A power supply elastic scheduling system and method for intelligent computing center

Abstract

The present invention discloses a power supply elastic dispatching system and method for an intelligent computing center, which relates to the technical field of electric energy management, including establishing a communication connection between a dispatching switch system and a remote control center; collecting real-time load power data, and determining the elastic dispatching load by setting a safety margin coefficient; based on the real-time load power data, using a long short-term memory network model to predict the load power, and detecting load anomalies in real time; extracting static data segments and dynamic data segments, and determining the power supply regulation rate by calculating the ratio of static power to dynamic change rate; defining multi-objective optimization, and generating an optimized dispatching strategy based on real-time load power, predicted load power, and static and dynamic data segments; calculating the capacity adjustment amount, generating specific dispatching instructions in combination with the optimized dispatching strategy, and executing corresponding capacity adjustment operations. The present invention improves the overall performance and operational reliability through load prediction and anomaly detection and multi-objective optimization algorithms.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power energy management, and in particular to a system and method for flexible power dispatching of an intelligent computing center. Background Art

[0002] With the rapid development of information technology, the power management system of intelligent computing centers, as core facilities supporting information processing and data storage in various fields of modern society, has become increasingly important. Traditional power management methods mostly use static configuration and manual adjustment strategies, relying on preset load thresholds and fixed adjustment parameters, which are difficult to cope with the dynamic changes in the load of computing centers. This static management method often exhibits problems of slow response and untimely adjustment when dealing with sudden load fluctuations, resulting in inefficient energy utilization and unstable system operation. In addition, with the continuous growth of computing needs, the power supply system of computing centers faces higher energy consumption and more complex load management challenges. There is an urgent need for more intelligent and flexible scheduling methods to optimize power configuration and improve the energy efficiency and stability of the overall system.

[0003] Although existing technologies have made some progress in power management of intelligent computing centers, they still have significant deficiencies in multi-objective optimization and real-time response capabilities. First, rule-based adjustment strategies lack flexibility and adaptability, making it difficult to accurately reflect complex changes in loads, and easily leading to over-adjustment or under-adjustment. Second, existing machine learning models have limited accuracy and real-time performance in load prediction and anomaly detection, and cannot effectively respond to rapid fluctuations in computing center loads and sudden abnormal events. In addition, in terms of multi-objective optimization, most methods fail to fully balance energy consumption, response time, and system stability, resulting in the performance of other key indicators being sacrificed when one goal is improved during the optimization process. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent computing center power elastic scheduling system to solve the problems that rule-based adjustment strategies lack flexibility and adaptability, are difficult to accurately reflect the complex changes in loads, and fail to fully balance energy consumption, response time, and system stability.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a power supply elastic scheduling system for an intelligent computing center, which comprises:

[0008] The communication module is responsible for establishing the communication connection between the dispatch switch system and the remote control center;

[0009] The data processing module is responsible for collecting real-time load power data and determining the elastic scheduling load by setting the safety margin coefficient;

[0010] The load prediction module is responsible for load power prediction based on real-time load power data using a long short-term memory network model and for detecting load anomalies in real time;

[0011] The data segment extraction module is responsible for extracting the static data segment and the dynamic data segment, and determining the power supply regulation rate by calculating the ratio of the static power to the dynamic change rate;

[0012] The multi-objective optimization module is responsible for defining multi-objective optimization and generating an optimal scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments;

[0013] The scheduling instruction module is responsible for calculating the capacity adjustment amount, generating specific scheduling instructions based on the optimized scheduling strategy, and executing the corresponding capacity adjustment operations based on the received scheduling instructions.

[0014] As a preferred solution of the power elastic scheduling system of the intelligent computing center described in the present invention, wherein: collecting real-time load power data, determining the elastic scheduling load by setting the safety margin coefficient, including the following steps,

[0015] Deploy power sensors at key power supply points in the intelligent computing center to collect load power data in real time;

[0016] The load power data collected by the power sensor is transmitted to the central data processing unit in real time;

[0017] Format the collected load power data;

[0018] According to the actual measurement results, determine the total load power capacity and calculate the real-time load power;

[0019] Set the safety margin coefficient. Based on the real-time load power and the set safety margin coefficient, calculate the current load that can be flexibly dispatched. The expression is:

[0020] ;

[0021] in, To flexibly dispatch the load, For time point The load power, is the total load power capacity, is the safety margin factor;

[0022] Verify the elastic scheduling load to ensure the calculated Do not exceed preset safety margin limits.

[0023] As a preferred solution of the power elastic dispatching system of the intelligent computing center described in the present invention, the load power prediction is performed based on the real-time load power data using a long short-term memory network model, including the following steps:

[0024] Get the past from the real-time database Minute-by-minute load power data to form a time series data set;

[0025] Perform data cleaning and normalization on the collected time series data set, and extract features for prediction from historical load power data;

[0026] Select the long short-term memory network model LSTM which includes output layer, hidden layer and output layer;

[0027] The mean square error is used as the loss function, and the Adam optimizer is used to optimize the model parameters;

[0028] Input the prepared training data into the LSTM model for multiple rounds of iterative training. After each round of training, evaluate the performance of the model on the validation set and determine the optimal number of training rounds through the early stopping mechanism.

[0029] Enter the latest The minute load power data is fed into the trained LSTM model to obtain the future Forecast load power for the minute.

[0030] As a preferred solution of the power elastic dispatching system of the intelligent computing center of the present invention, the real-time detection of load abnormality includes the following steps:

[0031] Select the Isolation Forest model, use the actual load power data and its related features as model input, and set the number of trees and subsample size;

[0032] The isolation forest model is trained using historical load power data. Multiple decision trees are constructed by randomly selecting features and randomly splitting values. Each tree is designed to "isolate" a data point in the data. After training, the isolation forest model generates an anomaly score for each data point.

[0033] The current load power data is input into the isolation forest model, and the isolation forest model outputs an anomaly score. The anomaly score is compared with a preset threshold to determine whether the load power data is abnormal.

[0034] When load abnormality is detected, the early warning mechanism is triggered and corresponding emergency measures are implemented.

[0035] As a preferred solution of the power elastic scheduling system of the intelligent computing center of the present invention, wherein: extracting static data segments and dynamic data segments includes the following steps:

[0036] Extract scheduling instruction logs from the power management system to obtain time series data including load changes;

[0037] Defining Variations is the continuous time period during which the load power changes in the dispatch instruction;

[0038] The acquired change segment Perform segmentation processing and divide it into multiple independent change data segments;

[0039] Apply linear regression analysis to each changing data segment, extract the intercept of each trend line, and form a static data segment set K;

[0040] Collect time series data of actual load power from the power monitoring system;

[0041] The slope of each time point is calculated using the difference method or moving regression method for each data point, and the slope is smoothed using a sliding window;

[0042] All smoothed slopes are averaged to obtain the dynamic data segment.

[0043] As a preferred solution of the power elastic scheduling system of the intelligent computing center described in the present invention, wherein: the power regulation rate is determined by calculating the ratio of static power to dynamic change rate, including the following steps:

[0044] All intercepts in the static data segment set are averaged to obtain the average power of the static data segment;

[0045] The average rate of change of the dynamic data segment is obtained by averaging all the smoothed slopes in the dynamic data segment. ;

[0046] The average power through the static data segment Average rate of change of dynamic data segment Calculates the power supply regulation rate.

[0047] As a preferred solution of the power elastic scheduling system of the intelligent computing center described in the present invention, wherein: defining multi-objective optimization, generating an optimized scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments, including the following steps,

[0048] Set minimizing energy consumption, minimizing response time, and maximizing stability as optimization goals;

[0049] According to the defined optimization goal, construct the objective function and set the constraints;

[0050] Combine the objective function and constraints to form the final optimization model;

[0051] Select hybrid genetic algorithm and particle swarm optimization as the optimization algorithm;

[0052] The real-time load power, predicted load power data, static data segment and average change rate of dynamic data segment are input into the optimization model and used as the input of GA-PSO algorithm;

[0053] Run the GA-PSO algorithm to search for the optimal scheduling strategy;

[0054] The optimal scheduling strategy is obtained through iterative optimization of the GA-PSO algorithm.

[0055] As a preferred solution of the power elastic scheduling system of the intelligent computing center described in the present invention, the hybrid genetic algorithm and particle swarm optimization algorithm are as follows:

[0056] Generate an initial population, assign velocity and position to each potential solution, and initialize the particle swarm;

[0057] Calculate the fitness value of each individual according to the objective function, select individuals with higher fitness as parents, and perform crossover and mutation operations;

[0058] Randomly select parent individuals for crossover to generate offspring individuals, and randomly mutate the offspring individuals to increase population diversity;

[0059] Update the particle's speed and position according to the current speed, individual best position and global best position, and recalculate the fitness value according to the updated particle position;

[0060] When the maximum number of iterations is reached or the fitness value converges to the preset threshold, the termination condition check;

[0061] The individual with the highest fitness is selected from the population as the optimized scheduling strategy.

[0062] As a preferred solution of the power elastic scheduling system of the intelligent computing center of the present invention, the capacity adjustment amount is calculated, a specific scheduling instruction is generated in combination with the optimization scheduling strategy, and the corresponding capacity adjustment operation is performed according to the received scheduling instruction, including the following steps:

[0063] By combining flexible scheduling of load , power regulation rate And the coefficients of the optimized scheduling strategy , calculate the capacity adjustment , the expression is,

[0064] ;

[0065] when >0, an increase capacity instruction is generated; when When <0, a capacity reduction instruction is generated;

[0066] Determine whether to generate a scheduling instruction to increase or decrease capacity according to the calculated capacity adjustment result;

[0067] Construct the scheduling instruction content and record the generated scheduling instruction in the scheduling log.

[0068] In a second aspect, the present invention provides a method for flexible scheduling of power supply in an intelligent computing center, comprising:

[0069] Establish communication connection between dispatch switch system and remote control center;

[0070] Collect real-time load power data and determine the flexible scheduling load by setting the safety margin coefficient;

[0071] The load prediction module is responsible for load power prediction based on real-time load power data using a long short-term memory network model and for detecting load anomalies in real time;

[0072] Extract static data segments and dynamic data segments, and determine the power supply regulation rate by calculating the ratio of static power to dynamic change rate;

[0073] Define multi-objective optimization and generate optimal scheduling strategies based on real-time load power, predicted load power, and static and dynamic data segments;

[0074] The beneficial effects of the present invention are as follows: by combining load power prediction and anomaly detection, the future load change trend can be accurately predicted, and the resource allocation strategy can be optimized in advance. At the same time, it can quickly identify and respond to abnormal load patterns, prevent potential system failures and energy consumption anomalies, enhance its adaptability and emergency response capabilities to sudden load changes, and ensure the scientificity and rationality of the regulation rate by comprehensively analyzing the static and dynamic characteristics of the load, which not only optimizes energy distribution and improves the accuracy of power management, but also improves the overall efficiency and responsiveness, making power regulation more efficient and intelligent. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0076] Figure 1 This is a system diagram of the power elastic scheduling system of the intelligent computing center in Example 1.

[0077] Figure 2This is a schematic diagram of determining the load capacity adjustment direction in Example 1. DETAILED DESCRIPTION

[0078] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0079] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0080] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0081] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and provides a power supply elastic scheduling system for an intelligent computing center, comprising the following steps:

[0082] The communication module is responsible for establishing the communication connection between the dispatch switch system and the remote control center;

[0083] Specifically, select an MQTT Broker with excellent performance, such as Eclipse Mosquitto, and deploy it in the control center to ensure that the Broker has high availability to prevent single point failures; configure the Broker IP address and port number, enable message persistence, ensure that messages are not lost after the Broker is restarted, and reasonably set the maximum number of connections based on the expected number of devices to avoid resource exhaustion; assign a unique client ID to each scheduling switch system, such as Switch001, Switch002, etc., set a reasonable keep-alive interval, such as 60 seconds, to ensure the real-time nature of the connection, configure the will message, and ensure that other systems are notified when the client is accidentally disconnected so that emergency measures can be taken.

[0084] Furthermore, ensure that the network connection between the control center and all dispatching switch systems has high bandwidth and low latency characteristics, preferably Ethernet or high-speed wireless network, to achieve redundant network paths, improve network reliability and fault tolerance; assign static IP addresses to the control center and dispatching switch systems to avoid connection interruptions caused by dynamic IP changes;

[0085] Generate a private key and certificate signing request (CSR), submit the CSR to the CA, obtain a signed certificate, install the certificate and private key to the MQTT Broker and the scheduling switch system; configure the TLS parameters of the MQTT Broker and the client, enable TLS on the MQTT Broker, and configure the client to use TLS connection;

[0086] Configure the Broker to require the client to provide a valid SSL certificate, configure the client to present the SSL certificate to the Broker for identity authentication, regularly update and maintain CRL, revoke lost or leaked certificates, and regularly replace private keys and certificates to prevent security risks caused by long-term use; each time a connection is established, the Broker verifies the client certificate to ensure that it is issued by a trusted CA and has not been revoked. The client also verifies the Broker certificate to ensure that the connection is established with the correct Broker.

[0087] After the configuration is completed, perform an initial connection test to ensure that a TLS-encrypted MQTT connection can be successfully established between the control center and the dispatch switch system. Specifically, use an MQTT client (such as MQTT.fx or a custom script) to try to connect to the Broker, publish and subscribe to test topics, verify the correct reception and sending of messages, and ensure the validity of the connection.

[0088] Simulate message transmission under high concurrency and test the performance of communication protocols under high load. Test indicators include throughput (messages / second), latency (the time from message publishing to message receiving) and successful connection rate. Perform vulnerability scans to detect potential security risks, such as unauthorized access and data leakage. Use penetration testing tools to simulate attacks and verify the effectiveness of TLS encryption and two-way authentication.

[0089] Set the initial reconnection interval (such as 1 second) and adopt an exponential backoff strategy to gradually increase the reconnection interval to avoid server overload caused by frequent reconnections. Set an upper limit on the number of reconnection attempts (such as 10 times) to prevent infinite reconnections. Implement automatic reconnection logic on both the client and the Broker to ensure that both parties can automatically restore the connection after the connection is interrupted. Monitor the connection status in real time. When multiple consecutive reconnection failures occur, an alarm is triggered to notify the operation and maintenance personnel to intervene manually.

[0090] The dispatch switch system usually contains the following main components:

[0091] Power Distribution Units: PDUs are used to distribute power to various servers and network devices. They have the ability to monitor and manage various power supplies. They usually support remote monitoring, load balancing, and overload protection. They can communicate with the central management system through a network interface to achieve real-time data collection and remote control.

[0092] Automated circuit breakers: When current anomalies (such as short circuits and overloads) are detected, they can quickly cut off the power supply to prevent equipment damage and fire risks. These circuit breakers integrate intelligent sensors and communication modules, which can provide real-time feedback of current status to the management system to achieve automated safety control.

[0093] Power module: responsible for receiving power from the main power supply and distributing it to different load devices, supporting capacity adjustment and redundant power supply configuration. Advanced power modules have high-efficiency conversion capabilities, modular design and high reliability, and can respond quickly when load demand changes;

[0094] Communication interface: The dispatch switch system exchanges data and transmits dispatch instructions with the remote control center through the communication interface. Common communication protocols include MQTT, Modbus, BACnet, etc. These protocols support low-latency, high-reliability real-time data transmission to ensure the timely execution of dispatch instructions;

[0095] Specifically, Schneider Electric's APC series PDUs and Emerson's ARM series automatic circuit breakers.

[0096] The remote control center usually consists of the following key modules: data acquisition and processing module, prediction and anomaly detection module, multi-objective optimization module, scheduling strategy generation and scheduling instruction sending module, and human-computer interaction interface;

[0097] Specifically, for example, Siemens' SCADA system provides comprehensive data collection, monitoring and control functions, integrates advanced prediction and analysis tools, and supports multi-objective optimization and intelligent scheduling strategy generation;

[0098] and Schneider Electric’s EcoStruxure platform: enables real-time data processing and intelligent decision-making through cloud and edge computing, and supports automatic generation and sending of multi-objective optimization and dynamic scheduling instructions.

[0099] The data processing module is responsible for collecting real-time load power data and determining the elastic scheduling load by setting the safety margin coefficient;

[0100] Deploy power sensors at key power supply points in the intelligent computing center (such as UPS, PDU, server racks, etc.) to collect load power data in real time;

[0101] The load power data collected by the power sensor is transmitted to the central data processing unit (CDPU) in real time via industrial Ethernet or wireless communication;

[0102] Format the collected real-time load power data to ensure that the transmitted data format is consistent, such as JSON format, including timestamp and value pairs;

[0103] According to the actual measurement results, the total load power capacity is determined as follows:

[0104] ;

[0105] in, is the total load power capacity (watt, W), For the The maximum capacity of each power supply unit (Watts), is the total number of power supply units, is the index variable of the power unit;

[0106] Calculate the real-time load power, the expression is,

[0107] ;

[0108] in, For time point The real-time load power (dimensionless), For time point The load power (watt, W).

[0109] Set the safety margin factor, =0.2, where is the safety margin factor, and its value range is 0< <1; It should be noted that the selection =0.2 (i.e. 20%) as a safety margin factor to ensure sufficient adjustment space under high load conditions to prevent overload risks;

[0110] Calculate the current flexibly dispatchable load based on the real-time load power and the set safety margin factor , the expression is,

[0111] ;

[0112] Verify the elastic scheduling load to ensure the calculated Not exceeding the preset safety margin limit,

[0113] ;

[0114] If the calculation result If the preset range is exceeded, the alarm mechanism is triggered, and the operation and maintenance personnel are notified to conduct a status check and adjust the safety margin coefficient according to the actual situation. Or re-evaluate the total load power capacity , ensuring the stable operation of the intelligent computing center.

[0115] The load prediction module is responsible for load power prediction based on load power data using a long short-term memory network model and for detecting load anomalies in real time;

[0116] Specifically, get the past from the real-time database The load power data of each minute is used to form a time series data set, which is specifically expressed as:

[0117] ;

[0118] in, Indicates time point Load power (watt, W), Indicates at a point in time Before The load power in the time interval is Indicates at a point in time Before Load power in a time interval;

[0119] The collected time series data set is preprocessed, including removing missing values, outlier processing and data smoothing to ensure data quality. Features that are helpful for prediction are extracted from historical load power data, such as time features (hours, days of the week), moving averages, load change rates, etc. The data is standardized or normalized to accelerate the model training process and improve prediction accuracy. Commonly used methods include Min-Max normalization and Z-score normalization.

[0120] The Long Short-Term Memory (LSTM) model is selected because it excels in capturing long-term dependencies in time series data. The input layer accepts preprocessed time series data sets, the hidden layer includes multiple LSTM units that can learn the temporal dynamic characteristics of the load data, and the output layer outputs future Load power forecast value for the time step;

[0121] The mean square error is used as the loss function, and the Adam optimizer is used to optimize the model parameters;

[0122] Input the prepared training data into the LSTM model for multiple rounds of iterative training. After each round of training, evaluate the performance of the model on the validation set to prevent overfitting, and determine the optimal number of training rounds through the early stopping mechanism.

[0123] Enter the latest The minute load power data is fed into the trained LSTM model to obtain the future Forecast load power for the minute.

[0124] Furthermore, the Isolation Forest model was chosen because of its efficient anomaly detection capability in high-dimensional data space and its insensitivity to noise;

[0125] The actual load power data and its related features (such as lag value, moving average, etc.) are used as model inputs, and the number of trees and subsample size are set. The number of trees is usually set to 100 to balance model performance and computational overhead, and the subsample size is 256 to ensure the diversity of each tree.

[0126] The Isolation Forest model is trained using historical load power data to enable it to learn the patterns of normal load behavior. Multiple decision trees are constructed by randomly selecting features and randomly splitting values. Each tree aims to "isolate" a data point in the data. After training, the Isolation Forest model generates an anomaly score for each data point. The lower the score, the more likely it is an anomaly point.

[0127] The current load power data is input into the isolation forest model, and the isolation forest model outputs anomaly scores. Compare and determine whether the load power data is abnormal. The expression is:

[0128] ;

[0129] ;

[0130] in, For time point The abnormal judgment result of represents the trained isolation forest detection model, For time point The abnormal score of Typically, a percentile based on the score distribution is chosen, such as the 5% quantile;

[0131] When detected When an abnormal event occurs, it will automatically record the abnormal event and send a warning notification (such as SMS, email or system notification) to the relevant operation and maintenance personnel to implement the corresponding emergency measures;

[0132] The specific emergency measures are as follows:

[0133] Automatically transfer part of the load to backup servers or other nodes with lower resource exhaustion to reduce the load pressure on the main system; adjust resource allocation strategies, such as dynamically expanding computing resources or limiting resource usage of non-critical tasks to ensure the stable operation of critical tasks; record the time when the abnormality occurs, load power value and related status for subsequent analysis and optimization.

[0134] The data segment extraction module is responsible for extracting the static data segment and the dynamic data segment, and determining the power supply regulation rate by calculating the ratio of the static power to the dynamic change rate;

[0135] Specifically, the scheduling instruction log is extracted from the power management system to obtain time series data including load changes;

[0136] Defining Variations is the continuous time period in which the load power changes in the scheduling instruction. For example, when the load power change rate exceeds the preset threshold, it is marked as a change period. The expression is:

[0137] ;

[0138] in, For time point Load power (watt, W), For the next time point The load power, is the threshold value of load change rate;

[0139] Furthermore, a standard deviation-based method is used to set the load change rate threshold The mean of the load change rate plus twice the standard deviation, for example,

[0140] ;

[0141] in, is the mean of the historical load change rate, is the standard deviation.

[0142] According to the time continuity and change characteristics, Divide into several sub-segments with consistent internal change trends, and use sliding windows or clustering-based methods (such as DBSCAN) to identify and divide the change segments;

[0143] Apply linear regression analysis to each variation data segment to fit the load power With time point The relationship between , extract the intercept of each trend line, and form a static data segment set K;

[0144] Furthermore, time series data of actual load power is collected from the power monitoring system;

[0145] The slope of each time point is calculated using the difference method or moving regression method for each data point to reflect the dynamic change of load power. To reduce the influence of noise, the slope is smoothed using a sliding window.

[0146] All smoothed slopes are averaged and the average slope is used as the dynamic feature of load change to obtain a dynamic data segment, which reflects the overall change trend of load power and is used for subsequent regulation rate determination and optimization of scheduling strategies.

[0147] Furthermore, by calculating the ratio of static power to dynamic change rate, the power supply regulation rate is determined, including the following steps:

[0148] All intercepts in the static data segment set are averaged to obtain the average power of the static data segment;

[0149] The average rate of change of the dynamic data segment is obtained by averaging all the smoothed slopes in the dynamic data segment. ;

[0150] The average power through the static data segment Average rate of change of dynamic data segment Calculating Power Regulation Rate , to guide the power supply regulation strategy.

[0151] It should be noted that Describes the load change rate per unit power of the power supply based on the static power. A lower value means that at static power, the power supply needs to respond faster to load changes, and vice versa.

[0152] The multi-objective optimization module is responsible for defining multi-objective optimization and generating an optimal scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments;

[0153] In order to ensure that the scheduling strategy can meet various requirements, the following three optimization goals are set:

[0154] Minimize energy consumption: reduce overall energy consumption, improve energy efficiency, and reduce operating costs; Minimize response time: ensure that dispatch instructions can be responded to in the shortest time and improve dynamic adjustment capabilities; Maximize stability: maintain operation within a safe load range, prevent overload or waste of resources, and improve reliability and stability;

[0155] According to the defined optimization goal, the following objective function is constructed:

[0156] ;

[0157] in, To comprehensively optimize the target value, the goal is to minimize this value. is the total energy consumption (watt-hour, Wh), reflecting the energy consumption during operation. The goal is to reduce this value as much as possible. Response time (seconds, s) measures the time interval from the issuance of a scheduling instruction to its execution. The goal is to shorten this time as much as possible. is a stability indicator (dimensionless) that evaluates the ability of the intelligent computing center to maintain operation within a safe load range. The goal is to increase this value as much as possible. is the energy consumption weight, which indicates the importance of energy consumption in the overall optimization. is the response time weight, indicating the importance of response speed in optimization. is the stability weight, which indicates the importance of the stability of the intelligent computing center in the optimization. The setting of the weight coefficient needs to be determined according to the specific application scenario and priority to ensure ;

[0158] In order to ensure the feasibility of the optimization results, a series of constraints need to be set, expressed as:

[0159] ;

[0160] in, , are the minimum and maximum allowable values ​​of load power, is the maximum load change rate that can be tolerated, is the upper limit of the load change rate. For time point The stability index, is the minimum required value for stability;

[0161] It should be noted that the power constraint is to ensure that the load power is within a safe range to avoid overload or waste of resources. The change rate constraint is to limit the load change rate to prevent instability due to rapid load changes. The stability constraint is to ensure that the stability index is not lower than the minimum requirement to ensure the reliability of operation.

[0162] Combine the objective function and constraints to form the final optimization model;

[0163] Select hybrid genetic algorithm and particle swarm optimization as the optimization algorithm;

[0164] The real-time load power, predicted load power data, static data segment and average change rate of dynamic data segment are input into the optimization model and used as the input of GA-PSO algorithm;

[0165] Run the GA-PSO algorithm to search for the optimal scheduling strategy, including:

[0166] Initialize the population and particle swarm: generate multiple initial scheduling strategy solutions; Iterative optimization: continuously optimize the scheduling strategy through crossover, mutation and particle update; Fitness evaluation: evaluate the fitness of the scheduling strategy in each generation to ensure that the objective function is gradually optimized;

[0167] Through iterative optimization of the GA-PSO algorithm, the optimal scheduling strategy is obtained, which achieves the best balance between energy consumption, response time and stability.

[0168] It should be noted that the hybrid genetic algorithm and particle swarm optimization algorithm are as follows:

[0169] Generate an initial population, including multiple potential scheduling strategy solutions, assign speed and position to each potential solution, and initialize the particle swarm;

[0170] Calculate the fitness value of each individual according to the objective function, select individuals with higher fitness as parents, and perform crossover and mutation operations;

[0171] Randomly select parent individuals for crossover to generate offspring individuals, and randomly mutate the offspring individuals to increase population diversity;

[0172] Update the particle's speed and position according to the current speed, individual best position and global best position. The expression is:

[0173] ;

[0174] ;

[0175] in, , is the updated particle speed and position, For particles In the The speed of generation, For particles In the The position of the generation, For particles The best location in history, is the historical best position of the group, is the inertia weight, , is the learning factor, , is a random number between [0,1], is the current iteration number, that is, the current number of steps executed by the particle swarm optimization algorithm, Represents the index of a specific particle in the particle swarm;

[0176] Recalculate the fitness value based on the updated particle position;

[0177] When the maximum number of iterations is reached or the fitness value converges to the preset threshold, the termination condition check;

[0178] The individual with the highest fitness is selected from the population as the optimized scheduling strategy.

[0179] It should be noted that the characteristics of Genetic Algorithm (GA) are strong global search capabilities, the ability to jump out of the local optimal solution and find the global optimal or better solution, and it is applicable to both discrete and continuous problems, with high flexibility and a wide range of applications; the characteristics of Particle Swarm Optimization (PSO) are fast convergence speed, the ability to quickly approach the optimal solution, and fewer parameters, easy to implement and debug. The hybrid GA-PSO algorithm combines the diversity of GA and the fast convergence of PSO, improves the optimization efficiency, reduces the possibility of falling into the local optimal solution, increases the probability of finding a better solution, and is more adaptable to multi-objective optimization, can handle complex multi-objective optimization problems, and meet the needs of practical applications.

[0180] The scheduling instruction module is responsible for calculating the capacity adjustment amount, generating specific scheduling instructions in combination with the optimized scheduling strategy, and executing the corresponding capacity adjustment operation according to the received scheduling instructions;

[0181] Furthermore, by combining the elastic scheduling of load , power regulation rate And the coefficients of the optimized scheduling strategy , calculate the capacity adjustment , the expression is,

[0182] ;

[0183] when >0, an increase capacity instruction is generated to increase the load capacity and improve energy supply; When <0, a capacity reduction instruction is generated to instruct to reduce the load capacity and reduce energy consumption;

[0184] Determine whether to generate a scheduling instruction to increase or decrease capacity according to the calculated capacity adjustment result;

[0185] Construct the dispatch instruction content and record the generated dispatch instruction in the dispatch log for subsequent tracking and auditing.

[0186] The generated dispatch instructions are formatted according to the predetermined communication protocol to ensure compatibility and correct parsing, and the dispatch instruction content is encrypted using an encryption algorithm (such as AES) to ensure communication security;

[0187] The encrypted dispatch instruction is sent to the dispatch switch system through the communication module. The specific process includes:

[0188] Connection establishment: ensuring that the communication link is established and active;

[0189] Data transmission: Send scheduling instruction data packets, following the sending order and rhythm of the communication protocol;

[0190] Confirmation of receipt: After receiving the dispatch instruction, the dispatch switch system sends a confirmation signal back to ensure that the dispatch instruction is delivered correctly;

[0191] If no confirmation signal is received within the scheduled time, the dispatch instruction is resent until confirmation or the retry limit is reached; the sending failure event is recorded and the operation and maintenance personnel are notified to perform manual intervention;

[0192] The dispatch switch system analyzes the dispatch instructions received, extracts the adjustment direction (increase or decrease) and the specific adjustment amount, and triggers the corresponding hardware operation according to the adjustment direction. If the capacity is increased, the backup generator is started or more power supply channels are opened. If the capacity is reduced, some power supply channels are closed or the generator output is reduced.

[0193] Monitor the capacity adjustment process in real time to ensure that operations are accurately executed according to scheduling instructions and remain stable during execution; feedback the execution results (such as whether the adjustment is successful, the actual adjustment amount, etc.) to the management system for subsequent optimization and adjustment.

[0194] This embodiment also provides a method for flexible scheduling of power supply in an intelligent computing center, including: establishing a communication connection between a scheduling switch system and a remote control center; collecting real-time load power data, and determining the flexible scheduling load by setting a safety margin coefficient; based on the real-time load power data, using a long short-term memory network model to predict the load power, and detecting load anomalies in real time; extracting static data segments and dynamic data segments, and determining the power supply regulation rate by calculating the ratio of static power to dynamic change rate; defining multi-objective optimization, and generating an optimized scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments; calculating the capacity adjustment amount, generating specific scheduling instructions in combination with the optimized scheduling strategy, and executing corresponding capacity adjustment operations according to the received scheduling instructions.

[0195] In summary, the present invention enables the control center to obtain the operating status and load data of the dispatching switch system in real time through communication connection, supports remote monitoring and management, significantly improves the response speed and operation convenience, realizes timely response to load fluctuations and effective management of adjustment space through real-time monitoring and dynamic calculation of load, avoids overload risks, improves energy utilization efficiency and stability, and can accurately predict future load change trends and optimize resource allocation strategies in advance through the combination of load power prediction and abnormality detection. At the same time, it can quickly identify and respond to abnormal load patterns to prevent potential system failures and energy consumption anomalies, which not only improves the level of intelligence, but also enhances its adaptability and emergency response capabilities to sudden load changes. By comprehensively analyzing the static and dynamic characteristics of the load, it ensures the scientificity and rationality of the adjustment rate, which not only optimizes energy distribution and improves the accuracy of power management, but also improves the overall efficiency and response capabilities, making power regulation more efficient and intelligent, and further enhancing the operational stability and energy utilization efficiency of the intelligent computing center.

[0196] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A power supply elastic scheduling system for an intelligent computing center, characterized in that: include, The communication module is responsible for establishing the communication connection between the dispatch switch system and the remote control center; The data processing module is responsible for collecting real-time load power data and determining the elastic scheduling load by setting the safety margin coefficient; The load prediction module is responsible for load power prediction based on real-time load power data using a long short-term memory network model and for detecting load anomalies in real time; The data segment extraction module is responsible for extracting the static data segment and the dynamic data segment, and determining the power supply regulation rate by calculating the ratio of the static power to the dynamic change rate; The multi-objective optimization module is responsible for defining multi-objective optimization and generating an optimal scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments; The scheduling instruction module is responsible for calculating the capacity adjustment amount, generating specific scheduling instructions in combination with the optimized scheduling strategy, and executing the corresponding capacity adjustment operation according to the received scheduling instructions; Collect real-time load power data and determine the elastic scheduling load by setting the safety margin coefficient, including the following steps: Deploy power sensors at key power supply points in the intelligent computing center to collect load power data in real time; The load power data collected by the power sensor is transmitted to the central data processing unit in real time; Format the collected load power data; According to the actual measurement results, determine the total load power capacity and calculate the real-time load power; Set the safety margin coefficient. Based on the real-time load power and the set safety margin coefficient, calculate the current load that can be flexibly dispatched. The expression is: ΔP(t)=λ(t)×P 总容量 ×M Among them, ΔP(t) is the flexibly dispatchable load, λ(t) is the real-time load power at time point t, and P 总容量 is the total load power capacity, M is the safety margin factor; Verify the flexibly dispatchable load and ensure that the calculated ΔP(t) does not exceed the preset safety margin limit; Define multi-objective optimization, generate an optimized scheduling strategy based on real-time load power, predicted load power, and static and dynamic data segments, including the following steps: Set minimizing energy consumption, minimizing response time, and maximizing stability as optimization goals; According to the defined optimization goal, construct the objective function and set the constraints; Combine the objective function and constraints to form the final optimization model; Select hybrid genetic algorithm and particle swarm optimization as the optimization algorithm; The real-time load power, predicted load power data, static data segment and average change rate of dynamic data segment are input into the optimization model and used as the input of GA-PSO algorithm; Run the GA-PSO algorithm to search for the optimal scheduling strategy; The optimal scheduling strategy is obtained through iterative optimization of the GA-PSO algorithm.

2. The power supply elastic scheduling system for the intelligent computing center according to claim 1, characterized in that: Based on real-time load power data, the long short-term memory network model is used to predict load power, including the following steps: Obtain the load power data of the past T minutes from the real-time database to form a time series data set; Perform data cleaning and normalization on the collected time series data set, and extract features for prediction from historical load power data; Select the long short-term memory network model LSTM which includes output layer, hidden layer and output layer; The mean square error is used as the loss function, and the Adam optimizer is used to optimize the model parameters; Input the prepared training data into the LSTM model for multiple rounds of iterative training. After each round of training, evaluate the performance of the model on the validation set and determine the optimal number of training rounds through the early stopping mechanism. Input the latest T-minute load power data into the trained LSTM model to obtain the predicted load power for the next Δt minutes.

3. The power supply elastic scheduling system for the intelligent computing center according to claim 2, characterized in that: Real-time detection of load anomalies includes the following steps: Select the Isolation Forest model, use the actual load power data and its related features as model input, and set the number of trees and subsample size; The Isolation Forest model is trained using historical load power data. Multiple decision trees are constructed by randomly selecting features and randomly splitting values. Each tree is designed to "isolate" a data point in the data. After training, the Isolation Forest model generates an anomaly score for each data point. The current load power data is input into the isolation forest model, and the isolation forest model outputs an anomaly score. The anomaly score is compared with a preset threshold to determine whether the load power data is abnormal. When load abnormality is detected, the early warning mechanism is triggered and corresponding emergency measures are implemented.

4. The power supply elastic scheduling system for the intelligent computing center as claimed in claim 3, characterized in that: Extracting static data segments and dynamic data segments includes the following steps: Extract scheduling instruction logs from the power management system to obtain time series data including load changes; Define the change segment D 变化 is the continuous time period during which the load power changes in the dispatch instruction; The acquired change segment D 变化 Perform segmentation processing and divide it into multiple independent change data segments; Apply linear regression analysis to each changing data segment, extract the intercept of each trend line, and form a static data segment set K; Collect time series data of actual load power from the power monitoring system; The slope of each time point is calculated using the difference method or moving regression method for each data point, and the slope is smoothed using a sliding window; All smoothed slopes are averaged to obtain the dynamic data segment.

5. The power supply elastic scheduling system for the intelligent computing center as claimed in claim 4, characterized in that: The power supply regulation rate is determined by calculating the ratio of static power to dynamic change rate, including the following steps: All intercepts in the static data segment set are averaged to obtain the average power of the static data segment; The average change rate R of the dynamic data segment is obtained by averaging all the smoothed slopes in the dynamic data segment. 动态 ; The average power P of the static data segment 静态 The average rate of change R of the dynamic data segment 动态 Calculates the power supply regulation rate.

6. The intelligent computing center power supply elastic scheduling system according to claim 5, characterized in that: Hybrid genetic algorithm and particle swarm optimization algorithm, as follows, Generate an initial population, assign velocity and position to each potential solution, and initialize the particle swarm; Calculate the fitness value of each individual according to the objective function, select individuals with higher fitness as parents, and perform crossover and mutation operations; Randomly select parent individuals for crossover to generate offspring individuals, and randomly mutate the offspring individuals to increase population diversity; Update the particle's speed and position according to the current speed, individual best position and global best position, and recalculate the fitness value according to the updated particle position; When the maximum number of iterations is reached or the fitness value converges to the preset threshold, the termination condition check; The individual with the highest fitness is selected from the population as the optimized scheduling strategy.

7. The power supply elastic scheduling system for the intelligent computing center according to claim 6, characterized in that: Calculate the capacity adjustment amount, generate specific scheduling instructions in combination with the optimized scheduling strategy, and execute the corresponding capacity adjustment operation according to the received scheduling instructions, including the following steps: The capacity adjustment ΔP is calculated by combining the flexibly dispatchable load ΔP(t), the power regulation rate η, and the coefficient S of the optimal scheduling strategy. 调节 , the expression is, ΔP 调节 =ΔP(t)×η×S When ΔP 调节 >0, an increase capacity instruction is generated; when ΔP 调节 When <0, a capacity reduction instruction is generated; Determine whether to generate a scheduling instruction to increase or decrease capacity according to the calculated capacity adjustment result; Construct the scheduling instruction content and record the generated scheduling instruction in the scheduling log.

8. A method for flexible scheduling of power supply in an intelligent computing center, based on the flexible scheduling system for power supply in an intelligent computing center according to any one of claims 1 to 7, characterized in that: include, Establish communication connection between dispatch switch system and remote control center; Collect real-time load power data and determine the flexible scheduling load by setting the safety margin coefficient; Based on real-time load power data, a long short-term memory network model is used to predict load power and detect load anomalies in real time; Extract static data segments and dynamic data segments, and determine the power supply regulation rate by calculating the ratio of static power to dynamic change rate; Define multi-objective optimization and generate optimal scheduling strategies based on real-time load power, predicted load power, and static and dynamic data segments; Calculate the capacity adjustment amount, generate specific scheduling instructions based on the optimized scheduling strategy, and execute corresponding capacity adjustment operations based on the received scheduling instructions.

Images