A load dynamic prediction evaluation method with deep coupling of computing power and electric power

Abstract

The present application relates to the technical field of computing power load prediction, and discloses a load dynamic prediction and evaluation method for computing power-electricity deep coupling, obtains a multi-dimensional computing power operation parameter set of a target computing power cluster, obtains an electricity state parameter set of a matched electricity supply network, constructs a computing power-electricity coupling correlation mapping model according to the multi-dimensional computing power operation parameter set and the electricity state parameter set, performs rolling prediction on the computing power load demand of the target computing power cluster in a continuous time window, generates an initial load prediction sequence group, determines the computing power supply-demand balance degree, the electricity cost sensitivity and the resource utilization efficiency of a prediction result unit, generates a three-tuple evaluation index structure, performs confidence screening on the initial load prediction sequence group, obtains a credible load prediction sequence set, inputs a dynamic feedback adjustment framework, generates a power supply adjustment suggestion strategy, and improves the cooperative operation level of a heterogeneous computing power cluster and a diversified electricity system under the premise of ensuring the prediction accuracy and the real-time regulation.

Description

Technical Field

[0001] This invention relates to the field of computing power load prediction technology, and more specifically, to a method for dynamic load prediction and evaluation with deep coupling of computing power and electricity. Background Technology

[0002] Computing power-electricity co-optimization refers to a technical system that improves both energy efficiency and computing service quality by coordinating the scheduling of computing resources and the allocation of power supply. It has become a key infrastructure support technology in the rapid development of the digital economy.

[0003] Existing technologies often employ separate management methods, such as independently predicting the load of computing clusters and autonomously scheduling the power network. However, existing solutions lack a dynamic correlation mechanism between computing demand prediction and power supply regulation, resulting in insufficient load prediction accuracy. Furthermore, static threshold strategies are unable to cope with the sudden fluctuations in computing tasks, causing a lag in power resource allocation. On the other hand, simple linear mapping methods ignore the nonlinear characteristics of computing equipment power consumption, leading to the dual problems of energy waste and decreased power supply reliability. Summary of the Invention

[0004] To address the aforementioned technical problems, this application aims to provide a load dynamic prediction and evaluation method with deep coupling of computing power and power, thereby solving the technical difficulties in the prior art where computing power planning and power dispatch are separated and unable to adapt to rapid load changes, resulting in low energy efficiency. It achieves the technical effect of dynamically mapping multi-dimensional computing power parameters and power state parameters based on a deep correlation model, thereby significantly improving the collaborative operation level of heterogeneous computing power clusters and diversified power systems while ensuring prediction accuracy and real-time control.

[0005] To achieve the above objectives, this invention provides a load dynamic prediction and evaluation method with deep coupling of computing power and electricity, comprising: Obtain a multi-dimensional set of computing power operation parameters of the target computing power cluster within a preset monitoring period, wherein the multi-dimensional set of computing power operation parameters includes the computing power resource occupancy sequence of each computing node, the task queue length sequence of each computing node, and the instruction throughput sequence of each computing node; Obtain a set of power status parameters of the power supply network that matches the target computing power cluster, wherein the set of power status parameters includes the load power sequence of the power supply node, the voltage stability sequence of the power supply node, and the power quality sequence of the power supply node. Based on the multi-dimensional computing power operation parameter set and the power state parameter set, a computing power-power coupling correlation mapping model is constructed. Based on the coupled correlation mapping model, the computing load demand of the target computing power cluster in multiple consecutive future time windows is predicted in a rolling manner, and an initial load prediction sequence group is generated. The initial load prediction sequence group includes the predicted computing load and predicted power consumption for each future time window. For each prediction result unit in the initial load prediction sequence group, determine the computing power supply and demand balance, power cost sensitivity, and resource utilization efficiency corresponding to the prediction result unit, and generate a triplet evaluation index structure. Based on the triplet evaluation index structure, the initial load prediction sequence group is screened by confidence level to obtain a reliable load prediction sequence set. The reliable load prediction sequence set is then input into the dynamic feedback adjustment framework to generate a computing power scheduling suggestion strategy for the target computing power cluster and a power supply adjustment suggestion strategy for the power supply network.

[0006] Furthermore, obtain a set of multi-dimensional computing power operation parameters for the target computing power cluster within a preset monitoring period, including: By deploying performance acquisition agents on each computing node in the target computing power cluster, and using a unified timestamp as a reference, time-series data of CPU utilization, GPU utilization, and memory bandwidth utilization are collected for each computing node. The timing data of the CPU utilization rate, the timing data of the GPU utilization rate, and the timing data of the memory bandwidth utilization rate are normalized to generate the computing resource utilization sequence. Record the task submission timestamp, task execution duration, and task resource requirement tag of each computing node within the preset monitoring period, and construct the task queue length sequence; The number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions completed by each computing node within the preset monitoring period is counted to generate the instruction throughput sequence.

[0007] Furthermore, based on the multi-dimensional computing power operation parameter set and the power state parameter set, a computing power-power coupling correlation mapping model is constructed, including: Time alignment is performed on the multi-dimensional computing power operation parameter set and the power status parameter set to obtain a synchronized computing power-power joint parameter matrix. Based on the computing power-power joint parameter matrix, a computing power-side feature vector group and a power-side feature vector group are extracted. The computing power-side feature vector group includes load fluctuation characteristics, task type characteristics, and resource contention characteristics. The power-side feature vector group includes power ramp-up characteristics, harmonic content characteristics, and frequency offset characteristics. Determine the mutual information entropy matrix between the computing power-side feature vector group and the power-side feature vector group. The mutual information entropy matrix is ​​used to quantify the synchronous correlation strength between computing power state changes and power state changes. Based on the mutual information entropy matrix, a set of key feature pairs with a correlation strength greater than a preset coupling threshold is selected; A deep temporal prediction network is trained using the key feature pairs. The deep temporal prediction network includes an encoder-decoder structure, wherein the encoder is used to extract the implicit coupling patterns of the key feature pairs, and the decoder is used to reconstruct the computing power-electricity joint state at future time moments. The trained deep temporal prediction network is identified as the computing power-electricity coupling correlation mapping model.

[0008] Furthermore, based on the aforementioned coupled correlation mapping model, the computing load demand of the target computing power cluster is predicted in rolling fashion over multiple consecutive time windows in the future, generating an initial load prediction sequence set, including: Based on the coupled correlation mapping model, the current state vector of the target computing power cluster is determined, wherein the current state vector includes the computing power resource occupancy value and power load value at the last monitoring time. Based on the current state vector, a context-aware feature representation is generated using the encoder part of the coupled association mapping model; The context-aware feature representation is input into the decoder of the coupled association mapping model to obtain the predicted state increment for the next time window; The predicted state increment is superimposed on the current state vector to generate the predicted computing power load and the predicted power consumption for the first future time window; Using the prediction result of the first future time window as the new current state vector, the operations of generating the context-aware feature representation, the input decoder, and the superimposed state increment are iteratively executed to obtain the initial load prediction sequence group of multiple consecutive time windows in sequence.

[0009] Furthermore, for each prediction result unit in the initial load prediction sequence group, the computing power supply and demand balance, electricity cost sensitivity, and resource utilization efficiency corresponding to the prediction result unit are determined, including: Based on the predicted computing power load and the rated computing power capacity of the target computing power cluster, the computing power capacity margin ratio is calculated, and the reciprocal of the computing power capacity margin ratio is used as the computing power supply and demand balance. Obtain the time-of-use electricity price parameters and carbon emission factor parameters of the power supply network in the time window corresponding to the prediction result unit; Based on the predicted power consumption, the time-of-use electricity price parameters, and the carbon emission factor parameters, the comprehensive power cost per unit computing power task is calculated, and the rate of change of the comprehensive power cost is used as the power cost sensitivity. Based on the predicted computing power load and the predicted power consumption, a computing power-to-power conversion efficiency index is calculated. The resource utilization efficiency is determined by the ratio of the computing power-to-power conversion efficiency index to the historical average conversion efficiency of the target computing power cluster.

[0010] Furthermore, based on the triplet evaluation index structure, the initial load forecast sequence set is subjected to confidence level grading and screening to obtain a reliable load forecast sequence set, including: Extract the triplet evaluation index structure of each prediction result unit and construct an evaluation index vector; Calculate the cluster center vector of the evaluation index vector of all prediction result units, and use it as the benchmark evaluation reference point; For each prediction result unit, the Euclidean distance between the evaluation index vector and the benchmark evaluation reference point is calculated to obtain the quantified value of the prediction deviation. The quantified value of the prediction deviation is compared with a preset grading threshold range for confidence level grading and filtering.

[0011] Furthermore, based on the aforementioned triplet evaluation index structure, the initial load forecast sequence set is subjected to confidence level grading and screening to obtain a reliable load forecast sequence set, which further includes: The preset tiered threshold range includes the upper limit of the first-level threshold, the upper limit of the second-level threshold, and the upper limit of the third-level threshold; In response to the prediction deviation quantification value being less than the upper limit of the first-level threshold, the corresponding prediction result unit is marked as a high-confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the first-level threshold and the upper limit of the second-level threshold, the corresponding prediction result unit is marked as a medium confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the second threshold and the upper limit of the third threshold, the corresponding prediction result unit is marked as a low confidence prediction result; If the quantified value of the prediction deviation is greater than the upper limit of the third-level threshold, the corresponding prediction result unit is removed. The prediction result units labeled as high-confidence prediction results and medium-confidence prediction results are integrated into the set of credible load prediction sequences.

[0012] Furthermore, the reliable load prediction sequence set is input into the dynamic feedback adjustment framework to generate a computing power scheduling recommendation strategy for the target computing power cluster and a power supply adjustment recommendation strategy for the power supply network, including: Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the predicted computing power load and the current actual load of the target computing power cluster is analyzed to generate a trend vector of computing power load demand change. Based on the computing power load demand change trend vector, the corresponding task migration scheme, node start / stop scheme and load balancing scheme are matched in the preset scheduling strategy library, and the computing power scheduling suggestion strategy is generated by combining them. Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the predicted power consumption and the current actual power supply of the power supply network is analyzed to generate a power supply and demand gap vector. Based on the power supply and demand gap vector and the time-of-use electricity price parameters, the corresponding generator output adjustment scheme, energy storage charging and discharging scheme and flexible load adjustment scheme are matched in the preset power supply strategy library, and the power supply adjustment suggestion strategy is generated by combining them. The computing power scheduling suggestion strategy and the power supply adjustment suggestion strategy are aligned on the time axis to generate a collaborative execution plan sequence, which contains a joint scheduling instruction set for each future time window.

[0013] Furthermore, it also includes: The unified monitoring panel of the target computing power cluster displays an associated view of the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy.

[0014] Furthermore, a view showing the association between the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy is presented, including: In response to receiving a command to display prediction details for the unified monitoring panel, a sub-window for displaying prediction details pops up in the unified monitoring panel; In the prediction details display sub-window, the generation basis of each prediction result unit in the trusted load prediction sequence set is displayed in layers. The generation basis includes: the original data source of the multi-dimensional computing power operation parameter set, the ranking of the key feature contribution of the coupled correlation mapping model, and the detailed scores of each dimension of the triplet evaluation index structure. In response to receiving a selection instruction for a specific feature in the ranking of the contribution of the key features, the time-series evolution curve of the specific feature within the preset monitoring period is expanded in the prediction details display sub-window, and the influence weight value of the specific feature on the current prediction result is marked.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a method for dynamic load forecasting and evaluation with deep coupling of computing power and power. The method involves obtaining a multi-dimensional set of computing power operation parameters for the target computing power cluster; obtaining a set of power status parameters for the matching power supply network; constructing a computing power-power coupling correlation mapping model based on the multi-dimensional computing power operation parameter set and the power status parameter set; performing rolling forecasts of the computing power load demand of the target computing power cluster within a continuous time window to generate an initial load forecast sequence set; determining the computing power supply-demand balance, power cost sensitivity, and resource utilization efficiency of the forecast result units to generate a triplet evaluation index structure; performing confidence screening on the initial load forecast sequence set to obtain a reliable load forecast sequence set; inputting this set into a dynamic feedback adjustment framework to generate power supply adjustment suggestions; and improving the collaborative operation level of heterogeneous computing power clusters and diversified power systems while ensuring forecast accuracy and real-time control. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The diagram shows a flowchart of a load dynamic prediction and evaluation method with deep coupling of computing power and electricity in an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0019] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] The following is a description of preferred embodiments of the present invention in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, an embodiment of the present invention discloses a load dynamic prediction and evaluation method with deep coupling of computing power and electricity, including: S110: Obtain a multi-dimensional computing power operation parameter set of the target computing power cluster within a preset monitoring period, wherein the multi-dimensional computing power operation parameter set includes the computing power resource occupancy sequence of each computing node, the task queue length sequence of each computing node, and the instruction throughput sequence of each computing node. S120: Obtain the power status parameter set of the power supply network that matches the target computing power cluster, wherein the power status parameter set includes the load power sequence of the power supply node, the voltage stability sequence of the power supply node, and the power quality sequence of the power supply node. S130: Construct a computing power-electricity coupling correlation mapping model based on the multi-dimensional computing power operation parameter set and the power state parameter set; S140: Based on the coupled correlation mapping model, the computing load demand of the target computing power cluster in multiple consecutive future time windows is predicted in a rolling manner, and an initial load prediction sequence group is generated, wherein the initial load prediction sequence group includes the predicted computing load and predicted power consumption corresponding to each future time window. S150: For each prediction result unit in the initial load prediction sequence group, determine the computing power supply and demand balance, power cost sensitivity and resource utilization efficiency corresponding to the prediction result unit, and generate a triplet evaluation index structure. S160: Based on the triplet evaluation index structure, the initial load prediction sequence group is screened by confidence level to obtain a reliable load prediction sequence set. The reliable load prediction sequence set is input into the dynamic feedback adjustment framework to generate a computing power scheduling suggestion strategy for the target computing power cluster and a power supply adjustment suggestion strategy for the power supply network.

[0023] In this embodiment, the coupled correlation mapping model is used to characterize the dynamic response relationship between changes in computing load and fluctuations in power consumption.

[0024] In this embodiment, the target computing cluster is a GPU computing cluster containing 100 servers in a data center, and the preset monitoring period is 15 minutes. The power supply network is a 10 kV dedicated substation, and the power supply nodes include the transformer output terminal and the UPS output terminal. The load power sequence records the active power value, such as [850kW, 865kW, ...]. The voltage stability sequence is the voltage deviation percentage, such as [2.1%, 2.3%, ...]. The power quality sequence includes harmonic distortion rate and frequency offset.

[0025] In some embodiments of this application, obtaining a multi-dimensional set of computing power operation parameters of the target computing power cluster within a preset monitoring period includes: By deploying performance acquisition agents on each computing node in the target computing power cluster, and using a unified timestamp as a reference, time-series data of CPU utilization, GPU utilization, and memory bandwidth utilization are collected for each computing node. The timing data of the CPU utilization rate, the timing data of the GPU utilization rate, and the timing data of the memory bandwidth utilization rate are normalized to generate the computing resource utilization sequence. Record the task submission timestamp, task execution duration, and task resource requirement tag of each computing node within the preset monitoring period, and construct the task queue length sequence; The number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions completed by each computing node within the preset monitoring period is counted to generate the instruction throughput sequence.

[0026] In this embodiment, the performance acquisition agent is a lightweight daemon deployed on each server, which obtains raw data by reading the memory controller registers. The unified timestamp uses a Unix timestamp synchronized via the NTP protocol. CPU utilization time-series data is obtained by reading the ratio of user-mode to system-mode time, for example, the sequence [0.45, 0.48, 0.52]. Graphics processor utilization is obtained by querying GPU utilization using the NVML library. Memory bandwidth utilization is obtained by reading the ratio of memory read / write bandwidth to peak bandwidth from the IMC (Integrated Memory Controller) performance counter. Normalization maps the three types of utilization to the [0, 1] interval, and the computing power resource utilization sequence is a three-dimensional vector sequence, for example, [(0.45, 0.60, 0.30), (0.48, 0.65, 0.35), ...]. The task queue length sequence is obtained by listening to task scheduler events, recording the arrival time, estimated execution time (e.g., 300 seconds), and resource tag (e.g., "GPU intensive") for each task. The queue length is the number of tasks currently awaiting scheduling, for example, [12, 15, 18]. The instruction throughput sequence is read from the CPU performance counter (PMC). The number of integer arithmetic instructions is counted by the FIXC0 register, floating-point arithmetic by the FPXC register, and vector arithmetic by the VMXC register, generating a three-dimensional throughput vector, for example, [(5e8, 3e8, 4e8), ...], in instructions per second.

[0027] The beneficial effects of the above technical solution are: the performance acquisition agent realizes distributed data acquisition, the unified timestamp ensures data synchronization of multiple nodes, the normalization process eliminates the difference in units, the task queue length reflects the load pressure, the instruction throughput quantifies the actual computing intensity, and the multi-dimensional computing power parameters provide high-quality input for subsequent modeling.

[0028] In some embodiments of this application, a computing power-power coupling correlation mapping model is constructed based on the multi-dimensional computing power operation parameter set and the power state parameter set, including: Time alignment is performed on the multi-dimensional computing power operation parameter set and the power status parameter set to obtain a synchronized computing power-power joint parameter matrix. Based on the computing power-power joint parameter matrix, a computing power-side feature vector group and a power-side feature vector group are extracted. The computing power-side feature vector group includes load fluctuation characteristics, task type characteristics, and resource contention characteristics. The power-side feature vector group includes power ramp-up characteristics, harmonic content characteristics, and frequency offset characteristics. Determine the mutual information entropy matrix between the computing power-side feature vector group and the power-side feature vector group. The mutual information entropy matrix is ​​used to quantify the synchronous correlation strength between computing power state changes and power state changes. Based on the mutual information entropy matrix, a set of key feature pairs with a correlation strength greater than a preset coupling threshold is selected; A deep temporal prediction network is trained using the key feature pairs. The deep temporal prediction network includes an encoder-decoder structure, wherein the encoder is used to extract the implicit coupling patterns of the key feature pairs, and the decoder is used to reconstruct the computing power-electricity joint state at future time moments. The trained deep temporal prediction network is identified as the computing power-electricity coupling correlation mapping model.

[0029] In this embodiment, the time alignment operation employs the nearest neighbor interpolation method to uniformly align the computing power data (sampling frequency 1Hz) and power data (sampling frequency 10Hz) to a 1Hz time base, forming a computing power-power joint parameter matrix. Rows in the matrix represent time steps, and columns represent features of each dimension. Computing power side feature vector extraction includes: load fluctuation features (calculating the standard deviation and coefficient of variation of the computing power resource occupancy sequence); task type features (one-hot encoding of resource tags in the task queue and statistically calculating their proportion); and resource contention features (calculating the Pearson correlation coefficient of resource occupancy among multiple nodes). Power side feature vector extraction includes: power ramp features (calculating the mean absolute value of the difference in the load power sequence); harmonic content features (extracting the proportion of the 3rd, 5th, and 7th harmonics after FFT transformation of the voltage signal); and frequency offset features (calculating the deviation of the grid frequency from the 50Hz standard value). The mutual information entropy matrix calculation uses the Kraskov-Stögbauer-Grassberger algorithm to estimate the mutual information value between computing power features and power features; for example, the mutual information between load fluctuation and power ramp is 0.82 bits. A preset coupling threshold of 0.5 bits was set to select feature pairs with mutual information greater than 0.5 to form a key feature pair set, such as {(load fluctuation, power ramp-up), (task type, harmonic content)}. The deep temporal prediction network adopts an LSTM encoder-decoder structure, with a 2-layer LSTM encoder with 128 hidden units and a 2-layer LSTM decoder with 64 hidden units. The input is the historical sequence of key feature pairs, and the output is the prediction of the computing power-electricity joint state at future time points. Training uses the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error. The training data consists of historical data from the past 30 days. After training, the network parameters are fixed and deployed as a coupling correlation mapping model.

[0030] The beneficial effects of the above technical solutions are: time alignment solves the problem of multi-source heterogeneous frequency data fusion, feature vector groups extract key state representations, mutual information entropy matrix quantifies nonlinear correlation strength, coupling threshold filters strongly correlated features to reduce dimensionality, encoder-decoder structure captures temporal dependencies, and model output can be directly used for prediction, significantly improving the modeling accuracy of computing power-electricity coupling relationship.

[0031] In some embodiments of this application, based on the coupled correlation mapping model, rolling predictions are made on the computing load demand of the target computing power cluster over multiple consecutive time windows in the future, generating an initial load prediction sequence group, including: Based on the coupled correlation mapping model, the current state vector of the target computing power cluster is determined, wherein the current state vector includes the computing power resource occupancy value and power load value at the last monitoring time. Based on the current state vector, a context-aware feature representation is generated using the encoder part of the coupled association mapping model; The context-aware feature representation is input into the decoder of the coupled association mapping model to obtain the predicted state increment for the next time window; The predicted state increment is superimposed on the current state vector to generate the predicted computing power load and the predicted power consumption for the first future time window; Using the prediction result of the first future time window as the new current state vector, the operations of generating the context-aware feature representation, the input decoder, and the superimposed state increment are iteratively executed to obtain the initial load prediction sequence group of multiple consecutive time windows in sequence.

[0032] In this embodiment, the current state vector is formed by concatenating the computing power resource utilization rate (e.g., CPU 65%, GPU 80%) and the power load value (e.g., 900kW) at the last moment of the monitoring period, with a dimension of 5. The encoder LSTM encodes the current state vector and the historical sequence of the previous 15 minutes into a 128-dimensional context-aware feature representation, which contains the implicit pattern of computing power-power coupling. The decoder LSTM uses the context-aware feature representation as the initial state and gradually decodes to generate the predicted state increment for the next 15 minutes (the first time window), for example, an increase of 0.2 PFLOPS in computing power load and an increase of 50kW in power consumption. The superposition operation uses vector addition. Using [1.7, 950] as the new current state vector, it is input into the encoder-decoder again to generate the prediction for the second time window (the next 15-30 minutes), and this process is iterated 96 times (24 hours) to form an initial load prediction sequence group. Each group contains 96 prediction result units, and each unit records the predicted computing power load and predicted power consumption for that window. During the iteration process, the encoder uses the output of the previous prediction as the historical input each time to achieve rolling updates.

[0033] The beneficial effects of the above technical solution are: the current state vector anchors the prediction starting point, the encoder extracts deep temporal features, the decoder generates future increments, the superposition operation realizes state recursion, the iterative mechanism realizes long-term multi-step prediction, the rolling update adapts to dynamic changes, the prediction sequence group covers the entire scheduling cycle, and provides a data foundation for subsequent evaluation and strategy generation.

[0034] In some embodiments of this application, for each prediction result unit in the initial load prediction sequence group, determining the computing power supply-demand balance, electricity cost sensitivity, and resource utilization efficiency corresponding to the prediction result unit includes: Based on the predicted computing power load and the rated computing power capacity of the target computing power cluster, the computing power capacity margin ratio is calculated, and the reciprocal of the computing power capacity margin ratio is used as the computing power supply and demand balance. Obtain the time-of-use electricity price parameters and carbon emission factor parameters of the power supply network in the time window corresponding to the prediction result unit; Based on the predicted power consumption, the time-of-use electricity price parameters, and the carbon emission factor parameters, the comprehensive power cost per unit computing power task is calculated, and the rate of change of the comprehensive power cost is used as the power cost sensitivity. Based on the predicted computing power load and the predicted power consumption, a computing power-to-power conversion efficiency index is calculated. The resource utilization efficiency is determined by the ratio of the computing power-to-power conversion efficiency index to the historical average conversion efficiency of the target computing power cluster.

[0035] In this embodiment, the predicted computing load is 1.7 PFLOPS, the rated computing capacity is 2.0 PFLOPS, the computing capacity margin ratio is (2.0-1.7) / 2.0=0.15, and the computing power supply-demand balance is 6.67. The larger the value, the tighter the supply and demand. The time-of-use electricity price parameter is obtained from the grid, for example, 0.8 yuan / kWh during peak hours, and the carbon emission factor parameter is 0.6 kgCO2 / kWh. The predicted power consumption is 950 kWh, and the comprehensive power cost is 950 × 0.8 + 950 × 0.6 × carbon tax coefficient (set to 0.1 yuan / kg) = 817 yuan. The cost change rate between adjacent time windows is calculated. For example, from 800 yuan to 817 yuan, the change rate is 2.125%, which is determined as the power cost sensitivity. The computing power-to-power conversion efficiency index is predicted computing load / predicted power consumption = 1.79e12FLOPS / kW. The historical average conversion efficiency is 1.5e12FLOPS / kW, and the resource utilization efficiency score is approximately 1.19 (1.79 / 1.5). A score greater than 1 indicates that the efficiency is better than the historical average. The three-factor evaluation index structure is [6.67, 2.125, 1.19], which respectively characterize the supply and demand tension, cost volatility, and energy efficiency level.

[0036] The beneficial effects of the above technical solutions are as follows: the reciprocal of capacity margin highlights the degree of supply and demand tension; time-of-use pricing and carbon emission factors quantify green costs; cost change rate captures price fluctuation sensitivity points; computing power-electricity conversion efficiency measures energy efficiency level; historical comparison realizes efficiency normalization; and the ternary structure evaluates and predicts results from multiple dimensions, providing a quantitative basis for confidence screening.

[0037] In some embodiments of this application, the initial load forecast sequence set is subjected to confidence level screening based on the triplet evaluation index structure to obtain a reliable load forecast sequence set, including: Extract the triplet evaluation index structure of each prediction result unit and construct an evaluation index vector; Calculate the cluster center vector of the evaluation index vector of all prediction result units, and use it as the benchmark evaluation reference point; For each prediction result unit, the Euclidean distance between the evaluation index vector and the benchmark evaluation reference point is calculated to obtain the quantified value of the prediction deviation. The quantified value of the prediction deviation is compared with a preset grading threshold range for confidence level grading and filtering.

[0038] In some embodiments of this application, the initial load forecast sequence set is screened based on confidence level according to the triplet evaluation index structure to obtain a reliable load forecast sequence set, and the method further includes: The preset tiered threshold range includes the upper limit of the first-level threshold, the upper limit of the second-level threshold, and the upper limit of the third-level threshold; In response to the prediction deviation quantification value being less than the upper limit of the first-level threshold, the corresponding prediction result unit is marked as a high-confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the first-level threshold and the upper limit of the second-level threshold, the corresponding prediction result unit is marked as a medium confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the second threshold and the upper limit of the third threshold, the corresponding prediction result unit is marked as a low confidence prediction result; If the quantified value of the prediction deviation is greater than the upper limit of the third-level threshold, the corresponding prediction result unit is removed. The prediction result units labeled as high-confidence prediction results and medium-confidence prediction results are integrated into the set of credible load prediction sequences.

[0039] In this embodiment, the evaluation index vector is the triplet evaluation index structure [computing power supply and demand balance, electricity cost sensitivity, and resource utilization efficiency], for example, a certain unit is [6.67, 2.125, 1.19]. K-Means clustering is performed on the evaluation index vectors of all 96 prediction result units, with cluster center K=3. The center of the largest cluster is taken as the benchmark evaluation reference point, for example, [5.5, 1.8, 1.05]. Euclidean distance is calculated. Preset tiered threshold ranges are set: upper limit 0.8 for the first level, 1.5 for the second level, and 2.5 for the third level. The unit with a distance of 1.23 is between 0.8 and 1.5, and is marked as having medium confidence. Distances less than 0.8 are marked as high confidence, 0.8-1.5 as medium confidence, 1.5-2.5 as low confidence, and values ​​greater than 2.5 are discarded. After screening, 60 high-confidence sequences and 30 medium-confidence sequences were integrated into a set of 90 reliable load prediction sequences, while 6 low-confidence sequences were removed or downgraded.

[0040] The beneficial effects of the above technical solution are: standardized evaluation index vectors provide multi-dimensional evaluation; cluster centers determine benchmark reference points; Euclidean distance quantifies prediction bias; three-level thresholds achieve confidence level classification; high / medium confidence results are retained to ensure reliability; low confidence results are eliminated to avoid misleading decisions; and the credible load prediction sequence set provides high-quality input for subsequent strategy generation.

[0041] In some embodiments of this application, the trusted load prediction sequence set is input to a dynamic feedback adjustment framework to generate a computing power scheduling recommendation strategy for the target computing power cluster and a power supply adjustment recommendation strategy for the power supply network, including: Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the predicted computing power load and the current actual load of the target computing power cluster is analyzed to generate a trend vector of computing power load demand change. Based on the computing power load demand change trend vector, the corresponding task migration scheme, node start / stop scheme and load balancing scheme are matched in the preset scheduling strategy library, and the computing power scheduling suggestion strategy is generated by combining them. Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the predicted power consumption and the current actual power supply of the power supply network is analyzed to generate a power supply and demand gap vector. Based on the power supply and demand gap vector and the time-of-use electricity price parameters, the corresponding generator output adjustment scheme, energy storage charging and discharging scheme and flexible load adjustment scheme are matched in the preset power supply strategy library, and the power supply adjustment suggestion strategy is generated by combining them. The computing power scheduling suggestion strategy and the power supply adjustment suggestion strategy are aligned on the time axis to generate a collaborative execution plan sequence, which contains a joint scheduling instruction set for each future time window.

[0042] In this embodiment, the high-confidence prediction result shows that the computing load in the next 4 hours will increase from the current 1.5 PFLOPS to 1.8 PFLOPS, a difference of 0.3 PFLOPS, with a trend vector of [+0.3, 0, 0] (indicating growth). Matching from the preset scheduling strategy library: the task migration scheme migrates low-priority tasks from the soon-to-be-fully-loaded node A to the idle node B; the node start / stop scheme starts the backup server node C; and the load balancing scheme adjusts the task scheduling algorithm weights. These are combined to form a computing load scheduling recommendation strategy. Simultaneously, the predicted power consumption increases from 900 kWh to 1050 kWh, while the current power supply is 1000 kWh, resulting in a gap vector of [-50] (indicating insufficient power supply). Matching from the preset power supply strategy library: the generator output adjustment scheme starts the diesel generator to increase output by 150 kW; the energy storage charging / discharging scheme discharges the lithium battery energy storage system by 50 kW; and the flexible load adjustment scheme postpones non-critical lighting loads. These are combined to form a power supply adjustment recommendation strategy. The timeline is aligned at 15-minute intervals to generate a sequence of collaborative execution plans, such as a joint instruction set for a T+4h window: {Computing power: Startup node C, Power: Energy storage discharge 50kW}, ensuring that computing power expansion and power supply increase are executed synchronously.

[0043] The beneficial effects of the above technical solutions are: load difference analysis quantifies changing demand, trend vectors represent the direction of change, strategy library matching enables rapid decision-making, task migration and node start-up and shutdown adjust computing power supply, energy storage discharge and unit regulation ensure power balance, time axis alignment ensures computing power-power strategy synergy, joint scheduling instruction set realizes closed-loop control, and improves overall energy utilization efficiency and computing power service quality.

[0044] In some embodiments of this application, it also includes: The unified monitoring panel of the target computing power cluster displays an associated view of the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy.

[0045] In some embodiments of this application, a related view is shown of the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy, including: In response to receiving a command to display prediction details for the unified monitoring panel, a sub-window for displaying prediction details pops up in the unified monitoring panel; In the prediction details display sub-window, the generation basis of each prediction result unit in the trusted load prediction sequence set is displayed in layers. The generation basis includes: the original data source of the multi-dimensional computing power operation parameter set, the ranking of the key feature contribution of the coupled correlation mapping model, and the detailed scores of each dimension of the triplet evaluation index structure. In response to receiving a selection instruction for a specific feature in the ranking of the contribution of the key features, the time-series evolution curve of the specific feature within the preset monitoring period is expanded in the prediction details display sub-window, and the influence weight value of the specific feature on the current prediction result is marked.

[0046] In this embodiment, the unified monitoring panel is a Vue.js-based web interface. When an operations and maintenance personnel click on a prediction curve node, a details display command is triggered, and a modal dialog box pops up as a sub-window displaying the prediction details. The layered display uses a three-column layout: the left side displays the original data source, listing the original CPU / GPU utilization values ​​of each node that the prediction unit depends on within the past 15 minutes; the middle displays the contribution ranking of key features, such as "Power Ramp Feature" contributing 32% and "Load Fluctuation Feature" contributing 28%, arranged in descending order; the right side displays the scores of each dimension of the triplet evaluation index, such as balance 6.67 (red warning), cost sensitivity 2.125 (yellow reminder), and efficiency 1.19 (green normal). After clicking on "Power Ramp Feature," the right side dynamically loads the time-series curve of this feature over the past 15 minutes, with time on the horizontal axis and ramp amplitude on the vertical axis. The current prediction point is marked in red, and the influence weight value of 0.32 is displayed in a floating position.

[0047] The beneficial effects of the above technical solution are: the prediction details display sub-window provides drill-down analysis capabilities, the hierarchical display clearly presents the prediction generation chain, the original data traceability ensures auditability, the feature contribution ranking explains the model decision basis, the dimension score details display evaluation details, the time series evolution curve provides historical context, the influence weight value quantifies the importance of features, improves system transparency and credibility, and supports human-machine collaborative decision-making.

[0048] In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0049] Although the invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the embodiments disclosed in this invention can be combined with each other in any way. The fact that not all of these combinations are described in this specification is merely for the sake of brevity and resource conservation.

[0050] It will be understood by those skilled in the art that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load dynamic prediction and evaluation method with deep coupling of computing power and electricity, characterized in that, include: Obtain a multi-dimensional set of computing power operation parameters of the target computing power cluster within a preset monitoring period, wherein the multi-dimensional set of computing power operation parameters includes the computing power resource occupancy sequence of each computing node, the task queue length sequence of each computing node, and the instruction throughput sequence of each computing node; Obtain a set of power status parameters of the power supply network that matches the target computing power cluster, wherein the set of power status parameters includes the load power sequence of the power supply node, the voltage stability sequence of the power supply node, and the power quality sequence of the power supply node. Based on the multi-dimensional computing power operation parameter set and the power state parameter set, a computing power-power coupling correlation mapping model is constructed. Based on the coupled correlation mapping model, the computing load demand of the target computing power cluster in multiple consecutive future time windows is predicted in a rolling manner, and an initial load prediction sequence group is generated. The initial load prediction sequence group includes the predicted computing load and predicted power consumption for each future time window. For each prediction result unit in the initial load prediction sequence group, determine the computing power supply and demand balance, power cost sensitivity, and resource utilization efficiency corresponding to the prediction result unit, and generate a triplet evaluation index structure. Based on the triplet evaluation index structure, the initial load prediction sequence group is screened by confidence level to obtain a reliable load prediction sequence set. The reliable load prediction sequence set is then input into the dynamic feedback adjustment framework to generate a computing power scheduling suggestion strategy for the target computing power cluster and a power supply adjustment suggestion strategy for the power supply network.

2. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, Obtain a set of multi-dimensional computing power operation parameters for the target computing power cluster within a preset monitoring period, including: By deploying performance acquisition agents on each computing node in the target computing power cluster, and using a unified timestamp as a reference, time-series data of CPU utilization, GPU utilization, and memory bandwidth utilization are collected for each computing node. The timing data of the CPU utilization rate, the timing data of the GPU utilization rate, and the timing data of the memory bandwidth utilization rate are normalized to generate the computing resource utilization sequence. Record the task submission timestamp, task execution duration, and task resource requirement tag of each computing node within the preset monitoring period, and construct the task queue length sequence; The number of integer arithmetic instructions, floating-point arithmetic instructions, and vector arithmetic instructions completed by each computing node within the preset monitoring period is counted to generate the instruction throughput sequence.

3. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, Based on the multi-dimensional computing power operation parameter set and the power state parameter set, a computing power-power coupling correlation mapping model is constructed, including: Time alignment is performed on the multi-dimensional computing power operation parameter set and the power status parameter set to obtain a synchronized computing power-power joint parameter matrix. Based on the computing power-power joint parameter matrix, a computing power-side feature vector group and a power-side feature vector group are extracted. The computing power-side feature vector group includes load fluctuation characteristics, task type characteristics, and resource contention characteristics. The power-side feature vector group includes power ramp-up characteristics, harmonic content characteristics, and frequency offset characteristics. Determine the mutual information entropy matrix between the computing power-side feature vector group and the power-side feature vector group. The mutual information entropy matrix is ​​used to quantify the synchronous correlation strength between computing power state changes and power state changes. Based on the mutual information entropy matrix, a set of key feature pairs with a correlation strength greater than a preset coupling threshold is selected; A deep temporal prediction network is trained using the key feature pairs. The deep temporal prediction network includes an encoder-decoder structure, wherein the encoder is used to extract the implicit coupling patterns of the key feature pairs, and the decoder is used to reconstruct the computing power-electricity joint state at future time moments. The trained deep temporal prediction network is identified as the computing power-electricity coupling correlation mapping model.

4. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, Based on the aforementioned coupled correlation mapping model, the computing load demand of the target computing power cluster is predicted in a rolling manner over multiple consecutive time windows in the future, generating an initial load prediction sequence set, including: Based on the coupled correlation mapping model, the current state vector of the target computing power cluster is determined, wherein the current state vector includes the computing power resource occupancy value and power load value at the last monitoring time. Based on the current state vector, a context-aware feature representation is generated using the encoder part of the coupled association mapping model; The context-aware feature representation is input into the decoder of the coupled association mapping model to obtain the predicted state increment for the next time window; The predicted state increment is superimposed on the current state vector to generate the predicted computing power load and the predicted power consumption for the first future time window; Using the prediction result of the first future time window as the new current state vector, the operations of generating the context-aware feature representation, the input decoder, and the superimposed state increment are iteratively executed to obtain the initial load prediction sequence group of multiple consecutive time windows in sequence.

5. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, For each prediction result unit in the initial load prediction sequence group, determine the computing power supply-demand balance, electricity cost sensitivity, and resource utilization efficiency corresponding to the prediction result unit, including: Based on the predicted computing power load and the rated computing power capacity of the target computing power cluster, the computing power capacity margin ratio is calculated, and the reciprocal of the computing power capacity margin ratio is used as the computing power supply and demand balance. Obtain the time-of-use electricity price parameters and carbon emission factor parameters of the power supply network in the time window corresponding to the prediction result unit; Based on the predicted power consumption, the time-of-use electricity price parameters, and the carbon emission factor parameters, the comprehensive power cost per unit computing power task is calculated, and the rate of change of the comprehensive power cost is used as the power cost sensitivity. Based on the predicted computing power load and the predicted power consumption, a computing power-to-power conversion efficiency index is calculated. The resource utilization efficiency is determined by the ratio of the computing power-to-power conversion efficiency index to the historical average conversion efficiency of the target computing power cluster.

6. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, Based on the aforementioned triplet evaluation index structure, the initial load forecast sequence set is subjected to confidence level grading and screening to obtain a reliable load forecast sequence set, including: Extract the triplet evaluation index structure of each prediction result unit and construct an evaluation index vector; Calculate the cluster center vector of the evaluation index vector of all prediction result units, and use it as the benchmark evaluation reference point; For each prediction result unit, the Euclidean distance between the evaluation index vector and the benchmark evaluation reference point is calculated to obtain the quantified value of the prediction deviation. The quantified value of the prediction deviation is compared with a preset grading threshold range for confidence level grading and filtering.

7. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 6, characterized in that, Based on the triplet evaluation index structure, the initial load prediction sequence set is screened according to confidence level to obtain a reliable load prediction sequence set, which also includes: The preset tiered threshold range includes the upper limit of the first-level threshold, the upper limit of the second-level threshold, and the upper limit of the third-level threshold; In response to the prediction deviation quantification value being less than the upper limit of the first-level threshold, the corresponding prediction result unit is marked as a high-confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the first-level threshold and the upper limit of the second-level threshold, the corresponding prediction result unit is marked as a medium confidence prediction result; In response to the prediction deviation quantification value being between the upper limit of the second threshold and the upper limit of the third threshold, the corresponding prediction result unit is marked as a low confidence prediction result; If the quantified value of the prediction deviation is greater than the upper limit of the third-level threshold, the corresponding prediction result unit is removed. The prediction result units labeled as high-confidence prediction results and medium-confidence prediction results are integrated into the set of credible load prediction sequences.

8. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, The reliable load prediction sequence set is input into the dynamic feedback adjustment framework to generate a computing power scheduling recommendation strategy for the target computing power cluster and a power supply adjustment recommendation strategy for the power supply network, including: Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the predicted computing power load and the current actual load of the target computing power cluster is analyzed to generate a vector of computing power load demand change. Based on the computing power load demand change trend vector, the corresponding task migration scheme, node start / stop scheme and load balancing scheme are matched in the preset scheduling strategy library, and the computing power scheduling suggestion strategy is generated by combining them. Based on each high-confidence prediction result in the set of reliable load prediction sequences, the difference between the corresponding predicted power consumption and the current actual power supply of the power supply network is analyzed to generate a power supply and demand gap vector. Based on the power supply and demand gap vector and the time-of-use electricity price parameters, the corresponding generator output adjustment scheme, energy storage charging and discharging scheme and flexible load adjustment scheme are matched in the preset power supply strategy library, and the power supply adjustment suggestion strategy is generated by combining them. The computing power scheduling suggestion strategy and the power supply adjustment suggestion strategy are aligned on the time axis to generate a collaborative execution plan sequence, which contains a joint scheduling instruction set for each future time window.

9. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 1, characterized in that, Also includes: The unified monitoring panel of the target computing power cluster displays an associated view of the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy.

10. The load dynamic prediction and evaluation method based on deep coupling of computing power and electricity as described in claim 9, characterized in that, A view showing the association between the trusted load prediction sequence set, the triplet evaluation index structure, the computing power scheduling recommendation strategy, and the power supply adjustment recommendation strategy includes: In response to receiving a command to display prediction details for the unified monitoring panel, a sub-window for displaying prediction details pops up in the unified monitoring panel; In the prediction details display sub-window, the generation basis of each prediction result unit in the trusted load prediction sequence set is displayed in layers. The generation basis includes: the original data source of the multi-dimensional computing power operation parameter set, the ranking of the key feature contribution of the coupled correlation mapping model, and the detailed scores of each dimension of the triplet evaluation index structure. In response to receiving a selection instruction for a specific feature in the ranking of the contribution of the key features, the time-series evolution curve of the specific feature within the preset monitoring period is expanded in the prediction details display sub-window, and the influence weight value of the specific feature on the current prediction result is marked.

Images