A full-tab capacitor control system and method for AI servers

Abstract

The present invention discloses a full-element capacitor control system and method for an AI server, relating to the field of data analysis technology. The system receives information content uploaded by users in a terminal system, extracts characteristic parameters of the information content and normalizes them to construct a characteristic vector, uses the characteristic vector and the operation level to construct a training sample and trains a random forest model to predict the operation level; sets a training cycle to form a training sample triple, constructs a dual-channel neural network fusion extraction feature based on LSTM and GRU, and adopts maximum likelihood estimation to optimize the model; generates a capacitor state training group and trains a capacitor state prediction model, constructs a digital twin system to simulate capacitor behavior, and outputs multi-dimensional operation indicators; collects the capacitor state vector of the target sampling period, and combines multiple candidate action vectors to predict the future capacitor state; screens characteristic actions according to thresholds, determines the final adjustment instruction, and improves the response speed.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and specifically to a full-tab capacitor control system and method for an AI server. Background Art

[0002] Currently, AI servers experience dramatic power consumption fluctuations and high uncertainty when executing large-scale parallel computing tasks, placing higher demands on the dynamic response capabilities of power supply systems. As the scale of AI computing continues to expand, full-element capacitors have been widely used in high-performance power supply voltage regulation due to their lower equivalent series resistance (ESR) and faster charge and discharge speeds. However, in actual applications, traditional capacitor array configurations often use preset parameters and cannot dynamically adjust according to load changes, making it difficult to meet the frequently changing current demands of current AI computing. Furthermore, they lack the ability to predict AI task loads in real time, often resulting in problems such as delayed capacitor response and inappropriate activation timing, which reduces energy efficiency and even causes performance fluctuations.

[0003] To improve the operating efficiency and stability of the power supply system, through real-time monitoring and analysis of AI computing tasks, the system can predict load change trends and adjust capacitor configurations in advance, thereby optimizing the energy supply path, improving response speed, reducing energy consumption, and avoiding the impact of sudden failures on system operation, thereby significantly improving the safety and stability of the server power supply system and promoting the efficient use of computing resources. Summary of the Invention

[0004] The purpose of the present invention is to provide a full-tab capacitor control system and method for AI servers to solve the problems raised in the prior art.

[0005] In order to solve the above technical problems, the present invention provides the following technical solution: a full-tab capacitor control method for an AI server, the method comprising:

[0006] Step S100: Receive user-uploaded information in the AI ​​server terminal system, set a sampling period, collect server power parameters and record capacitor adjustment parameters, extract feature parameters of the information content and normalize them to construct a feature vector, calculate an operation score, divide the operation level, use the feature vector and operation level to construct a training sample, and train a random forest model to predict the operation level;

[0007] Step S200: Set a training cycle, extract information content feature vectors and input them into the operation level prediction model, combine code analysis tools to generate structural context vectors, build an operation state matrix, calculate future power, form training sample triplets, build a dual-channel neural network based on LSTM and GRU to extract features, and predict future power through a hybrid density network, using maximum likelihood estimation to optimize the model;

[0008] Step S300: Collect the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combine the open state to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, build a digital twin system to simulate capacitor behavior, and output multi-dimensional operation indicators;

[0009] Step S400: extracting content features of user-uploaded information in real time, constructing a state matrix, structural context, and operating level labels, and inputting them into a prediction model to obtain future power values;

[0010] Step S500: collecting the capacitance state vector of the target sampling period, combining multiple candidate action vectors to predict the future capacitance state, calculating the cost score, and sorting the candidate action vectors according to the cost score to obtain a set of candidate action vectors;

[0011] Step S600: Input the real-time capacitor state, power value, predicted power and each candidate action vector into the digital twin system simulation, obtain the prediction index, filter the characteristic action according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

[0012] Furthermore, step S100 includes:

[0013] Step S101: In the terminal system of the AI ​​server, information content uploaded by the user is received, a sampling period is preset, power parameters of the server during the sampling period are monitored, capacitor adjustment parameters are recorded, and a sampling record for executing the information content is generated;

[0014] Step S102: Collect a piece of information uploaded by the user, extract the content type of the information, determine the characteristic parameters corresponding to the content type, calculate the characteristic value of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector A=(a1,a2,...ab), where a1,a2,...ab represent the normalized values ​​of the 1st, 2nd,...bth characteristic parameters respectively;

[0015] Step S103: In the historical sampling record set of the information content, each power response data of a certain historical sampling record is collected and normalized, and an operation score is calculated according to the following formula:

[0016] ;

[0017] Among them, B represents the running score, D c Expressed as the normalized value of the cth power response data, E cExpressed as the weight of the cth power response data, all historical sampling records are summarized and the average value of the running score is calculated;

[0018] Step S104: Preset a number of operation levels, set the average range of operation scores corresponding to each operation level, and automatically classify each piece of information into the corresponding operation level;

[0019] Step S105: In a set of information content of a certain content type, a content feature vector and an operation level of each information content are obtained, the operation level is used as a label, and a training sample set is constructed together with the content feature vector. The training sample set is input into a random forest model for training, with the content feature vector as input and the operation level label as output, to establish an operation level prediction model corresponding to the content type;

[0020] Power parameters include voltage, current, power, and energy consumption. Content types include text, image, and video. Text feature parameters include total word count, keyword density, grammatical complexity, and semantic vectors. Image feature parameters include image size, color histogram distribution, texture features, and image clarity. Video feature parameters include video duration, frame rate, resolution, and scene switching frequency. Power response data includes average power consumption, power fluctuation rate, and capacitor call frequency.

[0021] By collecting power parameters, content characteristics, and historical response data from the AI ​​server during a sampling period, this method can automatically analyze the operational status of the information content. Using normalization processing and a scoring mechanism, the evaluation process is highly standardized and consistent, avoiding human error.

[0022] By combining content feature vectors and operational level labels into training samples and inputting them into a random forest model for training, the system can automatically learn and predict the operational level of future uploaded content. Random forests have good generalization and anti-overfitting capabilities, making the prediction results highly robust and accurate.

[0023] It has strong adaptability in extracting content types and identifying corresponding feature parameters, and can be flexibly applied to various types of information content processing scenarios, such as text, images, videos, etc.

[0024] Furthermore, step S200 includes:

[0025] Step S201: Selecting a number of consecutive days as a training cycle, during which a piece of information uploaded by a user is collected, the content type of the information is extracted, a content feature vector of the information is constructed, and the feature vector is input into an operation level prediction model corresponding to the content type to obtain an operation level label for the information content;

[0026] Step S202: extracting the task computing density and memory bandwidth requirements of the information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the information content;

[0027] Step S203: Acquire a set of sampled records that execute the information content, collect the current mean value, power change rate, and load fluctuation coefficient in each sampled record to form an operation state vector of the sampled record, and arrange all the operation state vectors in chronological order of sampling time to form an operation state matrix of the information content;

[0028] Step S204: When the information content is executed, the number of future sampling periods is preset, the power value corresponding to each future sampling period is collected, and the future power is calculated according to the following formula:

[0029] ;

[0030] Among them, F represents the future power, G e is represented by the power value corresponding to the e-th future sampling period, f is represented by the number of future sampling periods, and the operating state matrix, structural context vector, operating level label and future power of the information content are combined to construct a training sample triplet S=[Y,(J,X),F], where Y is represented by the operating state matrix, J is represented by the structural context vector, X is represented by the operating level label, and F is represented by the future power;

[0031] Step S205: Based on the training sample triples, a dual-channel neural network model for future power prediction is constructed. The operating state matrix in the training sample triples is used as the first channel. The state vector representing the operating sequence fluctuation characteristics is extracted through the LSTM network. The structural context vector and the operating level label in the training sample triples are used as the second channel. The structural semantic feature vector is extracted through the GRU network. The state vector and the structural semantic feature vector are concatenated and fused to generate a joint representation vector. The future power is output using a mixed density network (MDN) structure.

[0032] Step S206: Obtain all training sample triplets within the training cycle, use maximum likelihood estimation as the objective function, calculate the negative log-likelihood loss between the predicted future power distribution and the actual future power value, perform model training, and obtain a prediction model for future power;

[0033] A dual-channel neural network architecture is used: the first channel uses LSTM to focus on extracting temporal fluctuation characteristics from the operating status matrix; the second channel uses GRU to model the semantic features between the code structure context and the operating level. The combined representation vector formed by fusing the two channels can better capture underlying patterns and significantly improve future power modeling capabilities.

[0034] By adopting a hybrid density network, the model no longer outputs a single power prediction value, but instead outputs a probability distribution of power values. This enhances the model's performance in multi-peak prediction and nonlinear uncertainty modeling scenarios, and is more suitable for scenarios where AI server loads are complex and power demand is volatile and uncertain.

[0035] Furthermore, step S300 includes:

[0036] Step S301: collecting state variables of each capacitor unit in the terminal system corresponding to a time point of a certain sampling record, wherein the state variables include equivalent series resistance, state of charge, and temperature, to form a capacitor state vector;

[0037] Step S302: obtaining the on-state of each capacitor unit, where the on-state includes enabled and disabled, and combining the on-states of each capacitor unit to construct an action vector;

[0038] Step S303: combining the capacitance state vector and action vector of a certain sampling record with the capacitance state vector of the next adjacent sampling record to form a capacitance state training group. Inputting all capacitance state training groups into the neural network model for training to obtain a capacitance state prediction model.

[0039] Step S304: Establishing a digital twin system based on the set of sampled records within the training cycle, inputting the capacitor state vector, action vector, current power value, and predicted power value corresponding to a certain sampled record into the digital twin system, performing simulation operations in the digital twin system, and obtaining output results including voltage output, power output, number of voltage drops, response delay, and capacitor temperature;

[0040] By combining the capacitor state vector and action vector at each sampling point and combining it with the capacitor state vector at the next sampling point, high-quality training samples are constructed. These effectively reflect the state evolution of the capacitor unit under different control strategies, helping to improve the prediction model's ability to fit capacitor behavior.

[0041] The obtained capacitor state prediction model can predict the state at the next moment based on the current state and operation, enabling the system to predict future state change trends, providing forward-looking support for the control system and reducing response lag;

[0042] The digital twin system not only reflects the state of the physical system, but also accepts real-time input to generate voltage output, power output, number of voltage drops, response delay, and capacitor temperature, enabling offline evaluation and optimization of action strategies.

[0043] Furthermore, step S400 includes:

[0044] Step S401: obtaining real-time information content uploaded by a user terminal, extracting the content type of the real-time information content, constructing a content feature vector of the real-time information content, and inputting the vector into an operation level prediction model corresponding to the content type to obtain an operation level label of the real-time information content;

[0045] Step S402: extracting the task computing intensity and memory bandwidth requirement of the real-time information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the real-time information content;

[0046] Step S403: The number of sampling periods for analysis and optimization is preset to n, a set of sampling periods of real-time information content is obtained, and an operation status matrix of the real-time information content is constructed;

[0047] Step S404: Inputting the operation state matrix, structural context vector, and operation level label of the real-time information content into the future power prediction model to obtain the predicted value of the future power;

[0048] Generate operating level labels based on the type of real-time information content, generate structural context vectors based on static code structure analysis, capture the real-time nature, complexity and resource dependency of the current task, and effectively improve the power prediction model's ability to recognize short-term behavioral fluctuations through structured representation vectors.

[0049] Furthermore, step S500 includes:

[0050] Step S501: obtaining a time point corresponding to the nth sampling period of the analysis and optimization, collecting the real-time state variables of each capacitor unit of the terminal system corresponding to the time point, and obtaining a real-time capacitor state vector;

[0051] Step S502: Preset a number of candidate motion vectors, input the real-time capacitance state vector and each candidate motion vector into a capacitance state prediction model, and predict the next capacitance state vector;

[0052] Step S503: extract the equivalent series resistance R and the state of charge H in a predicted next capacitance state vector, and calculate the cost score according to the following formula:

[0053] ;

[0054] Where P represents the cost score, R0 represents the standard value of equivalent series resistance, r1 and r2 represent the weights of equivalent series resistance and state of charge, respectively;

[0055] Step S504: Summarize the cost scores corresponding to all candidate action vectors, and arrange all candidate action vectors from low to high according to the cost scores to obtain a set of candidate action vectors;

[0056] By simulating candidate actions and predicting states, the system can predict in advance the specific impact of each control action on the future state of the capacitor. This is then quantitatively evaluated using a cost-scoring function, enabling an "evaluate first, then execute" strategy. This replaces the traditional "experience-based activation" model, significantly improving the robustness and foresight of the control strategy and reducing performance degradation or risk events caused by activating the wrong capacitor unit.

[0057] Based on the trained capacitor state prediction model, the equivalent series resistance and state of charge of the capacitor at the next moment can be predicted based on the current state and action combination. This improves the modeling ability of the capacitor's "behavioral evolution" and helps prevent high-temperature, low-battery, or high-impedance capacitor units from continuously participating in operation.

[0058] By presetting multiple candidate action vectors for parallel prediction and scoring, the optimal control strategy under the current conditions can be found.

[0059] Furthermore, step S600 includes:

[0060] Step S601: The real-time capacitor state vector, the real-time current power value, and the predicted value of future power at the time point corresponding to the nth sampling period are combined with each candidate action vector in the candidate action vector set, and input into the digital twin system. A simulation operation is performed to obtain the predicted number of voltage drops, predicted response delay, and predicted capacitor temperature corresponding to each candidate action vector.

[0061] Step S602: Preset a response delay threshold and a capacitor temperature threshold. If a candidate action vector satisfies all of the following conditions: the predicted number of voltage drops is 0; the predicted response delay is less than the response delay threshold; and the predicted capacitor temperature is less than the capacitor temperature threshold, then the candidate action vector is set as a feature action vector.

[0062] Step S603: extracting the position order of each characteristic action vector in the candidate action vector set, selecting the characteristic action vector with the smallest position order as the final adjustment instruction, and inputting it into the terminal system to adjust the capacitor unit;

[0063] The real-time capacitor state, current power, and predicted power values ​​are combined as inputs. Simulation is performed within the digital twin system, combining each candidate action to output predicted voltage drop, response delay, and capacitor temperature. This mechanism establishes a complete prediction chain from "current state → behavioral decision → system response," simulating the impact of control actions on the future system and enabling true "predictive control strategy screening."

[0064] Setting three conditions: voltage dips = 0 to ensure power supply continuity; response delay < threshold to ensure response speed; capacitor temperature < threshold to avoid thermal runaway; significantly enhancing operational stability and safety;

[0065] Among multiple qualified feature actions, the action vector with the smallest position order is selected first. The "two-stage screening" mechanism embodies: the first stage: security feasibility assessment; the second stage: performance optimization; ensuring that the selected strategy is not only safe, but also optimal in terms of performance and energy efficiency.

[0066] To better implement the above method, a full-tab capacitor control system for AI servers is proposed. The system includes a predicted operating level module, a future power module, a capacitor state module, a real-time prediction module, a candidate action vector module, and a real-time adjustment module.

[0067] Operation level prediction module: Receives user-uploaded information in the AI ​​server terminal system, sets a sampling period, collects server power parameters and records capacitor adjustment parameters, extracts feature parameters from the information and normalizes them to construct a feature vector, calculates the operation score, classifies the operation level, uses the feature vector and operation level to construct training samples, and trains a random forest model to predict the operation level.

[0068] Future Power Module: This module sets a training cycle, extracts information content feature vectors, and inputs them into the operating level prediction model. It uses code analysis tools to generate a structural context vector, constructs an operating state matrix, calculates future power, and forms training sample triplets. A dual-channel neural network based on LSTM and GRU is constructed to fuse and extract features. Future power is predicted using a hybrid density network, and the model is optimized using maximum likelihood estimation.

[0069] Capacitor state module: This module collects the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combines the on-state to form an action vector, generates a capacitor state training set, trains a capacitor state prediction model, builds a digital twin system to simulate capacitor behavior, and outputs multi-dimensional operating indicators.

[0070] Real-time prediction module: extracts content features of user-uploaded information in real time, constructs a state matrix, structural context, and operating level labels, and inputs them into the prediction model to obtain future power values;

[0071] Candidate action vector module: collects the target sampling period capacitor state vector, combines multiple candidate action vectors to predict the future capacitor state, calculates the cost score, and sorts the candidate action vectors according to the cost score to obtain a candidate action vector set;

[0072] Real-time adjustment module: The real-time capacitor state, power value, predicted power and candidate action vectors are input into the digital twin system simulation to obtain the prediction index, filter the characteristic actions according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

[0073] Furthermore, the operation level prediction module includes an operation score calculation unit and an operation level prediction model establishment unit:

[0074] Calculation and operation scoring unit: In the terminal system of the AI ​​server, the unit receives information content uploaded by the user terminal, presets a sampling period, monitors the power parameters of the server during the sampling period, records the capacitor adjustment parameters, generates a sampling record for executing the information content, collects a certain information content uploaded by the user terminal, extracts the content type of the information content, determines the characteristic parameters corresponding to the content type, calculates the characteristic value of each characteristic parameter in the query content, performs normalization processing, constructs a content feature vector, collects each power response data of a certain historical sampling record in the historical sampling record set of the information content, performs normalization processing, and calculates the operation score;

[0075] Establish an operation level prediction model unit: preset several operation levels, and set the average operation score range corresponding to each operation level, automatically divide each information content into the corresponding operation level, obtain the content feature vector and operation level of each information content in a certain content type of information content set, use the operation level as a label, and jointly construct a training sample set with the content feature vector, input the training sample set into the random forest model for training, wherein the content feature vector is used as input and the operation level label is used as output, and establish an operation level prediction model corresponding to the content type.

[0076] Furthermore, the future power module includes a unit for constructing a training sample triplet and a unit for establishing a future power prediction model:

[0077] Constructing a training sample triplet unit: Selecting a number of consecutive days as a training cycle, within the training cycle, collecting a piece of information content uploaded by a user terminal, extracting the content type of the information content, constructing a content feature vector for the information content, and inputting it into an operation level prediction model corresponding to the content type to obtain an operation level label for the information content. Using a static code analysis tool built into the terminal system, extracting the task computational intensity and memory bandwidth requirements of the information content, and performing normalization processing to construct a structural context vector for the information content. Obtaining a set of sampling records for executing the information content, collecting the current mean value, power change rate, and load fluctuation coefficient in each sampling record to form an operation state vector for the sampling record, arranging all the operation state vectors in chronological order of sampling time to form an operation state matrix for the information content. When the information content is executed, presetting the number of future sampling cycles, collecting the power value corresponding to each future sampling cycle, calculating the future power, and combining the operation state matrix, structural context vector, operation level label, and future power of the information content to construct a training sample triplet.

[0078] Establish a future power prediction model unit: based on the training sample triples, construct a dual-channel neural network model for future power prediction, use the operating state matrix in the training sample triples as the first channel, extract the state vector representing the operating sequence fluctuation characteristics through the LSTM network, use the structural context vector and the operating level label in the training sample triples as the second channel, extract the structural semantic feature vector through the GRU network, concatenate and fuse the state vector and the structural semantic feature vector to generate a joint representation vector, use a mixed density network MDN structure to output the future power, obtain all training sample triplets within the training cycle, use maximum likelihood estimation as the objective function, calculate the negative log-likelihood loss between the predicted future power distribution and the actual future power value, perform model training, and obtain a prediction model for future power.

[0079] Compared with the existing technology, the present invention has the following advantages: it introduces a neural network prediction model to predict the evolution of capacitor state in advance, combines future power prediction with simulation feedback, and implements a "prediction-simulation-optimization-execution" closed-loop control, thereby improving the accuracy and response foresight of capacitor control decisions, avoiding strategy lag and resource waste;

[0080] By introducing dimensions such as task content semantic features, static structure context features, and operating state matrices, and through dual-channel deep network fusion processing, it can mine the power consumption behavior of complex tasks, support the dynamic changes of different types of tasks and loads in AI servers, and enhance generalization and adaptability.

[0081] Build a capacitor digital twin system, take state, power and candidate control actions as input, output key indicators, and accurately simulate "the impact of a certain control action on the capacitor system in the future", providing visual, quantitative and reliable prediction support for decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flow chart of a method for controlling full-tab capacitors for AI servers according to the present invention;

[0083] Figure 2 This is a structural schematic diagram of a full-tab capacitor control system for an AI server according to the present invention. DETAILED DESCRIPTION

[0084] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0085] See also Figure 1 The present invention provides a technical solution: a full-tab capacitor control method for an AI server, the method comprising:

[0086] Step S100: Receive user-uploaded information in the AI ​​server terminal system, set a sampling period, collect server power parameters and record capacitor adjustment parameters, extract feature parameters of the information content and normalize them to construct a feature vector, calculate an operation score, divide the operation level, use the feature vector and operation level to construct a training sample, and train a random forest model to predict the operation level;

[0087] Wherein, step S100 includes:

[0088] Step S101: In the terminal system of the AI ​​server, information content uploaded by the user is received, a sampling period is preset, power parameters of the server during the sampling period are monitored, capacitor adjustment parameters are recorded, and a sampling record for executing the information content is generated;

[0089] Step S102: Collect a piece of information uploaded by the user, extract the content type of the information, determine the characteristic parameters corresponding to the content type, calculate the characteristic value of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector A=(a1,a2,...ab), where a1,a2,...ab represent the normalized values ​​of the 1st, 2nd,...bth characteristic parameters respectively;

[0090] Step S103: In the historical sampling record set of the information content, each power response data of a certain historical sampling record is collected and normalized, and an operation score is calculated according to the following formula:

[0091] ;

[0092] Among them, B represents the running score, D c Expressed as the normalized value of the cth power response data, E c Expressed as the weight of the cth power response data, all historical sampling records are summarized and the average value of the running score is calculated;

[0093] Step S104: Preset a number of operation levels, set the average range of operation scores corresponding to each operation level, and automatically classify each piece of information into the corresponding operation level;

[0094] Step S105: In a set of information content of a certain content type, a content feature vector and an operation level of each information content are obtained, the operation level is used as a label, and a training sample set is constructed together with the content feature vector. The training sample set is input into a random forest model for training, with the content feature vector as input and the operation level label as output, to establish an operation level prediction model corresponding to the content type;

[0095] For example, if the content type of information A is image, the parameters for extracting the information content are as follows: image size: 2MB; resolution: 1920×1080; number of channels: 3 (RGB); selected model: ResNet50; task label: image_classification; after normalization, the constructed content feature vector is (0.2, 0.23, 0.75, 0.66, 0.33);

[0096] In the b-th sampling record, the normalized power response data is as follows: the average power is 0.7, the weight is 0.4; the fluctuation coefficient is 0.12, the weight is 0.3; the peak power is 0.9, the weight is 0.3, and the calculated operation score is 0.586;

[0097] Assume that multiple task records collect feature vectors and run level labels:

[0098] The feature vector of the first task record is (0.2, 0.23, 0.75, 0.66, 0.33), and the operating level is L3; the feature vector of the second task record is (0.1, 0.15, 0.5, 0.33, 0.33), and the operating level is L2; ​​the feature vector of the third task record is (0.4, 0.45, 1.0, 1.0, 0.66), and the operating level is L4. Use these samples to train the random forest model. Input: content feature vector; Output: operating level label.

[0099] Step S200: Set a training cycle, extract information content feature vectors and input them into the operation level prediction model, combine code analysis tools to generate structural context vectors, build an operation state matrix, calculate future power, form training sample triplets, build a dual-channel neural network based on LSTM and GRU to extract features, and predict future power through a hybrid density network, using maximum likelihood estimation to optimize the model;

[0100] Wherein, step S200 includes:

[0101] Step S201: Selecting a number of consecutive days as a training cycle, during which a piece of information uploaded by a user is collected, the content type of the information is extracted, a content feature vector of the information is constructed, and the feature vector is input into an operation level prediction model corresponding to the content type to obtain an operation level label for the information content;

[0102] Step S202: extracting the task computing density and memory bandwidth requirements of the information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the information content;

[0103] Step S203: Acquire a set of sampled records that execute the information content, collect the current mean value, power change rate, and load fluctuation coefficient in each sampled record to form an operation state vector of the sampled record, and arrange all the operation state vectors in chronological order of sampling time to form an operation state matrix of the information content;

[0104] Step S204: When the information content is executed, the number of future sampling periods is preset, the power value corresponding to each future sampling period is collected, and the future power is calculated according to the following formula:

[0105] ;

[0106] Among them, F represents the future power, G eis represented by the power value corresponding to the e-th future sampling period, f is represented by the number of future sampling periods, and the operating state matrix, structural context vector, operating level label and future power of the information content are combined to construct a training sample triplet S=[Y,(J,X),F], where Y is represented by the operating state matrix, J is represented by the structural context vector, X is represented by the operating level label, and F is represented by the future power;

[0107] Step S205: Based on the training sample triples, a dual-channel neural network model for future power prediction is constructed. The operating state matrix in the training sample triples is used as the first channel. The state vector representing the operating sequence fluctuation characteristics is extracted through the LSTM network. The structural context vector and the operating level label in the training sample triples are used as the second channel. The structural semantic feature vector is extracted through the GRU network. The state vector and the structural semantic feature vector are concatenated and fused to generate a joint representation vector. The future power is output using a mixed density network (MDN) structure.

[0108] Step S206: Obtain all training sample triplets within the training cycle, use maximum likelihood estimation as the objective function, calculate the negative log-likelihood loss between the predicted future power distribution and the actual future power value, perform model training, and obtain a prediction model for future power;

[0109] For example, a user uploads an image task, and the system extracts a feature vector of (0.2, 0.23, 0.75, 0.66, 0.33). This is input into the operating level prediction model corresponding to the "image recognition class," and the model outputs an operating level label of L3.

[0110] The system call static code analysis tool analyzes model calls and resource distribution: the normalized computational density is 0.82, the normalized memory bandwidth requirement is 0.74, and the structure context vector J = (0.82, 0.74) is constructed;

[0111] Assume that the sampling records during the image recognition task are as shown in Table 1;

[0112]

[0113] Table 1

[0114] Each line constitutes an operating state vector, forming an operating state matrix;

[0115] Assume that the power sampled in the next five cycles is: 65.3W, 67.8W, 70.1W, 71.0W, 69.5W, calculate the future power to be 68.74, and construct the training sample triplet;

[0116] First channel: Input: running state matrix Y, network structure: LSTM, extracting running fluctuation features; output: state vector;

[0117] Second channel: Input: structural context vector J + operation level label X; Network structure: GRU network, extracting context semantic representation; Output: structural feature vector;

[0118] Concatenate the state vector and the structural feature vector, input the mixture density network, and output multiple Gaussian distribution parameters, including mean, variance, and weight;

[0119] The Adam optimizer is used for training using the negative log-likelihood loss function, and the weights of the LSTM / GRU / MDN joint model are adjusted to obtain the prediction function.

[0120] Step S300: Collect the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combine the open state to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, build a digital twin system to simulate capacitor behavior, and output multi-dimensional operation indicators;

[0121] Wherein, step S300 includes:

[0122] Step S301: collecting state variables of each capacitor unit in the terminal system corresponding to a time point of a certain sampling record, wherein the state variables include equivalent series resistance, state of charge, and temperature, to form a capacitor state vector;

[0123] Step S302: obtaining the on-state of each capacitor unit, where the on-state includes enabled and disabled, and combining the on-states of each capacitor unit to construct an action vector;

[0124] Step S303: combining the capacitance state vector and action vector of a certain sampling record with the capacitance state vector of the next adjacent sampling record to form a capacitance state training group. Inputting all capacitance state training groups into the neural network model for training to obtain a capacitance state prediction model.

[0125] Step S304: Establishing a digital twin system based on the set of sampled records within the training cycle, inputting the capacitor state vector, action vector, current power value, and predicted power value corresponding to a certain sampled record into the digital twin system, performing simulation operations in the digital twin system, and obtaining output results including voltage output, power output, number of voltage drops, response delay, and capacitor temperature;

[0126] For example, there are three capacitor units C1, C2, and C3. At sampling time T1, the following state parameters of each capacitor unit are collected: C1's equivalent series resistance is 0.12, state of charge is 0.65, and temperature is 48.2; C2's equivalent series resistance is 0.15, state of charge is 0.62, and temperature is 50.1; C3's equivalent series resistance is 0.11, state of charge is 0.68, and temperature is 46.9. The capacitor state vector is constructed as Lt(0.12, 0.65, 48.2, 0.15, 0.62, 50.1, 0.11, 0.68, 46.9).

[0127] Current action: C1 and C3 are enabled, C2 is disabled; the action code is enabled as 1, disabled as 0, and the action vector is Ut(1,0,1);

[0128] At the adjacent time point T2, the following states are collected: C1 has an equivalent series resistance of 0.13, a state of charge of 0.64, and a temperature of 48.7; C2 has an equivalent series resistance of 0.15, a state of charge of 0.61, and a temperature of 50.3; C3 has an equivalent series resistance of 0.12, a state of charge of 0.67, and a temperature of 47.1. The capacitance state vector Lt+1(0.13, 0.64, 48.7, 0.15, 0.61, 50.3, 0.12, 0.67, 47.1) is constructed.

[0129] Taking [Lt, Ut] as input and Lt+1 as output, a training sample is constructed. Multiple sampling records are aggregated to form a training set, and a neural network is trained as a capacitance state prediction model.

[0130] Step S400: extracting content features of user-uploaded information in real time, constructing a state matrix, structural context, and operating level labels, and inputting them into a prediction model to obtain future power values;

[0131] Wherein, step S400 includes:

[0132] Step S401: obtaining real-time information content uploaded by a user terminal, extracting the content type of the real-time information content, constructing a content feature vector of the real-time information content, and inputting the vector into an operation level prediction model corresponding to the content type to obtain an operation level label of the real-time information content;

[0133] Step S402: extracting the task computing intensity and memory bandwidth requirement of the real-time information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the real-time information content;

[0134] Step S403: The number of sampling periods for analysis and optimization is preset to n, a set of sampling periods of real-time information content is obtained, and an operation status matrix of the real-time information content is constructed;

[0135] Step S404: Inputting the operation state matrix, structural context vector, and operation level label of the real-time information content into the future power prediction model to obtain the predicted value of the future power;

[0136] For example, a user uploads an image recognition task with the following real-time content information: image size: 3.5MB; resolution: 2560×1440; number of channels: 3 (RGB); model type: EfficientNet-B3; task type: image_classification; construct a feature vector (0.35, 0.43, 0.75, 0.8, 0.33) and input it into the "image recognition class" operation level prediction model, outputting: operation level L4;

[0137] Static code analysis results: normalized computational density: 0.92; normalized memory bandwidth requirement: 0.85; structural context vector J = (0.92, 0.85);

[0138] Assume that the operating status data in the first 5 cycles is collected as Table 2

[0139]

[0140] Table 2

[0141] Build an operational status matrix;

[0142] The operation state matrix, structure context vector, and operation level label are input into the future power prediction model, and the predicted value of the future power is 84.6W.

[0143] Step S500: collecting the capacitance state vector of the target sampling period, combining multiple candidate action vectors to predict the future capacitance state, calculating the cost score, and sorting the candidate action vectors according to the cost score to obtain a set of candidate action vectors;

[0144] Wherein, step S500 includes:

[0145] Step S501: obtaining a time point corresponding to the nth sampling period of the analysis and optimization, collecting the real-time state variables of each capacitor unit of the terminal system corresponding to the time point, and obtaining a real-time capacitor state vector;

[0146] Step S502: Preset a number of candidate motion vectors, input the real-time capacitance state vector and each candidate motion vector into a capacitance state prediction model, and predict the next capacitance state vector;

[0147] Step S503: extract the equivalent series resistance R and the state of charge H in a predicted next capacitance state vector, and calculate the cost score according to the following formula:

[0148] ;

[0149] Where P represents the cost score, R0 represents the standard value of equivalent series resistance, r1 and r2 represent the weights of equivalent series resistance and state of charge, respectively;

[0150] Step S504: Summarize the cost scores corresponding to all candidate action vectors, and arrange all candidate action vectors from low to high according to the cost scores to obtain a set of candidate action vectors;

[0151] For example, at sampling time T5, the following state parameters of each capacitor unit are collected: C1 equivalent series resistance is 0.12, state of charge is 0.66, and temperature is 48.5; C2 equivalent series resistance is 0.14, state of charge is 0.64, and temperature is 49.8; C3 equivalent series resistance is 0.13, state of charge is 0.67, and temperature is 47.2; the capacitor state vector is constructed as Lt5(0.12, 0.66, 48.5, 0.14, 0.64, 49.8, 0.13, 0.67, 47.2);

[0152] Preset candidate motion vectors U1=(1, 0, 1); U2=(1, 1, 0); U3=(0, 1, 1); U4=(1, 1, 1);

[0153] Each candidate action vector is concatenated with Lt5 and input into the capacitance state prediction model to predict the capacitance state vector of the next cycle. The next capacitance state vector corresponding to U1 is (0.13, 0.65, 48.9, 0.14, 0.63, 50.1, 0.14, 0.66, 47.6); the next capacitance state vector corresponding to U2 is (0.13, 0.65, 48.8, 0.15, 0.62, 50.3, 0.13, 0.67, 47.1); the next capacitance state vector corresponding to U3 is (0.12, 0.64, 48.6, 0.14, 0.63, 50.0, 0.15, 0.65, 47.7); the next capacitance state vector corresponding to U4 is (0.13, 0.66, 49.1, 0.15, 0.63, 50.2, 0.14,0.66,47.8);

[0154] That is, the cost score of U1 is 0.191; the cost score of U2 is 0.226; the cost score of U3 is 0.254; the cost score of U4 is 0.224, and the set of candidate action vectors is (U1, U4, U2, U3).

[0155] Step S600: Input the real-time capacitor state, power value, predicted power and each candidate action vector into the digital twin system simulation, obtain the prediction index, filter the characteristic action according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction

[0156] Step S600 includes:

[0157] Step S601: The real-time capacitor state vector, the real-time current power value, and the predicted value of future power at the time point corresponding to the nth sampling period are combined with each candidate action vector in the candidate action vector set, and input into the digital twin system. A simulation operation is performed to obtain the predicted number of voltage drops, predicted response delay, and predicted capacitor temperature corresponding to each candidate action vector.

[0158] Step S602: Preset a response delay threshold and a capacitor temperature threshold. If a candidate action vector satisfies all of the following conditions: the predicted number of voltage drops is 0; the predicted response delay is less than the response delay threshold; and the predicted capacitor temperature is less than the capacitor temperature threshold, then the candidate action vector is set as a feature action vector.

[0159] Step S603: extracting the position order of each characteristic action vector in the candidate action vector set, selecting the characteristic action vector with the smallest position order as the final adjustment instruction, and inputting it into the terminal system to adjust the capacitor unit;

[0160] For example, Lt5, the current power value, and the predicted value of future power are combined with each candidate action vector and input into the digital twin system for simulation, and the following prediction indicators are obtained as Table 3;

[0161]

[0162] Table 3

[0163] The preset response delay threshold is 0.20s, the capacitor temperature threshold is 50.0, U1 satisfies all adjustments, and the final adjustment instruction is U1=(1, 0, 1).

[0164] In order to better implement the above method, a full-tab capacitor control system for AI servers is also proposed. The system includes a predicted operating level module, a future power module, a capacitor state module, a real-time prediction module, a candidate action vector module, and a real-time adjustment module.

[0165] Operation level prediction module: Receives user-uploaded information in the AI ​​server terminal system, sets a sampling period, collects server power parameters and records capacitor adjustment parameters, extracts feature parameters from the information and normalizes them to construct a feature vector, calculates the operation score, classifies the operation level, uses the feature vector and operation level to construct training samples, and trains a random forest model to predict the operation level.

[0166] The operation level prediction module includes an operation score calculation unit and an operation level prediction model establishment unit:

[0167] Calculation and operation scoring unit: In the terminal system of the AI ​​server, the unit receives information content uploaded by the user terminal, presets a sampling period, monitors the power parameters of the server during the sampling period, records the capacitor adjustment parameters, generates a sampling record for executing the information content, collects a certain information content uploaded by the user terminal, extracts the content type of the information content, determines the characteristic parameters corresponding to the content type, calculates the characteristic value of each characteristic parameter in the query content, performs normalization processing, constructs a content feature vector, collects each power response data of a certain historical sampling record in the historical sampling record set of the information content, performs normalization processing, and calculates the operation score;

[0168] Establish an operation level prediction model unit: preset several operation levels, and set the average operation score range corresponding to each operation level, automatically divide each information content into the corresponding operation level, obtain the content feature vector and operation level of each information content in a certain content type of information content set, use the operation level as a label, and jointly construct a training sample set with the content feature vector, input the training sample set into the random forest model for training, wherein the content feature vector is used as input and the operation level label is used as output, and establish an operation level prediction model corresponding to the content type.

[0169] Future Power Module: This module sets a training cycle, extracts information content feature vectors, and inputs them into the operating level prediction model. It uses code analysis tools to generate a structural context vector, constructs an operating state matrix, calculates future power, and forms training sample triplets. A dual-channel neural network based on LSTM and GRU is constructed to fuse and extract features. Future power is predicted using a hybrid density network, and the model is optimized using maximum likelihood estimation.

[0170] The future power module includes a unit for constructing a training sample triplet and a unit for establishing a future power prediction model:

[0171] Constructing a training sample triplet unit: Selecting a number of consecutive days as a training cycle, within the training cycle, collecting a piece of information content uploaded by a user terminal, extracting the content type of the information content, constructing a content feature vector for the information content, and inputting it into an operation level prediction model corresponding to the content type to obtain an operation level label for the information content. Using a static code analysis tool built into the terminal system, extracting the task computational intensity and memory bandwidth requirements of the information content, and performing normalization processing to construct a structural context vector for the information content. Obtaining a set of sampling records for executing the information content, collecting the current mean value, power change rate, and load fluctuation coefficient in each sampling record to form an operation state vector for the sampling record, arranging all the operation state vectors in chronological order of sampling time to form an operation state matrix for the information content. When the information content is executed, presetting the number of future sampling cycles, collecting the power value corresponding to each future sampling cycle, calculating the future power, and combining the operation state matrix, structural context vector, operation level label, and future power of the information content to construct a training sample triplet.

[0172] Establish a future power prediction model unit: based on the training sample triples, construct a dual-channel neural network model for future power prediction, use the operating state matrix in the training sample triples as the first channel, extract the state vector representing the operating sequence fluctuation characteristics through the LSTM network, use the structural context vector and the operating level label in the training sample triples as the second channel, extract the structural semantic feature vector through the GRU network, concatenate and fuse the state vector and the structural semantic feature vector to generate a joint representation vector, use a mixed density network MDN structure to output the future power, obtain all training sample triplets within the training cycle, use maximum likelihood estimation as the objective function, calculate the negative log-likelihood loss between the predicted future power distribution and the actual future power value, perform model training, and obtain a prediction model for future power.

[0173] Capacitor state module: This module collects the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combines the on-state to form an action vector, generates a capacitor state training set, trains a capacitor state prediction model, builds a digital twin system to simulate capacitor behavior, and outputs multi-dimensional operating indicators.

[0174] Real-time prediction module: extracts content features of user-uploaded information in real time, constructs a state matrix, structural context, and operating level labels, and inputs them into the prediction model to obtain future power values;

[0175] Candidate action vector module: collects the target sampling period capacitor state vector, combines multiple candidate action vectors to predict the future capacitor state, calculates the cost score, and sorts the candidate action vectors according to the cost score to obtain a candidate action vector set;

[0176] Real-time adjustment module: The real-time capacitor state, power value, predicted power and candidate action vectors are input into the digital twin system simulation to obtain the prediction index, filter the characteristic actions according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

[0177] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A full-tab capacitor control method for AI servers, characterized in that: Methods include: Step S100: Receive user-uploaded information in the AI ​​server terminal system, set a sampling period, collect server power parameters and record capacitor adjustment parameters, extract feature parameters of the information content and normalize them to construct a feature vector, calculate an operation score, divide the operation level, use the feature vector and operation level to construct a training sample, and train a random forest model to predict the operation level; Step S200: Set a training cycle, extract information content feature vectors and input them into the operation level prediction model, combine code analysis tools to generate structural context vectors, build an operation state matrix, calculate future power, form training sample triplets, build a dual-channel neural network based on LSTM and GRU to extract features, and predict future power through a hybrid density network, using maximum likelihood estimation to optimize the model; Step S300: Collect the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combine the open state to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, build a digital twin system to simulate capacitor behavior, and output multi-dimensional operation indicators; Step S400: extracting content features of user-uploaded information in real time, constructing a state matrix, structural context, and operating level labels, and inputting them into a prediction model to obtain future power values; Step S500: collecting the capacitance state vector of the target sampling period, combining multiple candidate action vectors to predict the future capacitance state, calculating the cost score, and sorting the candidate action vectors according to the cost score to obtain a set of candidate action vectors; Step S600: Input the real-time capacitor state, power value, predicted power and each candidate action vector into the digital twin system simulation, obtain the prediction index, filter the characteristic action according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

2. The method for controlling full-tab capacitance for an AI server according to claim 1, wherein: The step S100 includes the following steps: Step S101: In the terminal system of the AI ​​server, information content uploaded by the user is received, a sampling period is preset, power parameters of the server during the sampling period are monitored, capacitor adjustment parameters are recorded, and a sampling record for executing the information content is generated; Step S102: Collect a piece of information uploaded by the user, extract the content type of the information, determine the characteristic parameters corresponding to the content type, calculate the characteristic value of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector A=(a1,a2,...ab), where a1,a2,...ab represent the normalized values ​​of the 1st, 2nd,...bth characteristic parameters respectively; Step S103: In the historical sampling record set of the information content, each power response data of a certain historical sampling record is collected and normalized, and an operation score is calculated according to the following formula: ； Among them, B represents the running score, D c Expressed as the normalized value of the cth power response data, E c Expressed as the weight of the cth power response data, all historical sampling records are summarized and the average value of the running score is calculated; Step S104: Preset a number of operation levels, set the average range of operation scores corresponding to each operation level, and automatically classify each piece of information into the corresponding operation level; Step S105: In a set of information content of a certain content type, obtain the content feature vector and the operation level of each information content, use the operation level as a label, and construct a training sample set together with the content feature vector. Input the training sample set into the random forest model for training, wherein the content feature vector is used as input and the operation level label is used as output, and establish an operation level prediction model corresponding to the content type.

3. The method for controlling full-tab capacitance for an AI server according to claim 2, wherein: The step S200 includes the following steps: Step S201: Selecting a number of consecutive days as a training cycle, during which a piece of information uploaded by a user is collected, the content type of the information is extracted, a content feature vector of the information is constructed, and the feature vector is input into an operation level prediction model corresponding to the content type to obtain an operation level label for the information content; Step S202: extracting the task computing density and memory bandwidth requirements of the information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the information content; Step S203: Acquire a set of sampled records that execute the information content, collect the current mean value, power change rate, and load fluctuation coefficient in each sampled record to form an operation state vector of the sampled record, and arrange all the operation state vectors in chronological order of sampling time to form an operation state matrix of the information content; Step S204: When the information content is executed, the number of future sampling periods is preset, the power value corresponding to each future sampling period is collected, and the future power is calculated according to the following formula: ； Among them, F represents the future power, G e is represented by the power value corresponding to the e-th future sampling period, f is represented by the number of future sampling periods, and the operating state matrix, structural context vector, operating level label and future power of the information content are combined to construct a training sample triplet S=[Y,(J,X),F], where Y is represented by the operating state matrix, J is represented by the structural context vector, X is represented by the operating level label, and F is represented by the future power; Step S205: Based on the training sample triples, a dual-channel neural network model for future power prediction is constructed. The operating state matrix in the training sample triples is used as the first channel. The state vector representing the operating sequence fluctuation characteristics is extracted through the LSTM network. The structural context vector and the operating level label in the training sample triples are used as the second channel. The structural semantic feature vector is extracted through the GRU network. The state vector and the structural semantic feature vector are concatenated and fused to generate a joint representation vector. The future power is output using a mixed density network (MDN) structure. Step S206: Obtain all training sample triplets within the training cycle, use maximum likelihood estimation as the objective function, calculate the negative log-likelihood loss between the predicted future power distribution and the actual future power value, perform model training, and obtain a prediction model for future power.

4. The method for controlling full-tab capacitance for an AI server according to claim 3, wherein: The step S300 includes the following steps: Step S301: collecting state variables of each capacitor unit in the terminal system corresponding to a time point of a certain sampling record, wherein the state variables include equivalent series resistance, state of charge, and temperature, to form a capacitor state vector; Step S302: obtaining the on-state of each capacitor unit, where the on-state includes enabled and disabled, and combining the on-states of each capacitor unit to construct an action vector; Step S303: combining the capacitance state vector and action vector of a certain sampling record with the capacitance state vector of the next adjacent sampling record to form a capacitance state training group. Inputting all capacitance state training groups into the neural network model for training to obtain a capacitance state prediction model. Step S304: Establish a digital twin system based on the set of sampling records within the training cycle, input the capacitor state vector, action vector, current power value and predicted power value corresponding to a certain sampling record into the digital twin system, perform simulation operations in the digital twin system, and obtain output results including voltage output, power output, number of voltage drops, response delay and capacitor temperature.

5. The method for controlling full-tab capacitance for an AI server according to claim 4, wherein: The step S400 includes the following steps: Step S401: obtaining real-time information content uploaded by a user terminal, extracting the content type of the real-time information content, constructing a content feature vector of the real-time information content, and inputting the vector into an operation level prediction model corresponding to the content type to obtain an operation level label of the real-time information content; Step S402: extracting the task computing intensity and memory bandwidth requirement of the real-time information content through a static code analysis tool built into the terminal system, and performing normalization processing to form a structural context vector of the real-time information content; Step S403: The number of sampling periods for analysis and optimization is preset to n, a set of sampling periods of real-time information content is obtained, and an operation status matrix of the real-time information content is constructed; Step S404: inputting the operation state matrix, structural context vector, and operation level label of the real-time information content into the future power prediction model to obtain the predicted value of the future power.

6. The method for controlling full-tab capacitance for an AI server according to claim 5, wherein: The step S500 includes the following steps: Step S501: obtaining a time point corresponding to the nth sampling period of the analysis and optimization, collecting the real-time state variables of each capacitor unit of the terminal system corresponding to the time point, and obtaining a real-time capacitor state vector; Step S502: Preset a number of candidate motion vectors, input the real-time capacitance state vector and each candidate motion vector into a capacitance state prediction model, and predict the next capacitance state vector; Step S503: extract the equivalent series resistance R and the state of charge H in a predicted next capacitance state vector, and calculate the cost score according to the following formula: ； Where P represents the cost score, R0 represents the standard value of equivalent series resistance, r1 and r2 represent the weights of equivalent series resistance and state of charge, respectively; Step S504: Summarize the cost scores corresponding to all candidate motion vectors, and arrange all candidate motion vectors from low to high according to the cost scores to obtain a set of candidate motion vectors.

7. The method for controlling full-tab capacitance for an AI server according to claim 6, wherein: The step S600 includes the following steps: Step S601: The real-time capacitor state vector, the real-time current power value, and the predicted value of future power at the time point corresponding to the nth sampling period are combined with each candidate action vector in the candidate action vector set, and input into the digital twin system. A simulation operation is performed to obtain the predicted number of voltage drops, predicted response delay, and predicted capacitor temperature corresponding to each candidate action vector. Step S602: Preset a response delay threshold and a capacitor temperature threshold. If a candidate action vector satisfies all of the following conditions: the predicted number of voltage drops is 0; the predicted response delay is less than the response delay threshold; and the predicted capacitor temperature is less than the capacitor temperature threshold, then the candidate action vector is set as a feature action vector. Step S603: extracting the position order of each characteristic action vector in the candidate action vector set, selecting the characteristic action vector with the smallest position order as the final adjustment instruction, and inputting it into the terminal system to adjust the capacitor unit.

8. A full-tab capacitor control system for an AI server, used to implement the full-tab capacitor control method for an AI server as described in any one of claims 1 to 7, characterized in that: The system includes a predicted operating level module, a future power module, a capacitor state module, a real-time prediction module, a candidate action vector module, and a real-time adjustment module; The operation level prediction module receives user-uploaded information in the AI ​​server terminal system, sets a sampling period, collects server power parameters and records capacitor adjustment parameters, extracts characteristic parameters of the information content and normalizes them to construct a feature vector, calculates an operation score, divides the operation level, constructs a training sample using the feature vector and the operation level, and trains a random forest model to predict the operation level; The future power module sets a training cycle, extracts information content feature vectors and inputs them into the operation level prediction model, combines code analysis tools to generate structural context vectors, constructs an operation state matrix, calculates future power, forms training sample triplets, constructs a dual-channel neural network based on LSTM and GRU to extract features, and predicts future power through a mixed density network, using a maximum likelihood estimation optimization model. The capacitor state module collects the state variables of the capacitor unit at the time point corresponding to each sampling record to construct a capacitor state vector, combines the open state to form an action vector, generates a capacitor state training group and trains a capacitor state prediction model, builds a digital twin system to simulate capacitor behavior, and outputs multi-dimensional operation indicators; The real-time prediction module extracts the content features of user uploaded information in real time, constructs a state matrix, structural context and operation level label, and inputs them into the prediction model to obtain future power values; The candidate action vector module collects the target sampling period capacitance state vector, combines multiple candidate action vectors to predict the future capacitance state, calculates the cost score, and sorts the candidate action vectors according to the cost score to obtain a candidate action vector set; The real-time adjustment module inputs the real-time capacitor state, power value, predicted power and each candidate action vector into the digital twin system simulation, obtains the prediction index, filters the characteristic action according to the threshold, determines the final adjustment instruction, and adjusts the capacitor unit according to the final adjustment instruction.

9. The full-tab capacitor control system for an AI server according to claim 8, characterized in that: The operation level prediction module includes an operation score calculation unit and an operation level prediction model establishment unit: The calculation and operation scoring unit: in the terminal system of the AI ​​server, receives the information content uploaded by the user terminal, presets a sampling period, monitors the power parameters of the server during the sampling period, records the capacitance adjustment parameters, generates a sampling record for executing the information content, collects a certain information content uploaded by the user terminal, extracts the content type of the information content, determines the characteristic parameters corresponding to the content type, calculates the characteristic value of each characteristic parameter in the query content, performs normalization processing, constructs a content feature vector, collects each power response data of a certain historical sampling record in the historical sampling record set of the information content, performs normalization processing, and calculates the operation score; The operation level prediction model establishment unit presets several operation levels and sets the average range of operation scores corresponding to each operation level, automatically divides each piece of information content into the corresponding operation level, obtains the content feature vector and operation level of each piece of information content in a set of information content of a certain content type, uses the operation level as a label, and constructs a training sample set together with the content feature vector, inputs the training sample set into a random forest model for training, wherein the content feature vector is used as input and the operation level label is used as output, and establishes an operation level prediction model corresponding to the content type.

10. The full-tab capacitor control system for AI server according to claim 8, characterized in that: The future power module includes a unit for constructing a training sample triple and a unit for establishing a future power prediction model: The training sample triple construction unit includes: selecting a number of consecutive days as a training cycle; collecting a piece of information content uploaded by a user terminal during the training cycle; extracting the content type of the information content; constructing a content feature vector of the information content; and inputting the feature vector into an operation level prediction model corresponding to the content type to obtain an operation level label of the information content; extracting the task computing intensity and memory bandwidth requirement of the information content through a static code analysis tool built into the terminal system; performing normalization processing to construct a structural context vector of the information content; obtaining a set of sampling records for executing the information content; collecting the current mean value, power change rate, and load fluctuation coefficient in each sampling record to form an operation state vector of the sampling record; arranging all the operation state vectors in chronological order of sampling time to form an operation state matrix of the information content; and when the information content is executed, presetting the number of future sampling cycles, collecting the power value corresponding to each future sampling cycle, calculating the future power, and combining the operation state matrix of the information content, the structural context vector, the operation level label, and the future power to construct a training sample triple. The future power prediction model establishment unit: based on the training sample triples, constructs a dual-channel neural network model for future power prediction, uses the operating state matrix in the training sample triples as the first channel, extracts the state vector representing the operating sequence fluctuation characteristics through the LSTM network, uses the structural context vector and the operating level label in the training sample triples as the second channel, extracts the structural semantic feature vector through the GRU network, splices and fuses the state vector and the structural semantic feature vector to generate a joint representation vector, adopts a mixed density network MDN structure, outputs the future power, obtains all training sample triplets in the training cycle, adopts maximum likelihood estimation as the objective function, calculates the negative log-likelihood loss between the predicted future power distribution and the actual future power value, performs model training, and obtains a prediction model for future power.

Images