Energy consumption multi-dimensional analysis and load prediction method and system of computing power center

Abstract

The invention provides a computing power center energy consumption multi-dimensional analysis and load prediction method and system. A system collects multi-source heterogeneous data in real time, and after preprocessing, the contribution degree of each feature to energy consumption is deeply analyzed by using a Transform-based pre-trained large model, and a key driving factor is identified. And carrying out high-precision minute-level energy consumption rolling prediction by adopting a hybrid model combining LSTM and Transform, and introducing error feedback dynamic correction. And constructing an electricity-calculation cooperative scheduling agent based on deep reinforcement learning, and outputting a preliminary scheduling instruction. And establishing an optimization model taking the power matching degree, the energy consumption cost and the scheduling cost as targets, and solving a Pareto optimal scheme by adopting multi-target evolutionary algorithms such as NSGA-II and the like. The system analyzes an optimization scheme into a specific control instruction, issues and executes the specific control instruction through a safety interface, monitors the effect in real time, feeds back deviation data to a prediction and decision model, and achieves self-adaptive optimization and continuous evolution, thereby improving energy efficiency, reducing cost and guaranteeing power grid stability.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of smart energy and computing infrastructure, specifically involving a method and system for multi-dimensional analysis of energy consumption and load prediction of computing centers. Background Technology

[0002] With the rapid development of the digital economy and artificial intelligence technologies, computing centers, as core digital infrastructure, are experiencing explosive growth in scale and energy consumption. Statistics show that large computing centers consume hundreds of millions of kilowatt-hours of electricity annually, and their energy costs account for more than half of their total operating costs. Reducing energy consumption and improving energy efficiency have become key challenges for the industry's sustainable development. Currently, the energy management and power dispatching models of computing centers are still relatively rudimentary, struggling to adapt to the heterogeneous internal equipment, transient loads, and complex dynamic characteristics of their interaction with the power grid. This results in low energy utilization efficiency and generates significant economic and environmental costs.

[0003] Most existing energy consumption prediction technologies are based on single time-series analysis of historical power consumption data, such as using autoregressive integral moving average models. These methods treat computing centers as "black boxes," relying solely on historical power consumption values ​​for extrapolation, completely ignoring key intrinsic driving factors such as drastic fluctuations in server CPU and GPU utilization, memory and storage I / O activity, and dynamic cooling loads on data center precision air conditioning systems due to changes in ambient temperature and humidity. Due to a lack of perception and modeling capabilities for multi-dimensional influencing factors, traditional prediction models perform poorly when faced with sudden computing tasks or environmental disturbances, with prediction errors generally exceeding 15%, severely limiting the feasibility and reliability of proactive scheduling based on prediction results.

[0004] At the power supply and dispatch level, a static model based on long-term contracts or fixed quotas is still widely used. The power grid allocates power quotas to computing centers based on historical experience or rough estimates, while the operators of these centers passively manage their load under this rigid constraint, unable to flexibly adjust according to real-time electricity price signals and internal computing power needs. This leads to a double resource mismatch: during peak power consumption periods and when electricity prices are high, computing centers may be forced to reduce server frequency and load due to power quotas, resulting in the loss of valuable computing resources and affecting the completion of critical tasks; conversely, during off-peak power consumption periods and when electricity prices are low, the lack of proactive incentives and dispatching mechanisms leads to idle computing resources and wasted cheap electricity. Therefore, the industry urgently needs an intelligent technology solution that can deeply integrate multi-dimensional real-time data, achieve high-precision energy consumption perception and prediction, and support minute-level dynamic collaborative dispatching of power and computing resources, in order to fundamentally improve the energy efficiency management level and overall resource utilization rate of computing centers. Summary of the Invention

[0005] To overcome the problems of insufficient prediction model accuracy and low energy efficiency and high operating costs caused by static matching of computing resources in the background technology, this invention provides a method and system for multi-dimensional analysis of energy consumption and load prediction in computing centers. By real-time fusion of multi-dimensional data on energy consumption, equipment operation, time period and environment, and by using a pre-trained large model to analyze energy efficiency driving factors, a high-precision time series prediction model and a dynamic collaborative intelligent agent for computing based on reinforcement learning are constructed. Combined with a multi-objective optimization algorithm to generate optimal bidirectional scheduling instructions, minute-level high-precision energy consumption prediction and dynamic optimal matching of power supply and computing demand are achieved, thereby significantly improving the level of intelligent energy efficiency management and the efficiency of comprehensive resource utilization.

[0006] The specific technical solution of this application is as follows:

[0007] According to one aspect of this application, a method for multi-dimensional analysis of energy consumption and load forecasting of computing centers is provided, including:

[0008] Real-time collection of diverse and heterogeneous data within the computing center, followed by preprocessing to form a standardized sequence of feature vectors;

[0009] The standardized feature vector sequence is input into the pre-trained large-scale energy efficiency analysis model. The self-attention mechanism of the large-scale energy efficiency analysis model is used to analyze the contribution of different data features to the energy consumption and identify the key driving factors affecting energy consumption fluctuations.

[0010] Based on the historical sequence after feature parsing, a hybrid time series model combining a long short-term memory network and a Transformer architecture is used to perform rolling predictions of energy consumption in the next minute. The error between the real-time measurement value and the prediction value is dynamically corrected to generate a high-precision energy consumption prediction sequence for the future period.

[0011] Construct a decision-making intelligent agent with dynamic association between electricity and computing. The decision-making intelligent agent takes as input state information including real-time grid status, computing load status, environmental status and future energy consumption prediction sequence. Based on a deep reinforcement learning model, it outputs preliminary collaborative scheduling instructions for computing utilization adjustment and grid power request.

[0012] Based on the initial cooperative scheduling instructions, a multi-objective optimization model is constructed with the optimization objectives of maximizing the power-computer power matching degree, minimizing the total energy consumption cost, and minimizing the comprehensive scheduling cost. The model is then solved using a multi-objective evolutionary algorithm, and the final scheduling scheme is selected from the set of optimal solutions.

[0013] The final scheduling scheme is parsed into specific computing cluster control commands and power grid adjustment request commands, and then sent to the computing resource management platform and the power grid dispatching system for execution through secure interfaces. At the same time, the actual effect after the commands are executed is monitored, and the execution deviation data is fed back to the prediction model and decision-making agent for optimization and adaptive adjustment.

[0014] As a further option of the method of the present invention, the pre-trained energy efficiency analysis big model dynamically identifies and quantifies the influence pattern of key driving factors on energy consumption by parsing the weight matrix of its self-attention mechanism.

[0015] The analysis includes averaging and normalizing the attention weights along the time and feature dimensions to generate a structured report containing the relationship between feature contribution weights and nonlinear effects.

[0016] As a further option of the method of the present invention, the hybrid timing model adopts a cascaded architecture:

[0017] The preceding stage consists of stacked long short-term memory networks, used to extract temporal-dependent features from long historical sequences;

[0018] The latter stage is a Transformer decoder, which takes the feature sequence output from the former stage as input and performs multi-step prediction through a masked self-attention mechanism; the model output layer is directly mapped to the energy consumption prediction values ​​for multiple consecutive minutes in the future.

[0019] As a further option of the method of the present invention, the dynamic correction is achieved through an error feedback mechanism:

[0020] The instantaneous prediction error at the latest moment is added to the prediction sequence for future periods in an exponentially decaying manner. The correction formula is as follows: ,in, This is the corrected predicted value. These are the original predicted values. The attenuation coefficient is... For instantaneous error, To correct the strength coefficient.

[0021] As a further option of the method of the present invention, the decision agent is based on a deep Q-network architecture and uses a Dueling DQN structure to decompose the Q value into a state value function and an advantage function.

[0022] The exploration strategy employs an ε-greedy algorithm, and the exploration rate decreases linearly with the number of training steps.

[0023] Among them, the intelligent agent has a safety barrier when making online decisions. When the confidence of the output decision is too low or violates the preset safety rules, it automatically switches to the rule-based backup strategy.

[0024] As a further option of the method of the present invention, in the multi-objective optimization model, the computing power demand power consumption is modeled as a function of computing power utilization rate and ambient temperature, and the specific parameters in the function form are determined by the feature contribution weights and relationships parsed from the large energy efficiency analysis model.

[0025] As a further option of the method of the present invention, a non-dominated sorting genetic algorithm with an elitist strategy is used to solve the multi-objective optimization model;

[0026] The algorithm encodes candidate solutions with real numbers, performs population evolution by simulating binary crossover and polynomial mutation, and filters the Pareto optimal solution set based on non-dominated sorting and crowding calculation. Finally, it selects the final scheduling scheme from the solution set through a human-computer interactive visual interface.

[0027] As a further option of the method of the present invention, the step of parsing the final scheduling scheme into specific control instructions includes: generating a globally unique instruction identifier and timestamp for each instruction and performing logical verification in advance; the instructions are sent to the computing resource management platform and the power grid scheduling system respectively through a secure communication link with two-way authentication and end-to-end encryption, using a request-confirmation mechanism.

[0028] As a further option of the method of the present invention, the actual effect of the monitoring command after execution and the feedback optimization specifically include: establishing a historical case library to store complete scheduling experience trajectories, and periodically starting background learning tasks to update the policy network parameters of the decision-making agent offline using new data, fine-tuning the hybrid time series prediction model, and adjusting the weight preferences of the multi-objective optimization model to achieve the self-evolution of the system.

[0029] This application also provides a multi-dimensional analysis and load forecasting system for computing center energy consumption, the system including:

[0030] The data acquisition and preprocessing module is used to acquire diverse and heterogeneous data in the computing center in real time and preprocess it to form a standardized feature vector sequence.

[0031] The energy efficiency feature analysis module is equipped with a pre-trained large-scale energy efficiency analysis model, which is used to receive standardized feature vector sequences and analyze the contribution of different data features to energy consumption through the self-attention mechanism of the large-scale model, and identify the key driving factors affecting energy consumption fluctuations.

[0032] The energy consumption prediction module is equipped with a hybrid time series model that combines a long short-term memory network and a Transformer architecture. It is used to make rolling predictions of future energy consumption in minutes based on historical sequences after feature parsing, and to dynamically correct the error between real-time measurement values ​​and prediction values ​​to generate a high-precision energy consumption prediction sequence for future periods.

[0033] The collaborative scheduling agent module is a power-computer dynamic correlation decision-making agent built based on a deep reinforcement learning model. It is used to take state information including real-time power grid status, computing load status, environmental status and future energy consumption prediction sequence as input, and output preliminary collaborative scheduling instructions for computing utilization adjustment and power grid power request.

[0034] The multi-objective optimization module is used to construct a multi-objective optimization model based on the initial cooperative scheduling instructions, with the optimization objectives of maximizing the power-computer power matching degree, minimizing the total energy consumption cost, and minimizing the comprehensive scheduling cost. The model is solved using a multi-objective evolutionary algorithm, and the final scheduling scheme is selected from the set of optimal solutions.

[0035] The instruction execution and feedback module is used to parse the final scheduling scheme into specific computing cluster control instructions and power grid adjustment request instructions, and send them to the computing resource management platform and the power grid dispatching system for execution through a secure interface. At the same time, it monitors the actual effect after instruction execution and feeds back the execution deviation data to the energy consumption prediction module and the collaborative scheduling agent module for model optimization and adaptive adjustment.

[0036] The beneficial effects of this application are as follows:

[0037] By integrating multidimensional heterogeneous data and utilizing large models to deeply analyze energy efficiency characteristics, the minute-level energy consumption prediction error has been reduced from 15%~20% in traditional methods to below 5%, laying a reliable data foundation for refined scheduling.

[0038] It breaks the deadlock of static matching of electricity and computing resources. Through minute-level dynamic collaboration, it significantly improves the utilization rate of computing resources during off-peak hours and intelligently avoids peak hours. At the same time, it increases the time-series matching degree of power supply and computing demand to over 90%, significantly reducing energy waste and PUE value.

[0039] By tracking time-of-use electricity prices and optimizing scheduling strategies, peak electricity costs were effectively reduced, and computing power losses and task delay default costs caused by power constraints were minimized, resulting in significant direct economic benefits.

[0040] It has built a complete system from perception, prediction, decision-making, optimization to execution, monitoring, and feedback. The system can continuously iterate and optimize internal models and strategies based on operational data, adapt to changes in business and environment, and has long-term vitality.

[0041] The method fully considers secure integration with existing computing power management platforms and power grid dispatching systems, and sets up multi-layered security barriers and degradation strategies to ensure high availability and robustness of the technology in complex industrial environments. Attached Figure Description

[0042] Figure 1A schematic diagram of the overall process for multi-dimensional analysis of energy consumption and load forecasting methods in computing centers;

[0043] Figure 2 A detailed flowchart of the S100 method for multi-dimensional analysis of energy consumption and load forecasting in computing centers;

[0044] Figure 3 A detailed flowchart of the S200 method for multi-dimensional analysis of energy consumption and load forecasting in computing centers;

[0045] Figure 4 A detailed flowchart of the S300 method for multi-dimensional analysis of energy consumption and load forecasting in computing centers;

[0046] Figure 5 A detailed flowchart of the S400 method for multi-dimensional analysis of energy consumption and load forecasting in computing centers;

[0047] Figure 6 A detailed flowchart of the S500 method for multi-dimensional analysis of energy consumption and load forecasting in computing centers;

[0048] Figure 7 A detailed flowchart of the steps for the S600 method for multi-dimensional analysis of energy consumption and load prediction in computing centers. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0050] Computing centers contain numerous servers, cooling and power supply equipment, with complex energy interactions between systems. Traditional methods use single historical energy consumption data for prediction, ignoring the influence of multiple dimensions such as equipment load, ambient temperature, and grid electricity prices, leading to large prediction errors. Meanwhile, power dispatch uses fixed quotas, which cannot adapt to minute-level fluctuations in computing power demand, resulting in resource waste and increased costs. Existing technologies lack the ability to deeply integrate and analyze diverse data, and have not established a real-time, two-way collaborative decision-making mechanism.

[0051] Example 1

[0052] Please see Figure 1 The multi-dimensional analysis and load prediction method for computing center energy consumption provided in this embodiment of the invention includes the following steps:

[0053] S100: Real-time acquisition and preprocessing of diverse heterogeneous data.

[0054] S200: Deep analysis of energy efficiency features based on pre-trained large models.

[0055] S300: High-precision rolling energy consumption prediction based on a hybrid time series model.

[0056] S400: Constructing an intelligent agent for dynamic load forecasting based on electricity and computing.

[0057] S500: Power-computing power bidirectional matching based on multi-objective optimization.

[0058] S600: Generates and executes minute-level collaborative scheduling instructions.

[0059] The specific plan is as follows:

[0060] like Figure 2 As shown, the S100 constructs a high-frequency, multi-dimensional data acquisition and preprocessing pipeline to provide clean and well-organized input data for subsequent analysis.

[0061] S110: Synchronizes four types of streaming data from the data platform through standardized interfaces.

[0062] In this invention, the system acquires real-time data streams of energy consumption, equipment operation, time period tags, and environmental parameters in 1-10 minute intervals via agent programs deployed in the production environment or by directly calling the data platform API. Energy consumption data originates from smart meters deployed in each server rack or row-header, and includes timestamps, device identifiers, and minute-level active power consumption values ​​in kWh. Equipment operation data is collected through the server's out-of-band management interface, with collected indicators including at least CPU utilization percentage, memory utilization percentage, and hard disk temperature in degrees Celsius. The collection frequency is once per minute, and the reported value is the average value within that minute. Time period tag data is integrated from publicly available time-of-use pricing strategy documents released by the power grid company or from the internal energy management system. The system divides a 24-hour day into peak, flat, and valley periods, and assigns a corresponding tag and real-time electricity price to each time period. Environmental parameter data comes from a network of temperature and humidity sensors deployed in various functional areas of the computer room, with the sensors reporting data at a frequency of at least once per minute.

[0063] In one possible implementation, data transmission is achieved using the Apache Kafka message queue. The data producer publishes data to a designated topic, and the system of this invention acts as the data consumer, subscribing to the topic to achieve real-time streaming ingestion. To ensure data order and integrity, the system appends a monotonically increasing sequence number to each message and verifies it at the consumer end.

[0064] S120: Perform data quality control and missing value imputation.

[0065] In this invention, the system performs a series of validity checks on each incoming raw data record. The check rules include: numerical range check, mutation detection, calculating the absolute difference between the current value and the previous value, and marking it as an outlier if the difference exceeds three times the historical standard deviation; and time continuity check, checking whether the data timestamps are continuous and whether there are large-scale jumps or repetitions.

[0066] In one possible implementation, the system employs a context-based imputation strategy for data marked as missing or invalid. For short-term missing data, linear interpolation is used. For long-term data loss due to equipment failure, the average data from other servers of the same model and load level within the same cluster during the same period is used for imputation. All imputation operations are logged in the audit log.

[0067] S130: Normalize and encode features of heterogeneous data.

[0068] In this invention, to ensure that features of different dimensions and magnitudes can be treated equally by the model, all continuous numerical features are Z-score standardized. Categorical features, such as time period labels, are one-hot encoded. The mean and standard deviation parameters required for standardization are calculated and updated every 24 hours based on historical data from the past 7 days to adapt to the slow changes in data distribution. Finally, the processed four types of features are concatenated into a fixed-dimensional feature vector. It serves as the unified input for downstream models.

[0069] In one possible implementation, the system generates an additional set of statistical features, which are incorporated into the feature vector as derived features to provide richer temporal context information.

[0070] like Figure 3 As shown, the S200 utilizes a pre-trained large-scale pre-trained model to automatically mine key driving factors affecting energy consumption fluctuations and their quantitative contributions from the fused multidimensional data.

[0071] S210: Load and initialize the Guangming Power large model.

[0072] In this invention, the Guangming Power large-scale model is an encoder model based on the Transformer architecture, which has been pre-trained on massive amounts of historical data from power systems and data centers. The model file is stored in a persistent format and loaded into memory or GPU memory upon system startup. The model configuration includes 12 encoding layers, each with a hidden dimension of 768, and 12 attention heads. After loading, the model enters inference mode, and the parameters are frozen.

[0073] S220: Input the preprocessed time series data into the model for forward computation.

[0074] In this invention, the feature sequence output by S130, which includes a window of time in the past, is... The input sequence is transformed into a high-dimensional vector representation by the embedding layer of the model. Subsequently, the sequence undergoes processing through all 12 Transformer encoding layers. Each layer uses a self-attention mechanism to capture the complex dependencies between different time steps and different feature dimensions within the sequence. Finally, the model outputs the hidden state sequence of the last layer. and the weight matrices of each attention layer .

[0075] In one possible implementation, to simultaneously capture long-term trends and short-term fluctuations, the system employs a dual-input architecture. One input is downsampled data from the past 24 hours, used to extract daily-level cyclical patterns; the other input is raw-resolution data from the past hour, used to extract minute-level transient patterns. The two inputs are processed through independent embedding layers, concatenated along the feature dimension, and then input into a shared Transformer encoder.

[0076] S230: Analyze attention weights and generate a feature contribution report.

[0077] In this invention, the attention weight matrix output by the model It contains information about the relative importance of different features to the final representation. The system... Analysis: First, average the weights along the time dimension to obtain the weight distribution across the feature dimensions for each attention head; then, average the weights of all attention heads to obtain the global attention score for each input feature. Finally, the scores are normalized to obtain the contribution weights of each feature. .

[0078] In one possible implementation, the system automatically identifies features with contribution weights exceeding a preset threshold, marks them as key driving factors, and outputs them in the form of a structured report. The report not only lists the weights but also attempts to explain the impact patterns. For example, when GPU utilization is between 70% and 85%, its contribution weight to the change in total energy consumption reaches 0.45 and exhibits a non-linear positive correlation; after the data center temperature exceeds 25 degrees Celsius, for every 1 degree Celsius increase in temperature, its contribution weight to the energy consumption of the cooling subsystem increases by approximately 0.08. Simultaneously, the system generates a visual contribution heatmap, intuitively displaying the weight changes of key factors under different time periods and loads.

[0079] like Figure 4 As shown, the S300 constructs and uses a hybrid model that combines a long short-term memory network and a Transformer to make high-precision, rolling predictions of future energy consumption in minutes.

[0080] S310: Build and train the LSTM-Transformer hybrid prediction model.

[0081] In this invention, the prediction model consists of two parts. The first part is a two-layer stacked LSTM network with 128 units, responsible for extracting long-term trends, periodicity, and dependencies from a 24-hour historical sequence. The LSTM outputs the hidden state at each time step, forming a new sequence. The second part is a four-layer Transformer decoder that takes the sequence output from the first part as input and utilizes its masked self-attention mechanism and fully connected feedforward network to capture fine correlations within the sequence and perform multi-step predictions. The model concludes with a wired output layer that maps the decoder's final hidden state to the predicted power consumption value for each minute of the next 60 minutes.

[0082] In one possible implementation, model training uses historical normal operating data. The loss function is defined as the weighted sum of the mean squared error and the mean absolute error between the predicted and actual values. The optimizer uses Adam with an initial learning rate of 1e-4, and a cosine annealing scheduling strategy is employed. Early stopping is used during training; training is halted when the validation set loss no longer decreases for 10 consecutive epochs.

[0083] S320: Performs online rolling forecasting and dynamic error correction.

[0084] In this invention, in a production environment, the system initiates a prediction process every minute. The latest historical sequence, enhanced by S200 feature parsing, is input into a pre-trained hybrid model to obtain a power consumption prediction sequence for the next 60 minutes. At the same time, the system obtains the current time in real time. Actual power consumption measurement value and compared with the time in the previous prediction period Predicted value By comparing the results, the instantaneous prediction error is calculated. .

[0085] In one possible implementation, the system introduces a lightweight error feedback correction mechanism. This mechanism corrects the current instantaneous error. A portion of this is added to the newly generated future prediction sequence in an exponentially decaying manner, thereby quickly correcting potential systematic biases or lags in response to sudden changes in the model. The correction formula is: ,in , To correct the strength coefficient, This is the attenuation coefficient. This mechanism significantly improves the ability to predict and track sudden load changes.

[0086] S330: Monitor prediction performance and manage model iteration.

[0087] In this invention, the system continuously calculates key performance indicators for prediction online, including root mean square error, mean absolute percentage error, and the correlation coefficient between predicted and actual values. These key indicators are displayed in real-time in a dashboard format. The system sets performance thresholds; if the performance consistently exceeds these thresholds, an alarm is triggered.

[0088] In one possible implementation, model updates employ a strategy combining shadow mode and active learning. The new model version initially runs in shadow mode, where its predictions are not linked to actual production decisions and are only used for performance comparison with the current production model. Only after the new model consistently outperforms the production model in shadow mode and passes safety checks is a smooth transition implemented.

[0089] like Figure 5 As shown, the S400 is based on deep reinforcement learning technology to build an intelligent agent that can automatically learn to make computer-computer collaborative scheduling decisions in complex dynamic environments.

[0090] S410: Define the key elements of a reinforcement learning environment.

[0091] In this invention, the state space of the environment Precisely defined, its information all comes from real-time data streams and forecasting modules: the grid-side status includes real-time load rates obtained from the grid dispatching system. and transmission line regulation margin The computing power status includes the latest energy consumption prediction sequence obtained from the S300 module and the real-time average cluster utilization rate obtained from the resource management platform. ; Context state includes the current time period label and data center environment readings Action space Consider a set of discrete, executable scheduling instructions, primarily including instructions for adjusting the overall utilization of the computing cluster and instructions for adjusting power requests from the grid. Reward function. It is the key to guiding the learning strategy of the intelligent agent. It is constructed as a weighted sum of multiple objectives. Positive rewards come from the improvement of the matching degree of the power-computing curve, while negative rewards come from the total energy consumption exceeding the limit, the incurring task delay costs, or the violation of any physical and security constraints.

[0092] In one possible implementation, the specific weight coefficients of the reward function are determined through multiple rounds of simulation experiments and grid search, and can be dynamically adjusted at different operational stages. To encourage exploration, a regularization term based on action entropy is added to the reward function in the early stages of training.

[0093] S420: The agent is trained using a deep Q-network algorithm.

[0094] In this invention, the decision-making core of the agent is a deep Q-network, whose input is a state vector. The output is for each optional action. The Q-value is determined by the network architecture. The network structure typically contains 3 to 5 fully connected layers using the ReLU activation function. Training employs the classic DQN framework, including the use of an experience replay buffer to break data correlations and a target Q-network to stabilize the training objective. The agent interacts with the environment, tries different actions, observes state transitions and rewards, and stores these experiences in the buffer. During training, a batch of experience data is periodically sampled from the buffer to update the parameters of the online Q-network.

[0095] In one possible implementation, to address the continuity and complexity of the state space, a DuelingDQN network architecture is employed, decomposing the Q-value into the sum of a state value function and a dominance function to better evaluate the relative value of different actions. The exploration strategy uses an ε-greedy algorithm, with an initial exploration rate ε set to 0.5, which linearly decays to 0.05 with each training step to balance exploration and exploitation.

[0096] S430: Deploys intelligent agents and makes online decisions.

[0097] In this invention, after training is complete and the evaluation is satisfactory, the agent's policy network is deployed to the production environment. During online operation, the system collects the current state every minute. The action is input into a policy network, which outputs the Q-value for each action. The action with the highest Q-value is selected as the scheduling decision for the current time step. The decision result, along with its confidence level, is output.

[0098] In one possible implementation, the system sets up safety barriers. When the agent's output decision has too low confidence, or the decision triggers any preset safety rule, the low-confidence decision is intercepted, and a conservative, safe alternative strategy from the rule base is executed instead. All agent decisions and their final execution results are recorded, forming new experience samples.

[0099] like Figure 6 As shown, based on the initial decision made by the intelligent agent, the S500 establishes a multi-objective optimization model to find the optimal matching solution between power supply and computing power demand from a global and longer time scale.

[0100] S510: Constructing a computer-computer collaborative multi-objective optimization problem.

[0101] In this invention, the decision variable of the optimization problem is defined as the target utilization rate of the computing cluster at each time step within the future scheduling cycle. and the amount of power adjustment requested from the power grid. The objective function comprises three competing yet balanced aspects: the primary objective is to maximize the temporal alignment between the electricity supply curve and the computing power demand curve, using the Pearson correlation coefficient. The first objective is to quantify this; the second objective is to minimize the total energy consumption cost of the computing center during this cycle, taking into account time-of-use electricity pricing. The third objective is to minimize overall scheduling costs, including the costs of delayed computing tasks due to adjustments in computing load, and the additional maintenance costs caused by frequent device start-ups and shutdowns.

[0102] In one possible implementation, computing power requires power consumption. Modeled as utilization and ambient temperature The function, function form Parameters in The contribution weights and relationships provided by the S200 energy efficiency analysis module are used to establish a quantitative relationship between decision variables and objective functions.

[0103] S520: Solving for the Pareto optimal frontier using the NSGA-II algorithm.

[0104] In this invention, a non-dominated sorting genetic algorithm with an elitist strategy is used to solve the aforementioned multi-objective problem. Each candidate solution is encoded with real numbers, and its chromosome is directly represented as a vector. The algorithm initializes a population of size N and then iteratively evolves iteratively through genetic operations such as selection, crossover, and mutation. In each generation, individuals are non-dominated and their crowding is calculated based on the three objective function values ​​to simultaneously approximate the optimal solution set while maintaining solution diversity. After a set number of generations of evolution, the algorithm outputs a set of non-dominated solutions that cannot surpass each other in the three objectives, representing the optimal scheduling schemes under different trade-offs.

[0105] In one possible implementation, the crossover operation employs simulated binary crossover, and the mutation operation employs polynomial mutation, to enable efficient exploration and development within the solution space. Algorithm parameters are calibrated experimentally based on the problem size.

[0106] S530: Select the final scheduling scheme from the optimal solution set.

[0107] In this invention, the NSGA-II algorithm outputs an optimal set of solutions, rather than a single solution. The system selects the final execution plan from the optimal set of solutions based on the current operational strategy and preferences.

[0108] In one possible implementation, the system provides a human-computer interaction interface that visualizes the optimal solution set as a 3D scatter plot or parallel coordinate graph. Operators can directly select solutions of interest on the graph to view their corresponding detailed scheduling plans and performance forecasts, thus making a final selection based on expert experience.

[0109] like Figure 7 As shown, the S600 transforms optimization decisions into specific, executable control commands and sends them securely and reliably to the computing cluster and power grid dispatching system to complete the control.

[0110] S610: Analyzes and optimizes the scheme, generating dual-system control instructions.

[0111] In this invention, the target utilization rate sequence of the computing power cluster for each minute in the future period is extracted from the final optimization scheme selected by S530. and grid power adjustment request sequence For the computing power side, the instruction generation module converts these sequences into control instructions that the computing power resource management platform can recognize. These instructions may include: adjusting the CPU limits of a batch of containers to a specific value within a specific time range, or adjusting the overall cluster's job scheduling strategy to be more conservative or aggressive. For the power grid side, the instruction generation module encapsulates the power adjustment request sequences into standard messages conforming to the power dispatch communication protocol. These messages contain information such as the requested power value, start time, and duration.

[0112] In one possible implementation, the generated instructions will be accompanied by a globally unique instruction ID and timestamp, and will be logically validated in advance to ensure the rationality and security of the instructions.

[0113] S620: Sends and confirms commands synchronously through the security interface.

[0114] In this invention, instructions are issued separately through dedicated, highly available communication links. Instructions from the computing power side are typically issued via a RESTful API or message queue provided by the resource management platform. Instructions from the power grid side are transmitted to a designated interface of the power grid energy management system via the power dispatch data network using an encrypted tunnel. The issuance process employs a request-confirmation mechanism: after an instruction is issued, the system waits for confirmation from the recipient. If no confirmation is received within a specified time, the system will retry or escalate the alarm according to a preset strategy.

[0115] In one possible implementation, all command transmissions employ two-way authentication (and end-to-end encryption) to prevent command tampering or eavesdropping. The delivery module features a high-availability design to avoid command interruption due to single points of failure.

[0116] S630: Monitor execution results and implement feedback optimization.

[0117] Specifically, step S630 is responsible for full-link monitoring and evaluation of the execution effect after the scheduling command is issued. This step collects the actual utilization curve of the computing cluster and the actual power supply curve of the power grid in real time, and performs multi-granularity comparison and analysis with the expected targets in the optimization scheme. The system calculates the minute-level deviation and the achievement rate of key indicators within the cycle. When the deviation exceeds the preset threshold, the system automatically starts the root cause diagnosis program to investigate problems such as sudden changes in the external environment, execution link obstruction, or model inaccuracy, and can dynamically fine-tune the correction coefficients of the prediction model or the weight parameters of the agent reward function based on recent deviation data to achieve adaptive adjustment during operation.

[0118] In one possible implementation, the system stores the complete scheduling data from each round as an experience trajectory in a historical case library. The system periodically uses this new data to initiate background learning tasks: updating the experience replay buffer of the reinforcement learning agent with new experience and performing offline training to optimize the policy network parameters; simultaneously fine-tuning the time-series prediction model with new energy consumption data; and adjusting the weight settings of the multi-objective algorithm based on historical optimization preferences. Through this continuous feedback and iterative learning mechanism, the system achieves self-improvement and evolution from data to model, continuously adapting to changes in business load and external environment, and improving its long-term operational intelligence and overall efficiency.

[0119] Example 2

[0120] A multi-dimensional energy consumption analysis and load forecasting system for computing centers, the system comprising:

[0121] The data acquisition and preprocessing module is used to acquire diverse and heterogeneous data in the computing center in real time and preprocess it to form a standardized feature vector sequence.

[0122] The energy efficiency feature analysis module is equipped with a pre-trained large-scale energy efficiency analysis model, which is used to receive standardized feature vector sequences and analyze the contribution of different data features to energy consumption through the self-attention mechanism of the large-scale model, and identify the key driving factors affecting energy consumption fluctuations.

[0123] The energy consumption prediction module is equipped with a hybrid time series model that combines a long short-term memory network and a Transformer architecture. It is used to make rolling predictions of future energy consumption in minutes based on historical sequences after feature parsing, and to dynamically correct the error between real-time measurement values ​​and prediction values ​​to generate a high-precision energy consumption prediction sequence for future periods.

[0124] The collaborative scheduling agent module is a power-computer dynamic correlation decision-making agent built based on a deep reinforcement learning model. It is used to take state information including real-time power grid status, computing load status, environmental status and future energy consumption prediction sequence as input, and output preliminary collaborative scheduling instructions for computing utilization adjustment and power grid power request.

[0125] The multi-objective optimization module is used to construct a multi-objective optimization model based on the initial cooperative scheduling instructions, with the optimization objectives of maximizing the power-computer power matching degree, minimizing the total energy consumption cost, and minimizing the comprehensive scheduling cost. The model is solved using a multi-objective evolutionary algorithm, and the final scheduling scheme is selected from the set of optimal solutions.

[0126] The instruction execution and feedback module is used to parse the final scheduling scheme into specific computing cluster control instructions and power grid adjustment request instructions, and send them to the computing resource management platform and the power grid dispatching system for execution through a secure interface. At the same time, it monitors the actual effect after instruction execution and feeds back the execution deviation data to the energy consumption prediction module and the collaborative scheduling agent module for model optimization and adaptive adjustment.

[0127] Those skilled in the art will understand that the embodiments of this application are provided as methods, systems, or computer program products. Therefore, this application takes the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application takes the form of a computer program product implemented on one or more computer storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code. The solutions in the embodiments of this application are implemented using various computer languages, exemplified by the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0128] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, are implemented by computer program instructions. These computer program instructions are provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams.

[0129] These computer program instructions are also stored in a computer read-memory that can direct a computer or other programmed data processing device to operate in a particular manner, such that the instructions stored in the computer read-memory produce an article of manufacture including instruction means that implement the functions specified in the flowchart or multiple flowcharts and / or block diagram blocks or multiple block diagrams.

[0130] These computer program instructions are also loaded onto a computer or other programming data processing device to cause a series of operational steps to be performed on the computer or other programming device to produce a computer-implemented process, such that the instructions, which execute on the computer or other programming device, provide steps for implementing the functions specified in the flowchart flow or multiple flows and / or the block diagram blocks or multiple blocks.

[0131] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0132] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for multi-dimensional analysis of energy consumption and load forecasting in computing centers, characterized in that, The method includes: Real-time collection of diverse and heterogeneous data within the computing center, followed by preprocessing to form a standardized sequence of feature vectors; The standardized feature vector sequence is input into the pre-trained large-scale energy efficiency analysis model. The self-attention mechanism of the large-scale energy efficiency analysis model is used to analyze the contribution of different data features to the energy consumption and identify the key driving factors affecting energy consumption fluctuations. Based on the historical sequence after feature parsing, a hybrid time series model combining a long short-term memory network and a Transformer architecture is used to perform rolling predictions of energy consumption in the next minute. The error between the real-time measurement value and the prediction value is dynamically corrected to generate a high-precision energy consumption prediction sequence for the future period. Construct a decision-making intelligent agent with dynamic association between electricity and computing. The decision-making intelligent agent takes as input state information including real-time grid status, computing load status, environmental status and future energy consumption prediction sequence. Based on a deep reinforcement learning model, it outputs preliminary collaborative scheduling instructions for computing utilization adjustment and grid power request. Based on the initial cooperative scheduling instructions, a multi-objective optimization model is constructed with the optimization objectives of maximizing the power-computer power matching degree, minimizing the total energy consumption cost, and minimizing the comprehensive scheduling cost. The model is then solved using a multi-objective evolutionary algorithm, and the final scheduling scheme is selected from the set of optimal solutions. The final scheduling scheme is parsed into specific computing cluster control commands and power grid adjustment request commands, and then sent to the computing resource management platform and the power grid dispatching system for execution through secure interfaces. At the same time, the actual effect after the commands are executed is monitored, and the execution deviation data is fed back to the prediction model and decision-making agent for optimization and adaptive adjustment.

2. The method for multi-dimensional analysis of energy consumption and load forecasting of computing centers according to claim 1, characterized in that, The pre-trained energy efficiency analysis model dynamically identifies and quantifies the impact patterns of key driving factors on energy consumption by analyzing the weight matrix of its self-attention mechanism. The analysis includes averaging and normalizing the attention weights along the time and feature dimensions to generate a structured report containing the relationship between feature contribution weights and nonlinear effects.

3. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, The hybrid time series model adopts a cascaded architecture: The preceding stage consists of stacked long short-term memory networks, used to extract temporal-dependent features from long historical sequences; The latter stage is a Transformer decoder, which takes the feature sequence output from the former stage as input and performs multi-step prediction through a masked self-attention mechanism; the model output layer is directly mapped to the energy consumption prediction values ​​for multiple consecutive minutes in the future.

4. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 3, characterized in that, The dynamic correction is achieved through an error feedback mechanism: The instantaneous prediction error at the latest moment is added to the prediction sequence for future periods in an exponentially decaying manner. The correction formula is as follows: ,in, The corrected predicted value, These are the original predicted values. The attenuation coefficient is... For instantaneous error, To correct the strength coefficient.

5. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, The decision-making agent is based on a deep Q-network architecture and uses a Dueling DQN structure to decompose the Q value into a state value function and an advantage function. The exploration strategy employs an ε-greedy algorithm, and the exploration rate decreases linearly with the number of training steps. Among them, the intelligent agent has a safety barrier when making online decisions. When the confidence of the output decision is too low or violates the preset safety rules, it automatically switches to the rule-based backup strategy.

6. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, In the multi-objective optimization model, the computing power demand power consumption is modeled as a function of computing power utilization and ambient temperature. The specific parameters in the function form are determined by the feature contribution weights and relationships extracted from the large energy efficiency analysis model.

7. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, The multi-objective optimization model is solved using a non-dominated sorting genetic algorithm with an elitist strategy. The algorithm encodes candidate solutions with real numbers, performs population evolution by simulating binary crossover and polynomial mutation, and filters the Pareto optimal solution set based on non-dominated sorting and crowding calculation. Finally, it selects the final scheduling scheme from the solution set through a human-computer interactive visual interface.

8. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, The process of parsing the final scheduling scheme into specific control instructions includes: generating a globally unique instruction identifier and timestamp for each instruction and performing logical verification in advance; the instructions are sent to the computing resource management platform and the power grid scheduling system respectively through a secure communication link with two-way authentication and end-to-end encryption, using a request-confirmation mechanism.

9. The method for multi-dimensional analysis and load forecasting of computing center energy consumption according to claim 1, characterized in that, The actual effects of the monitoring commands after execution and the feedback optimization specifically include: establishing a historical case library to store complete scheduling experience trajectories, periodically starting background learning tasks, using new data to offline update the policy network parameters of the decision-making agent, fine-tuning the hybrid time series prediction model, and adjusting the weight preferences of the multi-objective optimization model to achieve the self-evolution of the system.

10. A computing center energy consumption multi-dimensional analysis and load forecasting system for performing the method as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire diverse and heterogeneous data in the computing center in real time and preprocess it to form a standardized feature vector sequence. The energy efficiency feature analysis module is equipped with a pre-trained large-scale energy efficiency analysis model, which is used to receive standardized feature vector sequences and analyze the contribution of different data features to energy consumption through the self-attention mechanism of the large-scale model, and identify the key driving factors affecting energy consumption fluctuations. The energy consumption prediction module is equipped with a hybrid time series model that combines a long short-term memory network and a Transformer architecture. It is used to make rolling predictions of future energy consumption in minutes based on historical sequences after feature parsing, and to dynamically correct the error between real-time measurement values ​​and prediction values ​​to generate a high-precision energy consumption prediction sequence for future periods. The collaborative scheduling agent module is a power-computer dynamic correlation decision-making agent built based on a deep reinforcement learning model. It is used to take state information including real-time power grid status, computing load status, environmental status and future energy consumption prediction sequence as input, and output preliminary collaborative scheduling instructions for computing utilization adjustment and power grid power request. The multi-objective optimization module is used to construct a multi-objective optimization model based on the initial cooperative scheduling instructions, with the optimization objectives of maximizing the power-computer power matching degree, minimizing the total energy consumption cost, and minimizing the comprehensive scheduling cost. The model is solved using a multi-objective evolutionary algorithm, and the final scheduling scheme is selected from the set of optimal solutions. The instruction execution and feedback module is used to parse the final scheduling scheme into specific computing cluster control instructions and power grid adjustment request instructions, and send them to the computing resource management platform and the power grid dispatching system for execution through a secure interface. At the same time, it monitors the actual effect after instruction execution and feeds back the execution deviation data to the energy consumption prediction module and the collaborative scheduling agent module for model optimization and adaptive adjustment.

Images