0008] According to the standardized data set and the electric computer coupling portrait set, the computing power network is cooperatively optimized and arranged to obtain integrated planning information of regulation and transaction;

[0009] The strategy elements of the integrated planning information of regulation and transaction are converted into instructions to obtain a multi-scale execution instruction set.

[0010] In a second aspect, the present application further provides an electric computer fusion data application device based on full life cycle quality guarantee, comprising:

[0011] The data integration module is configured to normalize and integrate the original computing data of the computing power network to obtain an initial data set of computing fusion.

[0012] The system processing module is configured to perform standard system processing on the spatio-temporal three-dimensional field in the initial data set of computing fusion to obtain a standardized data set.

[0013] The coupling representation module is configured to perform coupling representation learning on the computing-related data in the standardized data set to obtain a computing coupling portrait set.

[0014] The collaborative optimization module is configured to perform collaborative optimization arrangement on the computing power network according to the standardized data set and the computing coupling portrait set to obtain an integrated plan information of regulation and transaction.

[0015] The instruction generation module is configured to perform instruction conversion on the strategy elements of the integrated plan information of regulation and transaction to obtain a multi-scale execution instruction set.

[0016] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements any step of the method for applying computing fusion data based on full-life-cycle quality assurance when executing the computer program.

[0017] The method, device and computer device for applying computing fusion data based on full-life-cycle quality assurance can eliminate multi-source heterogeneous differences, improve data consistency and comparability by normalizing and integrating the original computing data of the computing power network and performing standard system processing on the spatio-temporal three-dimensional field. On this basis, the computing-related data is subjected to coupling representation learning to construct an accurate computing coupling portrait that accurately depicts the correlation between supply, demand and constraints, thereby significantly enhancing the robustness of prediction evaluation and decision-making. The standardized data set and the coupling portrait are used for collaborative optimization arrangement to form an integrated plan of regulation and transaction, thereby improving global resource utilization and service quality, reducing energy consumption and operation cost. Finally, the strategy elements are instructed to generate a multi-scale execution instruction set, thereby breaking the "data-model-arrangement-execution" closed loop, realizing end-to-end automation and real-time closed loop control, supporting elastic scaling and safe and reliable operation of cross-domain heterogeneous computing power, and ensuring that the utilization degree of computing fusion data cannot meet the current requirements of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed in the embodiment or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0019] Figure 1 An application environment diagram of a computer integrated data application method based on full life cycle quality assurance in an embodiment;

[0020] Figure 2 A flowchart of a computer integrated data application method based on full life cycle quality assurance in an embodiment;

[0021] Figure 3 A structural block diagram of a computer integrated data application device based on full life cycle quality assurance in an embodiment;

[0022] Figure 4 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0023] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0024] The computer integrated data application method based on full life cycle quality assurance provided by the embodiments of the present application can be applied in the application environment as shown in Figure 1 . Wherein, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. Wherein, the server 104 can be realized by an independent server or a server cluster composed of multiple servers.

[0025] In an exemplary embodiment, as shown in Figure 2 , a computer integrated data application method based on full life cycle quality assurance is provided. Taking the server in Figure 1 as an example, the method includes the following steps 202 to 210. Wherein:

[0026] Step 202, normalizing and integrating the computer original data of the computing power network to obtain a computer integrated initial data set.

[0027] Step 204, processing the space-time three-dimensional field in the computer integrated initial data set according to the standard system to obtain a standardized data set.

[0028] Step 206, coupling and characterizing the computer related data in the standardized data set to obtain a computer coupling portrait set.

[0029] Step 208, according to the standardized data set and the computer coupled image set, the algorithm network is cooperatively optimized and arranged, and the regulation and transaction integrated plan information is obtained.

[0030] Step 210, the strategy elements of the regulation and transaction integrated plan information are converted into instructions, and a multi-scale execution instruction set is obtained.

[0031] Among them, the algorithm network is a whole system and its operation control domain for algorithm supply and dispatch composed of data centers, edge nodes and their power supply and distribution, communication and control facilities.

[0032] Among them, the computer original data is the measurement and log data directly collected and unprocessed in the algorithm network, such as IT load, electromechanical / refrigeration power, PUE metering, energy storage SOC / power, timestamp and equipment / rack position, etc.

[0033] Among them, the standardization integration is a standardization integration process of clock alignment, deduplication, unit and caliber unification, anomaly / missing repair and establishment of main index and traceability label on multi-source heterogeneous original data.

[0034] Among them, the computer fusion initial data set is a basic data set with uniform structure and version and traceability information obtained after standardization integration.

[0035] Among them, the space-time three-dimensional field is a three-field set in the computer fusion initial data set for representing time dimension, space position dimension and index value dimension.

[0036] Among them, the standard system processing is a standardization mapping and caliber unification process on the space-time three-dimensional field according to the unified time zone / coordinate / index caliber.

[0037] Among them, the standardized data set is a data set obtained by standard system processing, which is time-aligned, space-coded consistent, index-caliber unified and directly modelable.

[0038] Among them, the computer related data is a subset of fields and records in the standardized data set directly related to the coupling of electricity and computer, such as IT load, PUE, electromechanical energy consumption and energy storage state, etc.

[0039] Among them, the coupling representation learning is a modeling process of joint representation learning and prediction generation on the computer related data under the constraints of time series and spatial topology, usually adding physical consistency constraints such as energy balance and power boundary.

[0040] Among them, the computer coupled image set is a result set encapsulated by the representation vectors and corresponding prediction curves generated by the coupling representation learning according to the unified identifier.

[0041] Among them, the collaborative optimization arrangement is a process of jointly considering power, computing power and carbon constraints in a unified feasible region, and solving and arranging resource interaction curves and regulation / trading schemes.

[0042] Among them, the regulation and transaction integrated plan information is an executable plan set output by the collaborative optimization arrangement, including time sequence plan curves, threshold values and execution parameters, etc.

[0043] Among them, the strategy element is a parameterized decision element in the regulation and transaction integrated plan information for specific execution units, including limits, ratios, curve segments, trigger conditions and priorities, etc.

[0044] Among them, the instruction conversion is a process of compiling strategy elements according to interface protocols and security policies into controllable and auditable control instructions.

[0045] Among them, the multi-scale execution instruction set is a control instruction set organized by seconds / minutes / hours and with version and effective / rollback labels, used for delivery and closed-loop execution.

[0046] Specifically, the electric and computing raw data of the computing power network are authenticated and listed as data sources (IT load, PUE metering, mechanical / electrical power, energy storage SOC and charging / discharging power, equipment and rack location information and time stamp), then the time zone is normalized and the event time is aligned (a unified sampling grid is established and the original resolution index is retained), and the main index and association key are constructed according to the five-level coding of park-building-machine room-rack-equipment / circuit. Subsequently, outlier detection and missing repair (outlier rejection, short gap forward / linear fill, long gap model fill) are implemented, and the power / energy / temperature and humidity dimensions and caliber are unified and consistent indexes are derived (such as power-energy integral consistency, PUE caliber uniformity), while quality labels and traceability metadata (source, version, confidence, processing flow) are generated. Finally, it is represented by time sequence fact table+dimension table+metadata table sedimentation, and the electric and computing fusion initial data set is obtained.

[0047] The standard system processing is performed on the space-time three-dimensional field in the computer fusion initial data set, that is, the time field, space field and index field in the computer fusion initial data set are processed in a standard system. First, three types of field lists and mapping rules are generated in the data set. First, the time dimension is unified in time zone and calendar, the format is standardized (YYYY-MM-DD HH:MM:SS), it is aligned to the unified sampling grid (such as 1 minute), and the original timestamp is kept as a secondary key, and year / season / month / day / time / minute level indexes are generated; second, the space dimension is bound with the device / circuit and “park-building-machine room-rack-equipment” five-level code, and the spatial coordinates and power supply / cooling topological relationship table is established; third, the index dimension is unified in unit and precision, the standard calculation range (such as PUE=total energy consumption of facilities / IT energy consumption) is defined, the dimension and enumeration constraint are completed, and the field is renamed. Finally, it is remapped and written according to {time key × space key × index key}, version number and traceability label are written, and the standardized data set with consistent range, unified coordinates and time sequence alignment is output.

[0048] The IT load, PUE, electromechanical power, energy storage SOC and other fields in the standardized data set are sampled by sliding window (aligning the unified time grid and distinguishing the input window / prediction window), and the time sequence sample set is obtained. Then, the cross-scale feature construction (trend / season and multi-order lag, start-stop / mutation indication, fluctuation intensity and IT load × PUE interaction) is performed on the time sequence sample set, and normalization is completed to obtain the feature sample set. According to the device-circuit-rack-machine room relationship table in the standardized data set, the spatial adjacency and coupling edge weight are generated to obtain the spatial topology graph. On this basis, the feature sample set is input as the node, and the spatial topology graph is input as the edge structure. The joint network of time sequence attention encoder-graph message passing is used for representation learning, and the physical consistency regularization is added according to the energy balance, power boundary and SOC upper and lower limit rules in the standardized data set. End-to-end training and verification are performed to obtain the representation vector group and the corresponding load / PUE prediction curve group. Finally, the representation vector group and the prediction curve group are unified and packaged (with version number, time / space anchor point and confidence interval), and the computer coupling image set is obtained.

[0049] The upper limit of capacity, climbing / starting and stopping, energy conservation, network / thermal constraints, SOC upper and lower bounds, and PUE indicators are extracted from the standardized data set, and a unified feasible region is constructed according to a unified time grid. Then, the prediction curve and confidence interval in the computer coupled image set are mapped to an uncertainty scenario set and a robust set is formed. The multi-objective robust solution and hierarchical coordination (such as first solving the Pareto approximation, and then refining according to priority / epsilon-constraint) are carried out in the unified feasible region with cost / energy efficiency / smoothness / carbon intensity and other multi-objectives. The executable resource interaction curve and strategy parameter are generated under the premise of ensuring physical consistency. The consistency of the results of seconds / minutes / hours is checked and versioned, and the integrated planning information of regulation and transaction (including time sequence planning curve, key threshold and execution parameter) is output.

[0050] The strategy elements are classified and grouped according to the object list and time granularity in the integrated planning information of regulation and transaction (forming a "second / minute / hour - execution unit - parameter domain" to be compiled list), and then the template compilation is completed according to the instruction template and parameter constraint attached to the integrated planning information of regulation and transaction (the plan curve is time-aligned and discretized, the threshold is quantized and tolerance packaged, and the trigger logic is conditionally filled). The candidate instruction set is obtained. According to the priority and dependency relationship built in the integrated planning information of regulation and transaction, the consistency check and conflict resolution of the candidate instruction set are carried out, and the cross-scale coordination is completed (to ensure that the second-minute-hour instructions do not conflict and can be executed in the same time slice). The instruction set that passes the check is formed. According to the interface mapping and safety strategy specified in the integrated planning information of regulation and transaction, the target mapping, batch packaging, signature and version marking (including effective / invalid time, rollback point and audit hash) of the instruction that passes the check are carried out. The multi-scale execution instruction set and its check summary / return specification are output as the only basis for issuance and closed loop.

[0051] In the above-mentioned method for computer fusion data application based on whole life cycle quality guarantee, the regularization integration of algorithm network computer original data and the standard system processing of space-time three-dimensional field eliminate the differences of multi-source heterogeneous data, improve the consistency and comparability of data. On this basis, the computer related data is coupled and characterized to learn, and the computer coupled image that accurately describes the correlation of supply, demand and constraints is constructed, which significantly enhances the prediction evaluation and decision robustness. With the help of standardized data set and coupled image, collaborative optimization arrangement is carried out to form integrated planning of regulation and transaction, which improves the overall resource utilization and service quality, reduces energy consumption and operation cost. Finally, the strategy elements are instructed, and the multi-scale execution instruction set is generated to break through the "data - model - arrangement - execution" closed loop, realize end-to-end automation and real-time closed loop control, support elastic expansion and safe and reliable operation of cross-domain heterogeneous computing power, and ensure that the utilization degree of computer fusion data cannot meet the current requirements of power grid.

[0052] In an exemplary embodiment, the computer-related data in the standardized data set is coupled with feature learning to obtain a computer coupling portrait set, including steps 302 to 306. Among them:

[0053] Step 302, the computer-related data in the standardized data set is processed by feature distillation to obtain a coupled feature vector set.

[0054] Step 304, the vector sequence in the coupled feature vector set is analyzed by sequence inference to obtain a coupled prediction curve set.

[0055] Step 306, according to the coupled feature vector set and the coupled prediction curve set, the computer-related data is constructed by portrait description to obtain a computer coupling portrait set.

[0056] Among them, the feature distillation process is to compress the computer-related multi-source features into low-dimensional representations with low noise and portability through self-supervised / constrained learning under the constraints of unified time grid and spatial topology.

[0057] Among them, the coupled feature vector set is a low-dimensional vector set output by feature distillation and organized by object and time slice, used to describe the coupling state of power and computing power.

[0058] Among them, the vector sequence is a continuous vector time sequence obtained by reorganizing the coupled feature vector set in chronological order.

[0059] Among them, the sequence inference analysis is an analysis process of causal multi-step prediction and uncertainty estimation of the vector sequence and projection to the physical feasible region.

[0060] Among them, the coupled prediction curve set is the time series prediction curve and its confidence interval of the target such as load, PUE, SOC obtained by sequence inference analysis, with time / space anchor points.

[0061] Among them, the portrait description construction is to input the coupled feature vector and the prediction curve, calculate the coupling strength, sensitivity and risk band, complete the typing and packaging, and form the portrait results that can be used for collaborative optimization.

[0062] Specifically, sliding window time series samples are constructed for the computer-related data in the standardized dataset (such as IT load, PUE, mechanical and electrical power, energy storage SOC, etc.) on a uniform sampling grid (distinguishing between input window L and prediction window H), then cross-scale and interaction features (trend / season, multi-order lag, rate of change and volatility intensity, IT load x PUE, energy balance difference, device start-stop indication, etc.) are calculated within the window and device-level robust normalization is completed. According to the device-rack-room relationship in the standardized dataset, spatial adjacency is generated, and the self-supervised distillation network of mask auto-encoding + contrastive learning is input with "time series segment as input, spatial adjacency as constraint", and physical consistency regularization is added, which is composed of energy conservation, power boundary and SOC upper and lower limit generated by the standardized dataset. The low-dimensional representation of each object / time slice is packaged with time / space anchor points, field mapping and quality weights to obtain a set of coupled representation vectors.

[0063] The individual representation vectors in the set of coupled representation vectors are reorganized into a vector sequence in time sequence with "object-time grid" (set prediction step H and sliding step S, and keep time / space anchor points and quality weights), then the vector sequence is predicted by a causal time series decoder (determine the hyper-participation early-stop threshold by sliding window rolling start cross-validation), and the confidence interval is obtained by split-conformal calibration of the training residual. Then sequence smoothing and feasible band projection are applied to the initial prediction (smoothing to suppress jitter with total variation constraint, feasible band generated from historical quantile intervals of the same vector sequence, used to remove out-of-range points), output coupled prediction curve set covering targets such as load, PUE, SOC, etc., with validity period, interval confidence and backtest indicators.

[0064] Based on the premise of the set of coupled representation vectors and the set of coupled prediction curves, align and fuse them one by one according to object-time slice (inherit time / space anchor points and confidence), form aligned samples, then calculate portrait indicators on the aligned samples, including coupled strength (such as mutual information / Granger causality), sensitivity and elasticity (such as ∂PUE / ∂IT load, ∂energy consumption / ∂computing power), steady-state / transient segment identification, and risk band obtained according to prediction interval width and feasible band boundary. Then, according to the above indicators, the objects are typed and prototype extraction is performed (such as unsupervised clustering to obtain "high load-high PUE" "peak shifting-charging" prototypes), and typical curves, key thresholds and applicable strategy domains are extracted for each prototype. Finally, prototype identification, typical curves, key thresholds, sensitivity matrix, risk band, validity period and quality score are unified and encapsulated according to a unified structure, and version number and provenance hash are written, outputting a computer-coupled portrait set that can be directly called by collaborative optimization.

[0065] In this embodiment, through the chain processing of "feature distillation → sequence inference → image construction", multi-source heterogeneous computer data can be compressed into a low-dimensional and noise-resistant coupled representation vector set under the unified space-time reference, significantly reducing the influence of caliber difference on modeling and reducing the computational complexity; On this basis, a coupled prediction curve set with confidence interval and physical consistency constraint is obtained, which improves the feasibility and robustness of short-term and medium-term prediction; Then the vector and curve are combined to form a computer coupled image set for image description, which structures and solidifies key elements such as coupling strength, sensitivity and risk, enhances explainability and auditability, so that subsequent collaborative optimization can directly consume high-quality images and predictions, shorten the link from data to executable plans, improve the adaptability to fluctuations and extreme working conditions, and realize end-to-end closed loop and traceability.

[0066] In one exemplary embodiment, sequence inference analysis is performed on the vector sequence in the coupled representation vector set to obtain a coupled prediction curve set, including steps 402 to 408. Among them:

[0067] Step 402, causal reversible coding is performed on the vector sequence in the coupled representation vector set to obtain a causal latent representation sequence.

[0068] Step 404, the causal latent representation sequence is substituted into the network space-time graph neural differential equation for continuous inference to obtain a continuous time latent trajectory.

[0069] Step 406, the energy conservation and power residual of the continuous time latent trajectory are projected for physical consistency to obtain a physically consistent latent trajectory.

[0070] Step 408, the physically consistent latent trajectory is diffused and consistent decoded to obtain a coupled prediction curve set.

[0071] Wherein, the causal reversible coding is a process of mapping the vector sequence to a stable and restorable latent representation at each time point by reversible transformation under the constraint of only relying on historical information.

[0072] Wherein, the causal latent representation sequence is a latent state sequence output by causal reversible coding and one-to-one corresponding to the original timestamp, which preserves causality and can be bidirectionally restored.

[0073] Wherein, the network space-time graph neural differential equation is a neural vector field that models the continuous evolution of state over time on a topology graph such as a device / rack, used to describe the space-time coupling dynamics.

[0074] Wherein, the continuous inference is a process of taking the latent initial state as the starting point, performing numerical integration along the space-time neural vector field, and obtaining the evolution result on the continuous time.

[0075] wherein the continuous-time latent trajectory is a set of continuous-time latent state curves obtained by continuous inference and continuously changing over time in the entire prediction interval.

[0076] wherein the physically consistent projection is a process of projecting the continuous-time latent trajectory back to the feasible set according to the endogenous constraints such as energy conservation, power boundary, SOC range, etc.

[0077] wherein the physically consistent latent trajectory is a continuous-time latent state trajectory satisfying the power and equipment physical constraints after the physically consistent projection.

[0078] wherein the diffusion-consistent decoding is a process of decoding the latent state into observable time series (such as load, PUE, SOC) using a consistent / diffusion generation model with the physically consistent latent trajectory as a condition.

[0079] Specifically, a "history-current" sliding window is constructed according to a unified time grid, and the history segment is input into a lightweight time series network to generate the affine parameters and masks required at each time. Then, the vector sequence in the coupled representation vector set is sequentially added at each time to perform the reversible flow of the causal mask (for example, a normalized flow with a lower triangular dependence structure), ensuring that the transformation at each time only depends on its historical information and maintains bidirectional reversibility. Perform with likelihood consistency as the main target, while cooperating with inverse mapping stability check and Lipschitz / norm regularization to improve numerical stability and reversibility, finally output a causal latent representation sequence corresponding to the original sequence for each object and each time slice, and attach time anchor points and quality weights.

[0080] Taking the last observed latent state in the causal latent representation sequence as the starting point, the spatio-temporal graph neural differential equation vector field model trained and parameter frozen under the same standardized caliber is called, and the latent state is gradually advanced in the prediction interval through an adaptive step numerical integrator. The vector field has built-in Hamiltonian-dissipative structure decomposition and Lipschitz stability constraints to ensure the interpretability and numerical stability of long-time integration. After integration, align the results on the continuous time back to the unified time grid and attach the time anchor points and quality weights, and output the continuous-time latent trajectory.

[0081] After discretizing the continuous-time latent trajectory according to the unified time grid and calculating the energy conservation and power boundary residuals of each time slice, the full-time-domain feasible set (including energy balance, power and ramping limits, SOC upper and lower bounds, and necessary start and end constraints) is constructed according to the endogenous caliber defined in the previous sequence. The projection solution is performed on the discrete trajectory according to the "minimum disturbance" criterion, and an interpretable proximal mapping / projection operator or small-scale iterative quadratic programming is preferentially used until the residual meets the threshold. To maintain continuity and stability in time, a lightweight time-domain smoothing control is added simultaneously to suppress jitter and jitter, and finally the physically consistent latent trajectory is output and the original time anchor points and confidence indicators are inherited.

[0082] Aligning the physically consistent latent trajectories to a uniform time grid as a conditional input, calling the diffusion consistency decoder trained and frozen at the same standardized caliber to perform one-shot / few-step decoding, obtaining the candidate sequences (load, PUE, SOC, etc.) in the observation domain. According to the feasible band generated by the pre-caliber, the candidate sequences are boundary cropped and light smoothing is applied to suppress local jitter, and the confidence interval and validity period label are obtained through the endogenous uncertainty or quantile calibration of the decoder, and the time / space anchor and quality weight are filled for the results, and the coupled prediction curve set is output.

[0083] In this embodiment, by mapping the vector sequence into stable and restorable latent states through causal reversible encoding, the problems of "only looking at the future / information leakage" and long-term error accumulation are significantly suppressed; then by continuous inference of network spatio-temporal graph neural ordinary differential equations, the coupled dynamics of devices-racks-machine rooms are described in the continuous time dimension, improving the ability to capture non-stationary and cross-scale behavior and reducing numerical bias caused by discrete step size; then the physically consistent projection is performed on the continuous trajectory, which endogenously implements energy conservation, power boundary and SOC range to the prediction results, avoiding infeasible solutions and "good-looking but unexecutable" curves; finally, the coupled prediction curve with confidence interval is output by diffusion consistency decoding, taking into account accuracy, stability and uncertainty quantification, and overall realizing an end-to-end closed loop from reversible representation to continuous dynamics to physical check to observable decoding, improving the explainability and auditability of the prediction, and ensuring that the results can be directly used for regulation and transaction arrangement.

[0084] In one exemplary embodiment, the causal latent representation sequence is substituted into the network spatio-temporal graph neural ordinary differential equation for continuous inference to obtain a continuous time latent trajectory, including steps 502 to 506. Among them:

[0085] Step 502, graph spatio-temporal Laplace embedding is performed on the causal latent representation sequence to obtain a graph spatio-temporal latent feature sequence.

[0086] Step 504, the graph spatio-temporal latent feature sequence is substituted into the Hamilton-dissipative neural ordinary differential equation to construct a vector field, obtaining a conservation-dissipation coupled vector field.

[0087] Step 506, measure-preserving symplectic integration is performed on the continuous time flow of the conservation-dissipation coupled vector field to obtain a reversible continuous time latent trajectory.

[0088] Step 508, counterfactual consistent reparameterization is performed on the reversible continuous time latent trajectory to obtain a continuous time latent trajectory.

[0089] The graph spatio-temporal Laplacian embedding is an embedding method that applies Laplacian / diffusion filtering to the node features of each time slice on device-rack-room topologies to reduce noise and strengthen structural consistency.

[0090] The graph spatio-temporal latent feature sequence is a latent feature vector sequence that is arranged in chronological order after graph spatio-temporal Laplacian embedding and corresponds one-to-one to time / space anchors.

[0091] The Hamiltonian-dissipative neural ordinary differential equation is a continuous-time model parameterized by a neural network on a graph topology that decomposes the system continuous dynamics into a Hamiltonian part that maintains structure and a dissipative part that introduces decay.

[0092] The conservative-dissipative coupled vector field is a state evolution direction field generated by the above equation and changing with time and graph structure, containing both conservative and dissipative components.

[0093] The continuous-time flow is a family of trajectories formed by the continuous evolution of the state along the coupled vector field over time.

[0094] The measure-preserving symplectic integral is an integral method that tries to maintain the phase space measure / symplectic structure as much as possible when numerically solving the continuous-time flow, in order to reduce drift and enhance reversibility.

[0095] The reversible continuous-time latent trajectory is a latent state trajectory that can be approximately consistently reproduced in forward and backward calculations through the measure-preserving symplectic integral.

[0096] The counterfactual consistent reparameterization is a process of minimum disturbance time and amplitude reparameterization of the reversible trajectory to meet the preset physical caliber and key scenario (counterfactual) constraints.

[0097] Specifically, according to the causal latent representation sequence, construct a spatio-temporal topology graph according to the hierarchical relationship and physical / logical connection of device / circuit-rack-room-park (determine nodes, edges, their weights and directions, and complete adjacency matrix normalization and timestamp alignment), then perform multi-order Laplacian spectrum / diffusion filtering embedding on the node representation of each time slice to enhance structural consistency and suppress noise, and through time consistent processing (such as sliding window smoothing and truncated alignment) to ensure the continuous stability of adjacent time slice representations while preserving mutation markers, output dimensions and additional time / space anchors and quality weights, forming a graph spatio-temporal latent feature sequence corresponding one-to-one to the original timestamp.

[0098] The generator trained under the same caliber and with frozen parameters is used to decouple the latent features of each time slice and each node into two channels: one channel generates Hamiltonian latent components and learns the antisymmetric mapping to obtain the Poisson tensor (ensuring structural conservation and reversible transmission), and the other channel generates dissipative latent components and learns the symmetric positive definite mapping to obtain the dissipative matrix (ensuring convergence and steady-state stability). Then, according to the established coupling rule, the two types of components are structured into a conservative-dissipative coupled vector field that changes with time and graph topology, and geometric consistency and stability constraints (such as antisymmetric / symmetric positive definite, Lipschitz bounded, causal dependence limited to the current and history) are imposed during generation to ensure interpretability, integrability, and numerical robustness. The generated vector field is supplemented with time / space anchors and version identifiers.

[0099] According to the unified time grid, the prediction interval and tolerance threshold are set, and the adaptive step-size measure-preserving symplectic integrator is enabled to time-propagate the conservative-dissipative coupled vector field; during integration, the conservative-dissipative splitting (symplectic update for conservative components and stabilization update for dissipative components) and step back / error control mechanism are adopted to continuously monitor the volume approximation conservation and energy drift limit, and unqualified steps are removed through forward-backward reversibility checking. When the full-interval error meets the threshold, the continuous results are aligned back to the unified time grid and supplemented with time / space anchors and quality labels, and the reversible continuous-time latent trajectory is output.

[0100] According to the established endogenous caliber and key event templates (such as energy conservation, power / climb boundary, SOC endpoint, and maintenance time window), a consistent target set is generated, and then under the "minimum disturbance" criterion, the reversible continuous-time latent trajectory is executed for time scale re-labeling and amplitude band alignment (preferably fine-tuning the phase and amplitude of key moments to make the reversible continuous-time latent trajectory fall within the feasible band and maintain smooth transition between segments), and the constraint satisfaction and continuity stability are checked through gradual checking cycles, and if necessary, only the local segment is slightly rolled back and realigned. When all consistency targets are met and the trajectory continuity passes the check, the results are aligned back to the unified time grid and supplemented with time / space anchors and version labels, and the continuous-time latent trajectory that meets the causal and physical caliber is output.

[0101] In this embodiment, by embedding structured denoising and consistency enhancement in topology and time dimension with graph space-time Laplacian, the coupling distinguishability across devices / rooms is significantly improved; then by constructing a vector field that maintains structural invariants and contains stable convergence mechanism with Hamiltonian-dissipative neural ordinary differential equation, the interpretable modeling of complex non-stationary dynamics is realized and divergence is suppressed; subsequently by significantly reducing numerical drift and maintaining approximate reversibility in long-time integration with measure-preserving symplectic integration, stable and replayable potential trajectories are obtained; finally by reparameterizing trajectories to align to physical and scenario constraints with counterfactual consistency, the results naturally satisfy executable criteria such as energy / power. It can have stronger robustness and interpretability in long-term and cross-scenario, significantly improve the accuracy and feasibility of subsequent decoding and prediction, reduce "good-looking but unexecutable" output, and facilitate auditing and tracing.

[0102] In one exemplary embodiment, the graph space-time latent feature sequence is substituted into the Hamiltonian-dissipative neural ordinary differential equation to construct a vector field, and a conservation-dissipation coupled vector field is obtained, including steps 602 to 608. Among them:

[0103] Step 602, symplectic-dissipative decomposition is performed on the graph space-time latent feature sequence to obtain Hamiltonian latent components and dissipative latent components.

[0104] Step 604, anti-symmetric parameterization mapping and differentiable decomposition mapping are performed on the graph space-time latent feature sequence to obtain a learnable Poisson tensor and a symmetric positive definite dissipation matrix.

[0105] Step 606, according to the Hamiltonian latent component and the dissipative latent component, the learnable Poisson tensor and the symmetric positive definite dissipation matrix are geometrically consistent fitting to obtain a geometrically consistent parameter set.

[0106] Step 608, according to the geometrically consistent parameter set, the Hamiltonian-dissipative vector field is structured synthesized to obtain the conservation-dissipation coupled vector field.

[0107] Wherein, the symplectic-dissipative decomposition is a process of decoupling and dividing the system or feature according to "conservative reversible" (Hamiltonian) and "decay stable" (dissipation) two channels.

[0108] Wherein, the Hamiltonian latent component is the energy pattern and dynamic mode extracted from the data that preserves structure and reversible transmission.

[0109] Wherein, the dissipative latent component is the stable dynamics mode extracted from the data that monotonically decays and tends to be stable.

[0110] Wherein, the anti-symmetric parameterization mapping is to map the input feature to a parameter object that satisfies the anti-symmetric structure, so as to carry the generation mode of Poisson (Hamiltonian) geometry.

[0111] wherein the differentiable decomposition mapping is a differentiable decomposition strategy that maps input features into a structured constrained parameter object (e.g., symmetric positive definite) for end-to-end learning.

[0112] wherein the learnable Poisson tensor is a trainable tensor parameterized by anti-symmetry for generating Hamiltonian vector field and preserving structural invariants.

[0113] wherein the symmetric positive definite dissipation matrix is a trainable matrix parameterized by differentiable decomposition for positive definiteness to describe energy dissipation and convergence stability.

[0114] wherein the geometric consistent fitting is a fitting process that projects and calibrates the above parameters under the constraints of anti-symmetry, symmetric positive definiteness, energy monotonicity, and spatiotemporal smoothness.

[0115] wherein the geometric consistent parameter set is a parameter set (including Poisson tensor and dissipation matrix) that satisfies all constraints and has spatiotemporal anchor points and version identifiers after completing the geometric consistent fitting.

[0116] wherein the structured synthesis is a step of combining the Poisson tensor driven conservation channel and the dissipation matrix driven stability channel to generate a conservation-dissipation coupled vector field according to the given coupling rules.

[0117] Specifically, the graph spatiotemporal latent feature sequence is aligned by "time slice-node" and the time / space anchor points and version labels are retained, and then it is sent into two complementary generation channels in parallel; wherein channel H implants conservation prior (reversible, sourceless, structure invariant) for extracting reversible energy patterns, and channel R implants dissipation prior (convergent, energy monotonic, steady-state reachable) for extracting decaying stable patterns. Channel selection is completed through gating / attention, and decoupling constraints (mutually exclusive, approximately orthogonal, cross-spatiotemporal smoothness) and stability constraints (causal dependence, amplitude bounded, Lipschitz control) are applied to ensure that the outputs of the two channels do not leak each other and stably evolve with topology and time. Finally, the output of channel H is defined as the Hamiltonian latent component, and the output of channel R is defined as the dissipation latent component, and they are output together with the anchor points and quality labels.

[0118] According to the graph spatiotemporal latent feature sequence, two parameterized mappings are established, one of which is to perform anti-symmetric parameterization mapping on the graph spatiotemporal latent feature sequence to generate a learnable Poisson tensor (through anti-symmetry, Poisson structure regularization, spatiotemporal smoothness, and causal dependence constraints to ensure structural conservation and reversible transmission); the other is to perform differentiable decomposition mapping on the graph spatiotemporal latent feature sequence to generate a symmetric positive definite dissipation matrix (through lower triangular decomposition or spectral mapping to ensure positive definiteness, and energy monotonicity and convergence stability constraints are applied), and the results of the two routes are aligned with the time / space anchor points and written with version and quality labels.

[0119] The Hamiltonian latent component and the learnable Poisson tensor are aligned on the conservation side, calibrating their structural consistency (such as antisymmetric structure, stable bearing of local conservation quantities, smooth variation with time and topology) at the same spacetime anchor point. Deviations are projected back to the feasible set with minimal perturbation. The dissipative latent component and the symmetric positive definite dissipation matrix are aligned on the dissipation side, checking and correcting their monotonic decay and steady-state accessibility properties (such as positive definiteness boundary, upper bound on convergence rate, spectral condition number constraint) to ensure stable consistency within different time slices and adjacent topological neighborhoods. Under the same coordinates and topology, the synergistic relationship between "Hamiltonian latent component - Poisson tensor" and "dissipative latent component - dissipation matrix" is jointly tested (such as mutual exclusion / exchangeable rules of conservation and dissipation channels, causal dependence limited to the current and history, cross-scale smoothness), and local violations are gently rolled back and re-projected with the source and magnitude of the correction recorded. The output is a geometric consistency parameter set with confidence scores, constraint satisfaction summaries, time / space anchors and version numbers, which explicitly retains and aligns the latest consistent form of the above four types of data (Hamiltonian latent component, dissipative latent component, learnable Poisson tensor, symmetric positive definite dissipation matrix).

[0120] After aligning and integrity checking the geometric consistency parameter set (including time / space anchors and version labels) by "time slice - node", the learnable Poisson tensor drives the conservation channel and the symmetric positive definite dissipation matrix drives the stable channel. According to the preset coupling rules (such as series-parallel coupling, gated weighting and topologically consistent message aggregation), local evolution operators are generated node by node and consistent fusion is performed along the graph topology. During the fusion process, structural conservation and stability barriers (antisymmetric / symmetric positive definite, energy monotonicity and upper bound, causal dependence limited to the current and history, Lipschitz bounded) and spatiotemporal smoothness constraints are continuously enforced. Local fragments that do not meet the threshold are gently rolled back and re-fused. After completion, a global consistency and version encapsulation (write quality score, anchor and summary) is performed on the global result, and finally an evolution-consistent, interpretable and numerically robust conservation-dissipation coupled vector field is output over time and topology.

[0121] In this embodiment, the reversible structural dynamics is decoupled from the monotonically convergent stabilized dynamics by the sin-dissipation decomposition, so that noise and drift are not misinterpreted as conservation laws; then the antisymmetric parameterized mapping and the differentiable decomposition mapping are used to generate the learnable Poisson tensor and the symmetric positive definite dissipation matrix, respectively, to embed the geometric constraints of "conservation / reversibility" and "dissipation / stability" at the model level, avoiding the generation of physically infeasible vector fields; subsequently, the geometric consistency fitting is used to constrain and project the above four types of data on the entire space-time domain, unifying the requirements of antisymmetry / positivity, energy monotonicity, smoothness, and causality, significantly reducing the parameter drift across time periods and topologies, and improving interpretability and auditability; finally, the resulting conservation-dissipation coupled vector field has both physical consistency and numerical robustness, supporting long-term integration and multi-scale linkage, and can output a more stable, executable, and continuous dynamics description that can be used for subsequent optimization scheduling.

[0122] In an exemplary embodiment, the algorithmic network is collaboratively optimized according to the standardized data set and the computer-coupled image set, obtaining the integrated regulation and transaction planning information, including steps 702 to 706. Among them:

[0123] Step 702, according to the standardized data set and the computer-coupled image set, the physical constraints and carbon boundaries of the algorithmic network are isomorphic modeled to obtain a unified feasible region.

[0124] Step 704, according to the unified feasible region, the collaborative optimization arrangement is solved by diffusion Nash to obtain a strategy-curve joint solution set.

[0125] Step 706, according to the strategy-curve joint solution set, the resource interaction curve and the regulation instruction of the algorithmic network are solved to obtain the integrated regulation and transaction planning information.

[0126] Among them, the physical constraints are the boundary conditions and balance relations in the algorithmic network determined by the physical properties of the devices and the power grid, such as capacity upper limit, climbing / start-stop limit, network and thermal constraint, energy conservation, and upper and lower bounds of energy storage SOC, etc.

[0127] Among them, the carbon boundary is the compliance boundary related to carbon management, including emission factor and accounting caliber, quota / intensity upper limit, green electricity proportion, and traceability requirement, etc.

[0128] Among them, isomorphic modeling is to convert physical constraints, computer rules, and carbon boundaries into equations / inequalities and logic atoms with consistent caliber under the unified time grid and space coding for calculation and solution.

[0129] Among them, the unified feasible region is the decision-making feasible set evolved over time obtained by consistent checking and intersection of various constraints and image information after isomorphic modeling.

[0130] Wherein, the diffusion Nash solution is self-consistent learning and fixed point / homotopy solution in the unified feasible region with diffusion kernel as the update operator to obtain the Nash equilibrium of multi-agent collaborative game.

[0131] Wherein, the strategy-curve joint solution set is the equilibrium solution set containing executable strategy vector and its corresponding time domain running trajectory output by the diffusion Nash solution.

[0132] Wherein, the resource interaction curve is the time sequence curve of power / energy such as electricity purchase and sale, output, charging and discharging, and interruptible load of each agent in the unified feasible region over time.

[0133] Wherein, the regulation instruction is to compile the planned strategy elements into controllable commands according to the interface protocol and safety policy, including effective window, priority, threshold and rollback identifier.

[0134] Specifically, according to the unified time grid and space coding merging caliber (unit, precision, calculation formula is consistent), physical constraints (capacity upper limit, ramping / start-stop, network / thermal constraints, energy storage SOC interval and end point, etc.) and carbon boundaries (emission factor, quota / intensity upper limit, green electricity proportion requirement) are extracted from the standardized data set, and the predicted curve and its confidence interval are mapped from the electrical coupling image set to the time period feasible band and scenario set. Then, the three types of information are uniformly expressed as constraint atoms (equation, inequality, logic switch) and consistency checking and conflict resolution (rule priority, cross-second / minute / hour multi-scale consistency, boundary bottom and abnormal elimination) are performed. The atoms that pass the checking are intersected according to object-period to form a unified feasible region evolving over time, and are written into the effective period, version number and traceability label.

[0135] The source / grid / load / storage / computing participants are instantiated according to the "object-period" and the revenue composition (cost, energy efficiency, smoothness, carbon target, etc.) and the constraints are extracted from the unified feasible region to form the revenue tensor. Self-consistent learning is performed on the diffusion kernel induced by the revenue tensor to satisfy the row randomness, detailed balance and upper bound of spectral radius, and the diffusion Nash transfer operator is obtained. With this transfer operator as the update operator, the relaxation feasible region is gradually tightened to the target feasible region along the homotopy path, and the Anderson acceleration fixed point solver is used for iteration. The residual convergence, KKT feasibility and stability indicators are dynamically monitored until convergence is obtained, and the Nash solution manifold is obtained. Finally, the solution manifold is uniformly decoded and the strategy vector and resource interaction curve facing the execution object are extracted to form the strategy-curve joint solution set with effective period, confidence and compliance checking summary.

[0136] The strategy-curve joint solution set is mapped back to the "object-time period-execution unit" (such as power / load / energy storage / refrigeration / computing power pool), and time alignment and interpolation are performed on the multi-granularity of seconds / minutes / hours to form candidate resource interaction curves and strategy parameters. Then, according to the pre-sequenced unified feasible region and physical caliber, consistency review and conflict resolution are performed (such as power upper limit preceding charging and discharging strategy, curve coverage preceding threshold triggering, and cross-scale non-interference), and dependency sorting and validity / invalidity window arrangement are completed. On this basis, the candidate results are solidified into executable plans, generating purchase and sale of electricity / green electricity clearing, energy storage charging and discharging, interruptible load and computing power arrangement, etc. Plan curves, and simultaneously extracting control parameters such as limit value, threshold, start-stop time window, and priority, and setting tolerance / rollback strategy and emergency occupation to enhance robustness. The versioned encapsulation of the reviewed curves and parameters (including validity period, audit summary, traceability label, and quality score) outputs the integrated control and transaction plan information.

[0137] In this embodiment, through the three-step link of "isomorphic modeling → diffusion Nash solving → plan generation", the prediction and uncertainty in the standardized data and the computer coupled portrait can be uniformly mapped into computable constraint atoms and intersected into a unified feasible region, the physical boundaries and carbon compliance caliber are simultaneously solidified at the model level to avoid caliber inconsistency and rule conflicts; On this basis, the strategy-curve joint solution set obtained by diffusion Nash solving takes into account the cost, energy efficiency and carbon target, which can more stably converge to an interpretable equilibrium solution in multi-agent game and non-convex feasible region compared to traditional linear / heuristic methods; Finally, according to this, resource interaction curves and control instructions are generated synchronously to ensure consistency and executability across multiple scales of seconds / minutes / hours, significantly reducing plan-execution deviation and out-of-bound risk, improving the adaptability to load fluctuations and extreme working conditions, and realizing the full-process auditability, traceability and rapid closed-loop landing.

[0138] In one exemplary embodiment, according to the unified feasible region, the cooperative optimization arrangement is subjected to diffusion Nash solving to obtain a strategy-curve joint solution set, including steps 802 to 808. Among them:

[0139] Step 802, performing a benefit tensor analysis on the unified feasible region to obtain benefit tensor information.

[0140] Step 804, performing self-consistent learning on the diffusion kernel induced by the benefit tensor information to obtain a diffusion Nash transfer operator.

[0141] Step 806, using the Anderson acceleration fixed point solver to solve the diffusion Nash transfer operator to obtain a Nash solution manifold.

[0142] Step 808, decoding the Nash solution manifold for consistency to obtain a strategy-curve joint solution set.

[0143] Wherein, the yield tensorization analysis is the process of unifying the cost, energy efficiency, risk and carbon and other multi-dimensional yield factors in the unified feasible region according to the "subject-object-time period" caliber, and aggregating them into a computable tensor representation.

[0144] Wherein, the yield tensor information is a multi-dimensional yield data set obtained by yield tensorization analysis, with a feasible mask, a scene weight, and an effective period label.

[0145] Wherein, the diffusion kernel is a transition probability matrix (or operator) defined on the object and time period, satisfying the properties of non-negativity and line randomness, used to describe the diffusion migration rule of policy update.

[0146] Wherein, the self-consistent learning is a learning mechanism that iteratively adjusts parameters until it is consistent with the observed (or set) distribution, with the steady state / preference induced by the current model as the reference.

[0147] Wherein, the diffusion Nash transfer operator is a diffusion kernel obtained under the constraints of self-consistent learning and stability, used as an update operator for policy iteration in multi-agent games.

[0148] Wherein, the Anderson acceleration fixed point solver is a fixed point iteration method that uses linear combination of historical iteration information to accelerate convergence.

[0149] Wherein, the homotopy fixed point solving is a process of gradually solving the fixed point from the relaxation feasible region to the target feasible region, in order to improve the convergence in non-convex / multi-solution cases.

[0150] Wherein, the Nash solution manifold is a collection of Nash equilibrium solutions (which can be regarded as a family of equilibrium points) continuously changing with the object and time period on the same problem.

[0151] Wherein, the consistency decoding is the process of restoring the solution manifold to an executable policy vector and time series curve according to the given caliber, and completing the feasible band clipping and cross-scale consistency checking.

[0152] Specifically, according to the index of "subject (source / net / load / storage / algorithm) - object - time period", the standardized cost items (purchase and sale of electricity / start and stop and climbing / life loss, etc.), energy efficiency items (energy consumption intensity / PUE caliber under the use of energy cost), smoothness and volatility risk items, carbon items (emission factor / carbon cost corresponding to quota or intensity upper limit) and constraint violation penalty items in the unified feasible region are sequentially executed caliber unification → dimensionless → scale standardization, and each item is weighted and aggregated according to the scene / confidence weight of the feasible region; At the same time, the mask is applied to the infeasible region (prohibited from participating in solving) and the effective period, version and traceability label are written, and finally the yield tensor which can directly drive the subsequent learning and solving is packaged.

[0153] According to the income tensor information combined with the "object-time period-feasible mask" and the adjacency relationship, the non-negative and row random initial diffusion kernel is constructed (zero for unfeasible positions and smooth start), and then enters the self-consistent learning closed loop; that is, the current diffusion kernel is used to derive the preference / steady state distribution to evaluate the income consistency residual, and the row randomness and detailed balance are forced to be satisfied by row normalization and Schrodinger bridge / Sinkhorn projection, while the spectral radius and mixing rate constraints are applied to prevent degradation, and then the time and space smoothing regularization and entropy regularization are used to suppress noise and overfitting. The loop is iterated until the self-consistent residual, stability and reversibility indicators are met at the same time, and the parameters are frozen, and the diffusion Nash transfer operator (with version number, validity period and verification summary) is output as the strategy update operator.

[0154] According to the diffusion Nash transfer operator (and inherit the unified feasible domain mask and constraint range), the homotopy path and temperature / penalty schedule from "relaxed feasible domain to target feasible domain" are set, and the starting strategy-curve is generated by neutral initialization in the feasible domain. The Anderson acceleration fixed point solver is enabled to advance in the loop of "prediction-update-projection-acceleration": specifically, the next step strategy update is generated using the diffusion Nash transfer operator, the out-of-bound components are removed by projecting according to the feasible domain mask, the numerical stability is controlled by applying damping and step size adaptation, and the accelerated solution is formed by Anderson mixing; During the iteration process, the residual convergence, KKT feasibility, stability / monotonicity indicators are continuously monitored and the acceleration restart is triggered when oscillation occurs, and the time consistency check is performed when the cross-period is advanced, until the homotopy path converges to the target feasible domain and all thresholds are met, and the Nash solution manifold with time / object anchor points, validity period and confidence score is output.

[0155] The equilibrium points of the Nash solution manifold are aligned and instantiated according to "object-time period-granularity (seconds / minutes / hours)", and each is interpreted as a strategy vector and candidate curve (including power / energy, SOC, electricity purchase / sale and interruptible load, etc.) according to the pre-agreed decoding rules / decoders. Then, according to the range of the pre-sequenced unified feasible domain, the candidate results are subjected to feasible band clipping and boundary projection (eliminating out-of-bound and conflicting items), and cross-scale consistency integration is completed (ensuring that seconds-minutes-hours do not contradict each other, priority and dependency are met); For the case of multiple solutions or degenerate solutions, disambiguation and redundancy removal are performed according to the selection criteria (minimum deviation, cost priority or minimum risk), and light smoothing is applied to the time series to suppress jitter. Finally, the valid period, confidence, audit traceability and version label are supplemented for the verified strategy and curve to form a strategy-curve joint solution set.

[0156] In this embodiment, by means of the progressive processing of "benefit tensor quantization → self-consistent diffusion kernel → Anderson accelerated homotopy fixed point → consistent decoding", the multi-dimensional objectives of cost, energy efficiency, risk and carbon in the unified feasible region are converted into a calculable benefit tensor, and the diffusion Nash transfer operator satisfying randomness, detailed balance and spectrum control is learned accordingly, so that stable and fast equilibrium solution is realized in non-convex and multi-agent game scenarios with the help of Anderson acceleration and homotopy path, and the Nash solution manifold evolving with time and object is obtained; further, the consistent decoding directly outputs executable strategy vector and resource interaction curve, ensuring the consistency and auditability of multi-scale and physical caliber across seconds / minutes / hours, significantly reducing the risk of plan boundary and oscillation, shortening the link from constraint to executable scheme, and further improving the convergence robustness and result interpretability.

[0157] In an exemplary embodiment, self-consistent learning is performed on the diffusion kernel induced by the benefit tensor information to obtain a diffusion Nash transfer operator, including steps 902 to 906. Among them:

[0158] Step 902, Markov transformation is performed on the benefit tensor information to obtain random benefit tensor information.

[0159] Step 904, Schrodinger bridge projection is performed on the diffusion kernel induced by the random benefit tensor information to obtain a detailed balance reversible diffusion kernel.

[0160] Step 906, Wasserstein natural gradient calculation is performed according to the reversible diffusion kernel to obtain a diffusion Nash transfer operator.

[0161] Wherein, the Markov transformation is to make the benefit tensor zero after the feasible mask, and then do non-negative and row normalization, so that the sum of each row is 1, thereby obtaining a row random representation that can be used for transfer modeling.

[0162] Wherein, the random benefit tensor information is the benefit tensor obtained by Markov transformation, which satisfies the row sum of 1 (row random) and carries the scene weight / valid period / mask label.

[0163] Wherein, the Schrodinger bridge projection is to project the initial transfer kernel into a set that matches the target edge distribution while maintaining row randomness and reversibility through bidirectional scaling and reweighting iteration.

[0164] Wherein, the detailed balance reversible diffusion kernel is a diffusion transfer kernel that satisfies non-negativity and row randomness, and has equal forward and reverse flow for a certain steady-state distribution (detailed balance), thus being reversible.

[0165] Wherein, the Wasserstein natural gradient calculation is to update the transfer kernel parameters along the natural gradient direction under the optimal transport (Wasserstein) metric, and project back to the constraint set after each step to maintain row randomness and reversibility.

[0166] Specifically, read the feasible mask, scenario weight and confidence in the benefit tensor information, aggregate the benefit tensor information (cost data, energy efficiency, volatility risk information, carbon items and penalty data) into a comprehensive utility score for each "object-period" according to its weight and risk aversion coefficient, and apply weight reduction to high uncertainty items. Then perform monotonic temperature mapping on the comprehensive utility score to get non-negative preference score (mask position is zero, extreme value is clipped mildly), and perform "feasible item + scenario weight" on the all-zero row to get the minimum positive lower bound to avoid absorbing state. Finally, normalize each row so that the sum is 1, and inherit the effective period, version and used hyperparameter record, output random benefit tensor information that meets row randomness.

[0167] After inducing the initial diffusion kernel from the random benefit tensor information according to "object-period" (keeping non-negative and row random, and inheriting the feasible mask), enter the Schrodinger bridge projection iteration: that is, use bidirectional scaling / reweighting to make the start and end edge distribution of the diffusion kernel consistent with the preference distribution implied by the benefit tensor, while using detailed balance constraints to pair the forward and reverse flows into a reversible form, and suppress numerical spikes and out-of-bounds through spectral radius and condition number upper bound, time / space smoothing regularization and mask projection; evaluate the edge consistency residual, flow conservation error and reversibility deviation in each iteration, and continue scaling and moderate temperature adjustment until convergence if the threshold is not reached. After convergence, freeze the parameters and write the effective period, version and check summary, output the detailed balance reversible diffusion kernel that meets detailed balance and is stable and usable.

[0168] Calculate the current path flow and benefit consistency error with the steady-state distribution of the reversible diffusion kernel and the symmetric edge weight, and determine the Wasserstein natural gradient direction on the optimal transport metric induced by its kernel (use the reversible diffusion kernel as the initial solution and geometric preconditioner, and define the local scale of the update with its steady-state distribution weight). Then perform backtracking line search / damped update along the natural gradient to get the candidate transition operator, and perform projection relying on the structure of the reversible diffusion kernel: that is, force to keep row random, detailed balance and spectral radius upper bound, zero out the infeasible positions according to the mask of the kernel, and perform time / space smoothing according to the adjacency weight of the kernel for adjacent periods and topological adjacency; when iterating, continue to evaluate the steady state and flow with the reversible diffusion kernel, monitor the monotonic decrease of consistency error, reversibility and stability indicators, and trigger the trust region / restart according to the kernel structure if necessary to avoid oscillation, and freeze the parameters when the thresholds are met, get the diffusion Nash transition operator.

[0169] In this embodiment, through the three-stage pipeline of "Markov transformation Schroding bridge projection Wasserstein natural gradient calculation", the yield tensor can be first standardized into a row-random probability representation, and the dimensional difference and the risk of absorption state can be eliminated under the constraint of a feasible mask; then the initial diffusion kernel induced by the representation is projected into a reversible kernel set that satisfies edge matching, row randomness and detailed balance, thereby guaranteeing the physical / statistical reversibility and numerical stability of the transition process from the source; finally, the Wasserstein natural gradient update is performed under the optimal transport geometry, and the constraint set is projected back at each step, thereby accelerating the convergence, suppressing oscillation and overfitting, maintaining reversibility and spectral control, and finally obtaining the diffusion Nash transition operator, so that the policy iteration of multi-agent cooperative game is more robust and more interpretable in a non-convex feasible region.

[0170] It should be understood that although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or stages or steps or stages in other steps.

[0171] Based on the same inventive concept, the embodiments of the present application also provide a full-life-cycle quality assurance based computer fusion data application device for implementing the full-life-cycle quality assurance based computer fusion data application method as described above. As shown in the full-life-cycle quality assurance based computer fusion data application device for implementing the full-life-cycle quality assurance based computer fusion data application method as described above. Figure 3 The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, and therefore the specific limitations in one or more full-life-cycle quality assurance based computer fusion data application device embodiments provided below can refer to the limitations of the full-life-cycle quality assurance based computer fusion data application method described above, and will not be described here again.

[0172] Each module in the full-life-cycle quality assurance based computer fusion data application device described above can be realized by software, hardware, and combinations thereof, in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0173] In an example embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in FIG. 1. Figure 4 The computer device includes a processor, a memory, an input / output interface, and a communication interface. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store server data. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with terminals outside through a network connection. The computer program is executed by the processor to implement a computer integrated data application method based on full life cycle quality assurance.

[0174] Those skilled in the art can understand that Figure 4 The structure shown in FIG. 1 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0175] In an example embodiment, a computer device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0176] In an example embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0177] In an example embodiment, a computer program product or a computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the steps in the above method embodiments.

[0178] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0179] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0180] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0181] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A computer-integrated data application method based on full lifecycle quality assurance, characterized in that, The method includes: The raw computing data of the computing network is normalized and integrated to obtain the initial dataset of computing convergence. The raw computing data includes IT load, PUE metering, electromechanical power, cooling power, energy storage SOC and charging / discharging power, equipment and rack location information and timestamps. The normalization and integration is a standardized integration process that performs clock alignment, deduplication, unit and caliber unification, anomaly and missing data repair on the raw computing data, and constructs a master index and association key according to the five-level coding of park-building-computer room-rack-equipment and loop, while generating quality tags and traceability metadata. The spatiotemporal three-dimensional fields in the initial dataset of the computer-integrated computing system are processed according to a standard system to obtain a standardized dataset. The spatiotemporal three-dimensional fields are a set of three types of fields in the initial dataset of the computer-integrated computing system used to represent the time dimension, spatial location dimension, and index value dimension. The spatial dimension binds equipment, loops, and five-level codes to establish a table of spatial coordinates and supply and distribution, and cooling topology relationships. The index dimension unifies the units and precision, standardizes the calculation caliber, defines the dimensions and enumeration constraints, and completes the field renaming. Finally, it is remapped according to {time key × spatial key × index key}, and the version number and traceability label are written to output a standardized dataset with consistent caliber, unified coordinates, and time sequence alignment. Coupled representation learning is performed on the computer-related data in the standardized dataset to obtain a computer-coupled profile set. The computer-related data is a subset of fields and records in the standardized dataset that are directly related to computer-computer coupling. The coupled representation learning is a modeling process that performs joint representation learning on the computer-related data and generates predictions under temporal and spatial topological constraints. It includes: sampling the computer-related data through a sliding window to obtain a temporal sample set; constructing cross-scale features on the temporal sample set and completing normalization to obtain a feature sample set; generating spatial adjacency and coupling edge weights based on the device-loop-rack-data center relationship table in the standardized dataset to obtain a spatial topology graph; using the feature sample set as node input and the spatial topology graph as edge structure, a joint network of temporal attention encoder and graph message passing is used for representation learning. The computer-coupled profile set is a set of results encapsulated with a unified identifier after the representation vectors produced by the coupled representation learning and the corresponding IT load and PUE prediction curves. Based on the standardized dataset and the computer-coupled profile set, the computing power network is collaboratively optimized and orchestrated to obtain integrated control and trading plan information. The collaborative optimization and orchestration is a process of jointly considering electricity, computing power and carbon constraints within a unified feasible domain, and solving and orchestrating resource interaction curves and control in an integrated manner, generating electricity purchase and sale, green electricity clearing, energy storage charging and discharging, interruptible load and computing power orchestration plan curves respectively. The strategy elements of the integrated regulation and trading plan information are converted into instructions to obtain a multi-scale execution instruction set.

2. The method according to claim 1, characterized in that, The process of performing coupled representation learning on the computer-related data in the standardized dataset to obtain a computer-coupled profile set includes: The computer-related data in the standardized dataset are subjected to feature distillation to obtain a coupled representation vector set; Sequence inference analysis is performed on the vector sequences in the coupling characterization vector set to obtain a set of coupling prediction curves; Based on the set of coupling characterization vectors and the set of coupling prediction curves, a profile description is constructed for the computer-related data to obtain the computer coupling profile set.

3. The method according to claim 2, characterized in that, The step of performing sequence inference analysis on the vector sequences in the coupling representation vector set to obtain a coupling prediction curve set includes: The vector sequence in the coupled representation vector set is causally reversibly encoded to obtain a causal latent representation sequence; Substituting the causal latent representation sequence into the network spatiotemporal graph constant differential equation for continuous inference, a continuous time latent trajectory is obtained; Physically uniform latent trajectories are obtained by physically uniformizing the energy conservation and power residuals of the continuous-time latent trajectory. The physical consistency latent trajectory is subjected to diffusion consistency decoding to obtain the set of coupled prediction curves.

4. The method according to claim 3, characterized in that, The step of substituting the causal latent representation sequence into the network spatiotemporal graph frequent differential equation for continuous inference to obtain the continuous-time latent trajectory includes: The causal latent representation sequence is subjected to graph-spatiotemporal Laplace embedding to obtain the graph-spatiotemporal latent feature sequence; Substituting the spatiotemporal latent feature sequence of the graph into the Hamiltonian-dissipative neural ordinary differential equation to construct a vector field, we obtain a conserved-dissipative coupled vector field. The invertible continuous-time latent trajectory is obtained by performing a symplectic integral on the continuous-time flow of the conserved-dissipative coupled vector field. The reversible continuous-time latent trajectory is reparameterized by counterfactual consistency to obtain the continuous-time latent trajectory.

5. The method according to claim 4, characterized in that, The step of substituting the spatiotemporal latent feature sequence of the graph into the Hamiltonian-dissipative ordinary differential equation to construct a vector field, thereby obtaining a conserved-dissipative coupled vector field, includes: The spatiotemporal latent feature sequence of the graph is subjected to symplectic dissipative decomposition to obtain Hamiltonian latent components and dissipative latent components; By performing antisymmetric parameterization mapping and differentiable decomposition mapping on the spatiotemporal latent feature sequence of the graph, a learnable Poisson tensor and a symmetric positive definite dissipation matrix are obtained. Based on the Hamiltonian latent component and the dissipation latent component, the learnable Poisson tensor and the symmetric positive definite dissipation matrix are geometrically consistent and fitted to obtain a geometrically consistent parameter set. Based on the geometrically consistent parameter set, the Hamiltonian-dissipative vector field is structurally synthesized to obtain the conserved-dissipative coupled vector field.

6. The method according to claim 1, characterized in that, The step of collaboratively optimizing and arranging the computing power network based on the standardized dataset and the computer-coupled profile set to obtain integrated regulation and transaction plan information includes: Based on the standardized dataset and the computer-coupled image set, the physical constraints and carbon boundaries of the computing network are modeled in a homogeneous manner to obtain a unified feasible region. Based on the unified feasible region, the collaborative optimization orchestration is solved using diffusion Nash to obtain the joint solution set of policy and curve; Based on the strategy-curve joint solution set, the resource interaction curve and control instructions of the computing power network are solved to obtain the integrated control and transaction plan information.

7. A computer-integrated data application device based on full lifecycle quality assurance, characterized in that, The device includes: The data integration module is used to normalize and integrate the raw computing data of the computing power network to obtain the initial dataset of computing fusion. The raw computing data includes IT load, PUE metering, electromechanical power, cooling power, energy storage SOC and charging / discharging power, equipment and rack location information and timestamps. The normalization and integration is a standardized integration process that performs clock alignment, deduplication, unit and caliber unification, anomaly and missing data repair on the raw computing data, and constructs a master index and association key according to the five-level coding of park-building-computer room-rack-equipment and loop, while generating quality tags and traceability metadata. The system processing module is used to perform standard system processing on the spatiotemporal three-dimensional fields in the initial dataset of the computer-integrated computing system to obtain a standardized dataset. The spatiotemporal three-dimensional fields are a set of three types of fields in the initial dataset of the computer-integrated computing system, which are used to represent the time dimension, spatial location dimension, and index value dimension. The spatial dimension binds equipment, loops, and five-level codes to establish a table of spatial coordinates and supply and distribution, and cooling topology relationships. The index dimension unifies the units and precision, standardizes the calculation caliber, defines the dimensions and enumeration constraints, and completes the field renaming. Finally, it is remapped according to {time key × spatial key × index key}, writes the version number and traceability label, and outputs a standardized dataset with consistent caliber, unified coordinates, and time sequence alignment. The coupling representation module is used to perform coupling representation learning on the computer-related data in the standardized dataset to obtain a computer coupling profile set. The computer-related data is a subset of fields and records in the standardized dataset that are directly related to computer-computer coupling. The coupling representation learning is a modeling process that performs joint representation learning on the computer-related data under temporal and spatial topological constraints and generates predictions. It includes: sampling the computer-related data through a sliding window to obtain a temporal sample set; constructing cross-scale features on the temporal sample set and completing normalization to obtain a feature sample set; generating spatial adjacency and coupling edge weights based on the device-loop-rack-data center relationship table in the standardized dataset to obtain a spatial topology graph; and performing representation learning using a joint network of temporal attention encoder and graph message passing with the feature sample set as node input and the spatial topology graph as edge structure. The computer coupling profile set is a set of results encapsulated with a unified identifier for the representation vectors produced by the coupling representation learning and the corresponding IT load and PUE prediction curves. The collaborative optimization module is used to perform collaborative optimization and orchestration of the computing power network based on the standardized dataset and the computer-coupled profile set to obtain integrated control and trading plan information. The collaborative optimization and orchestration is a process of jointly considering electricity, computing power and carbon constraints within a unified feasible domain, and solving and orchestrating resource interaction curves and control in an integrated manner, generating electricity purchase and sale, green electricity clearing, energy storage charging and discharging, interruptible load and computing power orchestration plan curves respectively. The instruction generation module is used to convert the strategy elements of the integrated regulation and trading plan information into instructions to obtain a multi-scale execution instruction set.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements any one of claims 1 to 6.

Images