Full-tab capacitance regulation and control system and method for AI server

Abstract

The invention discloses a full-tab capacitance regulation and control system and method for an AI server, and relates to the technical field of data analysis. Information content uploaded by a user is received in a terminal system, feature parameters of the information content are extracted, feature vectors are constructed in a normalized mode, and the feature vectors are stored in the terminal system; constructing a training sample by using the feature vector and the operation level, and training a random forest model to predict the operation level; a training period is set, a training sample triple is formed, a two-channel neural network is constructed based on LSTM and GRU to fuse and extract features, and a maximum likelihood estimation optimization model is adopted; generating a capacitance state training group and training a capacitance state prediction model, constructing a digital twin system to perform capacitance behavior simulation, and outputting a multi-dimensional operation index; collecting a capacitance state vector of a target sampling period, and predicting a future capacitance state in combination with a plurality of candidate action vectors; and screening characteristic actions according to a threshold value, determining a final adjustment instruction, and improving the response speed.

Description

Technical Field

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

[0002] Currently, when an AI server executes large-scale parallel computing tasks, the power consumption changes violently and has a high degree of uncertainty, which puts higher requirements on the dynamic response ability of the power supply system. With the continuous expansion of the scale of artificial intelligence computing, full-tab capacitors have been widely used in the field of high-performance power supply voltage regulation due to their lower equivalent series resistance (ESR) and faster charge and discharge speed. However, in actual applications, traditional capacitor array configuration methods mostly use preset parameters and cannot dynamically adjust according to load changes, making it difficult to meet the frequently changing current requirements in current AI computing. At the same time, the lack of real-time prediction ability for AI task loads often leads to problems such as capacitor response lag and improper enabling timing, resulting in reduced energy efficiency and even performance fluctuations. In order 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 the load change trend, adjust the capacitor configuration in advance, thereby optimizing the energy supply path, improving the response speed, reducing energy consumption, and avoiding the impact of sudden failures on system operation, thus significantly enhancing the security and stability of the server power supply system and promoting the efficient utilization of computing resources. Summary of the Invention

[0003] The purpose of the present invention is to provide a full-tab capacitor regulation system and method for an AI server to solve the problems raised in the prior art.

[0004] To solve the above technical problems, the present invention provides the following technical solution: A full-tab capacitor regulation method for an AI server, the method comprising: Step S100: Receive the information content uploaded by the user in the AI server terminal system, set the sampling period, collect the server power parameters and record the capacitor adjustment parameters, extract the characteristic parameters of the information content and normalize them to construct a feature vector, calculate the operation score, divide the operation level, and use the feature vector and the operation level to construct a training sample and train a random forest model to predict the operation level; Step S200: Set the training period, extract the information content feature vector and input it into the operation level prediction model, generate a structural context vector in combination with a code analysis tool, construct an operation state matrix, calculate the future power, form a training sample triple, construct a dual-channel neural network fusion based on LSTM and GRU to extract features, and predict the future power through a mixture density network, and optimize the model using maximum likelihood estimation; Step S300: Collect the state variables of the capacitor unit at the time points corresponding to each sampling record to construct a capacitor state vector, combine the on states to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, construct a digital twin system for capacitor behavior simulation, and output multi-dimensional operation indicators; Step S400: Extract the content features of the user-uploaded information in real time, construct a state matrix, a structural context, and an operation level label, and input them into the prediction model to obtain the future power value; Step S500: Collect the capacitor state vector of the target sampling period, predict the future capacitor state in combination with multiple candidate action vectors, calculate the cost score, sort the candidate action vectors according to the cost score, and 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 for simulation, obtain the prediction indicators, screen the characteristic actions according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

[0005] Furthermore, Step S100 includes: Step S101: In the terminal system of the AI server, receive the information content uploaded by the user terminal, preset the sampling period, monitor the power parameters of the server during the sampling period, record the capacitor adjustment parameters, and generate a sampling record for executing the information content; Step S102: Collect a piece of information content uploaded by the user terminal, extract the content type of the information content, determine the characteristic parameters corresponding to the content type, calculate the characteristic values of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector as A=(a1,a2,...ab), where a1,a2,...ab respectively represent the normalized values of the 1st, 2nd,...bth characteristic parameters; Step S103: In the historical sampling record set of the information content, collect each power response data of a historical sampling record, and perform normalization processing. Calculate the operation score according to the following formula: ; where B represents the operation score, D c represents the normalized value of the cth power response data, and E c represents the weight value of the cth power response data. Summarize all historical sampling records and calculate the average value of the operation score; Step S104: Preset several operation levels and set the operation score average value range corresponding to each operation level, and automatically divide each piece of information content into the corresponding operation level; Step S105: In the information content set of a certain content type, obtain the content feature vector and the running level of each piece of information content. Use the running level as a label, and jointly construct a training sample set with the content feature vector. Input the training sample set into a random forest model for training, where the content feature vector is used as the input and the running level label is used as the output, to establish a running level prediction model corresponding to the content type; The power parameters include voltage, current, power, energy consumption, etc. The content types include text, image, video, etc. The characteristic parameters of the text type include the total number of words, keyword density, syntactic complexity, semantic vector, etc. The characteristic parameters of the image type include image size, color histogram distribution, texture features, image clarity, etc. The characteristic parameters of the video type include video duration, frame rate, resolution, scene switching frequency, etc. The power response data includes average power consumption, power volatility, capacitor call frequency, etc.; By collecting the power parameters, content features, and historical response data of the AI server during the sampling period, this method can automatically analyze the running status of information content. Using normalization processing and a scoring mechanism, the evaluation process has a high degree of standardization and consistency, avoiding human errors; By forming a training sample with the content feature vector and the running level label and inputting it into a random forest model for training, the system can automatically learn and predict the running level of future uploaded content. The random forest has good generalization ability and anti-overfitting ability, making the prediction results have high robustness and accuracy; The extraction of content types and the identification of corresponding characteristic parameters have strong adaptability and can be flexibly applied to various types of information content processing scenarios, such as text, image, video, etc.

[0006] Further, step S200 includes: Step S201: Select a continuous number of days as the training period. During the training period, collect a certain piece of information content uploaded by the user side, extract the content type of the information content, construct the content feature vector of the information content, and input it into the running level prediction model corresponding to the content type to obtain the running level label of the information content; Step S202: Through the static code analysis tool built into the terminal system, extract the task computational intensity and memory bandwidth requirements of the information content, and perform normalization processing to form the structural context vector of the information content; Step S203: Obtain the sampling record set for executing the information content, collect the current mean value, power change rate, and load fluctuation coefficient in each sampling record, form the running state vector of the sampling record, and arrange all the running state vectors in the order of sampling time to form the running state matrix of the information content; Step S204: When the execution of the information content is completed, preset the number of future sampling periods, collect the power values corresponding to each future sampling period, and calculate the future power according to the following formula: ; where F represents the future power, G e represents the power value corresponding to the e-th future sampling period, f represents the number of future sampling periods, combine the operation status matrix, structure context vector, operation level label and future power of the information content to construct a training sample triple as S = [Y, (J, X), F], where Y represents the operation status matrix, J represents the structure context vector, X represents the operation level label, and F represents the future power; Step S205: Based on the training sample triple, construct a two-channel neural network model for future power prediction. Use the operation status matrix in the training sample triple as the first channel, and extract the state vector representing the operation time series fluctuation characteristics through the LSTM network. Use the structure context vector and operation level label in the training sample triple as the second channel, and extract the structure semantic feature vector through the GRU network. Concatenate and fuse the state vector and the structure semantic feature vector to generate a joint representation vector, and adopt the mixture density network MDN structure to output the future power; Step S206: Obtain all the training sample triples within the training period, use the 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 the prediction model of the future power; Use a two-channel neural network structure: The first channel uses LSTM and focuses on extracting the time series fluctuation characteristics in the operation status matrix; the second channel uses GRU to model the semantic features between the code structure context and the operation level. The joint representation vector after the fusion of the two channels can better capture the potential laws and significantly improve the modeling ability of the future power; Adopt a mixture density network, which no longer outputs a single power prediction value, but outputs the probability distribution of the power value, enhancing the performance of the model in scenarios such as multi-modal prediction and non-linear uncertainty modeling, and is more suitable for scenarios where the AI server load is complex and the power demand is volatile and uncertain.

[0007] Further, step S300 includes: Step S301: Collect the state variables of each capacitor unit of the terminal system corresponding to the time point of a certain sampling record. The state variables include equivalent series resistance, state of charge, and temperature, and constitute a capacitor state vector; Step S302: Obtain the on state of each capacitor unit. The on state includes enabled and disabled, and combine the on states of each capacitor unit to construct an action vector; Step S303: Combine the capacitance state vector and the action vector of a certain sampling record with the capacitance state vector of the next adjacent sampling record to form a capacitance state training set, and input all the capacitance state training sets into the neural network model for training to obtain a prediction model for the capacitance state; Step S304: Establish a digital twin system based on the sampling record set within the training period. Input the capacitance state vector, action vector, current power value, and predicted power value corresponding to a certain sampling record into the digital twin system, and perform a simulation operation in the digital twin system to obtain output results including voltage output, power output, number of voltage dips, response delay, and capacitance temperature; By combining the capacitance state vector and the action vector of each sampling point and combining with the capacitance state vector of the next sampling point, high-quality training samples are constructed, effectively reflecting the state evolution of the capacitance unit under different control strategies, which helps to improve the fitting ability of the prediction model to the capacitance behavior; The obtained prediction model for the capacitance state can predict the next moment's state based on the current state and operation, enabling the system to have the ability to predict the future state change trend, providing forward-looking support for the regulation system, and reducing response lag; The digital twin system not only reflects the state of the physical system but also can accept real-time input, generate outputs related to voltage output, power output, number of voltage dips, response delay, and capacitance temperature, and can realize offline evaluation and optimization of the action strategy.

[0008] Further, step S400 includes: Step S401: Obtain the real-time information content uploaded by the user terminal, extract the content type of the real-time information content, construct a content feature vector of the real-time information content, and input it into the operation level prediction model corresponding to the content type to obtain the operation level label of the real-time information content; Step S402: Extract the task computational intensity and memory bandwidth requirements of the real-time information content through the static code analysis tool built into the terminal system, and perform normalization processing to form the structural context vector of the real-time information content; Step S403: Preset the number of sampling periods for analysis and optimization as n, obtain the sampling period set of the real-time information content, and construct the operation state matrix of the real-time information content; Step S404: Input the operation state matrix, structural context vector, and operation level label of the real-time information content into the prediction model of future power to obtain the predicted value of future power; Generate operating level labels according to the type of real-time information content, generate structural context vectors based on static code structure analysis, capture the real-time, complexity and resource dependence of the current task, and effectively improve the power prediction model's ability to recognize short-term behavior fluctuations through structured representation vectors.

[0009] Furthermore, step S500 includes: Step S501: obtaining the 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 the real-time capacitor state vector; Step S502: Preset a number of candidate action vectors, input the real-time capacitance state vector and each candidate action vector into a capacitance state prediction model, and predict the next capacitance state vector; Step S503: extracting the equivalent series resistance R and the state of charge H in a predicted next capacitance state vector, and calculating 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 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; Through candidate action simulation and state prediction, the system can determine in advance the specific impact of each control action on the future state of the capacitor, and then conduct quantitative evaluation through the cost scoring function, thus realizing the strategy of "evaluation first, then execution", replacing the traditional "experience activation" mode, which can significantly improve the robustness and foresight of the control strategy, and reduce performance degradation or risk events caused by activating the wrong capacitor unit; Based on the trained capacitor state prediction model, the equivalent series resistance and charge state of the capacitor at the next moment can be predicted according to the current state and action combination, which improves the modeling ability of the "behavior evolution" of the capacitor and helps prevent high-temperature, low-power or high-impedance capacitor units from continuing to operate. By presetting multiple candidate action vectors for parallel prediction and scoring, the optimal control strategy under current conditions can be found.

[0010] Furthermore, step S600 includes: 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 respectively combined with each candidate action vector in the candidate action vector set, and input into the digital twin system, and a simulation operation is performed to obtain the predicted voltage drop times, 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 certain candidate action vector meets all of the following conditions: the predicted number of voltage dips is 0; the predicted response delay is less than the response delay threshold; the predicted capacitor temperature is less than the capacitor temperature threshold; then set the candidate action vector as the characteristic action vector; Step S603: Extract the rank of each characteristic action vector in the candidate action vector set, and select the characteristic action vector with the smallest rank as the final adjustment instruction, and input it into the terminal system to adjust the capacitor unit; Take the real-time capacitor state, the current power and the predicted power value as the combined input, combine each candidate action, perform simulation in the digital twin system, and output the predicted voltage dip, response delay, and capacitor temperature. This mechanism constructs a complete prediction chain from "current state → behavior decision → system reaction", simulates the impact of control actions on the future of the system, and realizes the real "prediction-based control strategy screening"; Set three conditions: the number of voltage dips = 0 to ensure power supply continuity; the response delay < threshold to ensure response speed; the capacitor temperature < threshold to avoid thermal runaway; significantly enhance operation stability and safety; Among multiple qualified characteristic actions, preferentially select the action vector with the smallest rank. The "two-stage screening" mechanism reflects: the first stage: safety and feasibility evaluation; the second stage: performance optimization; ensure that the selected strategy is not only safe but also optimal in terms of performance and energy efficiency.

[0011] To better implement the above method, a full-pole ear capacitor regulation system for an AI server is also proposed. The system includes a predicted operation 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; Predicted operation level module: Receive the information content uploaded by the user in the AI server terminal system, set the sampling period, collect the server power parameters and record the capacitor adjustment parameters, extract the characteristic parameters of the information content and normalize them to construct a characteristic vector, calculate the operation score, divide the operation level, and use the characteristic vector and the operation level to construct a training sample and train a random forest model to predict the operation level; Future power module: Set the training period, extract the characteristic vector of the information content and input it into the operation level prediction model, combine the code analysis tool to generate a structural context vector, construct an operation state matrix, calculate the future power, form a training sample triple, construct a dual-channel neural network fusion based on LSTM and GRU to extract features, and predict the future power through a mixture density network, and optimize the model using maximum likelihood estimation; Capacitance status module: Collect the status variables of the capacitance units at the time points corresponding to each sampling record to construct a capacitance status vector, combine the opening status to form an action vector, generate a capacitance status training set and train a capacitance status prediction model, construct a digital twin system for capacitance behavior simulation, and output multi-dimensional operation indicators; Real-time prediction module: Extract the feature of the information content uploaded by the user in real time, construct a status matrix, a structural context, and an operation level label, and input them into the prediction model to obtain the future power value; Candidate action vector module: Collect the capacitance status vector of the target sampling period, combine multiple candidate action vectors to predict the future capacitance status, calculate the cost score, sort the candidate action vectors according to the cost score, and obtain a set of candidate action vectors; Real-time adjustment module: Input the real-time capacitance status, power value, predicted power, and each candidate action vector into the digital twin system for simulation, obtain the prediction index, screen the characteristic actions according to the threshold, determine the final adjustment instruction, and adjust the capacitance unit according to the final adjustment instruction.

[0012] Furthermore, the predicted operation level module includes a computing operation score unit and an establishing operation level prediction model unit: Computing operation score unit: In the terminal system of the AI server, receive the information content uploaded by the user terminal, preset the sampling period, monitor the power parameters of the server during the sampling period, record the capacitance adjustment parameters, generate a sampling record for executing the information content, collect a certain piece of information content uploaded by the user terminal, extract the content type of the information content, determine the characteristic parameters corresponding to the content type, calculate the characteristic values of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector. In the historical sampling record set of the information content, collect each power response data of a certain historical sampling record, and perform normalization processing to calculate the operation score; Establishing operation level prediction model unit: Preset several operation levels, and set the average operation score range corresponding to each operation level. Automatically divide each piece of information content into the corresponding operation level. In the information content set of a certain content type, obtain the content feature vector and operation level of each piece of information content, 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, where the content feature vector is used as the input and the operation level label is used as the output, to establish the operation level prediction model corresponding to the content type.

[0013] Furthermore, the future power module includes a constructing training sample triple unit and an establishing future power prediction model unit: Constructing a training sample triplet unit: selecting a number of consecutive days as a training cycle, collecting a certain information content uploaded by the user terminal within the training cycle, extracting the content type of the information content, constructing a content feature vector of the information content, and inputting it 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 density and memory bandwidth requirement 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, 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 the 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, 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 triplet; Establish a future power prediction model unit: 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, a mixed density network MDN structure is used to output the future power, all training sample triplets in the training cycle are obtained, the maximum likelihood estimation is used as the objective function, the negative log-likelihood loss between the predicted future power distribution and the actual future power value is calculated, the model is trained, and a prediction model for future power is obtained.

[0014] Compared with the prior art, the beneficial effects of the present invention are: introducing a neural network prediction model to predict the evolution of capacitor state in advance, combining future power prediction with simulation feedback, implementing a "prediction-simulation-optimization-execution" closed-loop control, improving the accuracy and response foresight of capacitor regulation decisions, and avoiding strategy lag and resource waste; Introducing dimensions such as task content semantic features, static structure context features, and operating state matrix, based on dual-channel deep network fusion processing, it mines the power consumption behavior of complex tasks, supports the dynamic changes of different types of tasks and loads in AI servers, and enhances generalization and adaptability; Build a capacitive digital twin system that takes status, power, and candidate control actions as inputs, outputs key metrics, and accurately simulates "the impact of a certain control action on the capacitive system over a period of time in the future" to provide visual, quantitative, and reliable prediction support for decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is a schematic flow chart of a full-pole-ear capacitor regulation method for an AI server according to the present invention; Figure 2 FIG. is a schematic structural diagram of a full-pole-ear capacitor regulation system for an AI server according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] Please refer to Figure 1 , the present invention provides a technical solution: a full-pole-ear capacitor regulation method for an AI server, the method comprising: Step S100: Receive the information content uploaded by the user in the AI server terminal system, set the sampling period, collect the server power parameters and record the capacitor regulation parameters, extract the feature parameters of the information content and normalize them to construct a feature vector, calculate the operation score, divide the operation level, and use the feature vector and the operation level to construct a training sample and train a random forest model to predict the operation level; Among them, step S100 includes: Step S101: In the terminal system of the AI server, receive the information content uploaded by the user terminal, preset the sampling period, monitor the server power parameters during the sampling period, record the capacitor regulation parameters, and generate a sampling record for executing the information content; Step S102: Collect a certain piece of information content uploaded by the user terminal, extract the content type of the information content, determine the feature parameters corresponding to the content type, calculate the feature values of each feature parameter in the query content, and perform normalization processing to construct a content feature vector as A=(a1,a2,...ab), where a1,a2,...ab respectively represent the normalized values of the 1st, 2nd,...bth feature parameters; Step S103: In the historical sampling record set of the information content, collect each power response data of a certain historical sampling record, and perform normalization processing, and calculate the operation score according to the following formula: ; Among them, B represents the running score, and D c represents the normalized value of the c-th power response data, and E c represents the weight of the c-th power response data. Summarize all historical sampling records and calculate the average value of the running score; Step S104: Preset several running levels and set the range of the average running score corresponding to each running level, and automatically divide each piece of information content into the corresponding running level; Step S105: In the set of information content of a certain content type, obtain the content feature vector and running level of each piece of information content. Use the running 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, where the content feature vector is used as the input and the running level label is used as the output to establish the running level prediction model corresponding to the content type; For example, the content type of the A-th piece of information content collected is image class, and the information content parameters are extracted as follows: image size: 2MB; resolution: 1920×1080; number of channels: 3 (RGB); selected model: ResNet50; task label: image_classification; after normalization, the content feature vector is constructed as (0.2, 0.23, 0.75, 0.66, 0.33); In the b-th sampling record, the normalized power response data is as follows: average power is 0.7, weight is 0.4; fluctuation coefficient is 0.12, weight is 0.3; peak power is 0.9, weight is 0.3, and the calculated running score is 0.586; Suppose multiple task records collect feature vectors and running level labels: The feature vector of the first task record is (0.2, 0.23, 0.75, 0.66, 0.33), and the running level is L3; the feature vector of the second task record is (0.1, 0.15, 0.5, 0.33, 0.33), and the running level is L2; the feature vector of the third task record is (0.4, 0.45, 1.0, 1.0, 0.66), and the running level is L4. Use these samples to train the random forest model, input: content feature vector; output: running level label.

[0018] Step S200: Set the training period, extract the information content feature vector and input it into the running level prediction model, generate a structural context vector in combination with the code analysis tool, construct a running state matrix, calculate the future power, form a training sample triple, construct a dual-channel neural network fusion based on LSTM and GRU to extract features, and predict the future power through a mixture density network, and optimize the model using maximum likelihood estimation; Among them, step S200 includes: Step S201: Select a continuous number of days as the training period. During the training period, collect a certain piece of information content uploaded by the user side, extract the content type of the information content, construct the content feature vector of the information content, and input it into the operation level prediction model corresponding to the content type to obtain the operation level label of the information content; Step S202: Extract the task calculation intensity and memory bandwidth requirements of the information content through the static code analysis tool built in the terminal system, and perform normalization processing to form the structural context vector of the information content; Step S203: Obtain the sampling record set for executing the information content, collect the average current, power change rate, and load fluctuation coefficient in each sampling record, form the operation state vector of the sampling record, and arrange all the operation state vectors in the order of sampling time to form the operation state matrix of the information content; Step S204: When the execution of the information content is completed, preset the number of future sampling periods, collect the power values corresponding to each future sampling period, and calculate the future power according to the following formula: ; Among them, F represents the future power, G e represents the power value corresponding to the e-th future sampling period, f represents the number of future sampling periods, combine the operation state matrix, structural context vector, operation level label, and future power of the information content to construct a training sample triple S = [Y, (J, X), F], where Y represents the operation state matrix, J represents the structural context vector, X represents the operation level label, and F represents the future power; Step S205: Based on the training sample triple, construct a dual-channel neural network model for future power prediction. Use the operation state matrix in the training sample triple as the first channel, and extract the state vector representing the operation time series fluctuation characteristics through the LSTM network. Use the structural context vector and operation level label in the training sample triple as the second channel, and 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, and adopt the mixture density network MDN structure to output the future power; Step S206: Obtain all the training sample triples within the training period, use the 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, and perform model training to obtain the prediction model of the future power; For example, a user uploads an image task. The system extracts the feature vector as (0.2, 0.23, 0.75, 0.66, 0.33) and inputs it into the running level prediction model corresponding to the "image recognition class". The model outputs the running level label as L3; The system calls a static code analysis tool to analyze model calls and resource distribution: the normalized computational intensity is 0.82, and the normalized memory bandwidth requirement is 0.74, forming the structural context vector J = (0.82, 0.74); Assume that the sampling records during the operation of the image recognition task are shown in Table 1;

[0019] Table 1 Each one forms a running state vector to form a running state matrix; Assume that the powers sampled in the subsequent 5 cycles are: 65.3W, 67.8W, 70.1W, 71.0W, 69.5W. Calculate the future power as 68.74 and construct a training sample triple; The first channel: Input: the running state matrix Y, Network structure: LSTM, Extract running fluctuation features; Output: state vector; The second channel: Input: the structural context vector J + the running level label X; Network structure: GRU network, Extract context semantic representations; Output: structural feature vector; Concatenate the state vector and the structural feature vector, input them into a mixture density network, and output the means, variances, and weights of multiple Gaussian distribution parameters; Use the negative log-likelihood loss function and train with the Adam optimizer to adjust the weights of the LSTM / GRU / MDN joint model, and finally obtain the prediction function.

[0020] Step S300: Collect the state variables of the capacitor units at the time points corresponding to each sampling record to construct a capacitor state vector, combine the on states to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, construct a digital twin system for capacitor behavior simulation, and output multi-dimensional operation indicators; Among them, step S300 includes: Step S301: Collect the state variables of each capacitor unit of the terminal system corresponding to the time point of a certain sampling record. The state variables include equivalent series resistance, state of charge, and temperature, and form a capacitor state vector; Step S302: Obtain the on state of each capacitor unit. The on state includes enabled and disabled, and combine the on states of each capacitor unit to construct an action vector; Step S303: Combine the capacitance state vector and the action vector of a certain sampling record with the capacitance state vector of the next adjacent sampling record to form a capacitance state training set, and input all the capacitance state training sets into the neural network model for training to obtain a prediction model for the capacitance state; Step S304: Establish a digital twin system based on the sampling record set within the training period, input the capacitance state vector, action vector, current power value, and predicted power value corresponding to a certain sampling record into the digital twin system, and perform a simulation operation in the digital twin system to obtain output results including voltage output, power output, number of voltage dips, response delay, and capacitance temperature; For example, there are 3 capacitance units C1, C2, C3; at sampling time T1, collect the following state parameters of each capacitance unit: the equivalent series resistance of C1 is 0.12, the state of charge is 0.65, and the temperature is 48.2; the equivalent series resistance of C2 is 0.15, the state of charge is 0.62, and the temperature is 50.1; the equivalent series resistance of C3 is 0.11, the state of charge is 0.68, and the temperature is 46.9; construct a capacitance state vector Lt(0.12, 0.65, 48.2, 0.15, 0.62, 50.1, 0.11, 0.68, 46.9); Current action: Enable C1 and C3, disable C2; the action encoding for enabling is 1, for disabling is 0, and the action vector is Ut(1, 0, 1); The states collected at the adjacent time point T2: the equivalent series resistance of C1 is 0.13, the state of charge is 0.64, and the temperature is 48.7; the equivalent series resistance of C2 is 0.15, the state of charge is 0.61, and the temperature is 50.3; the equivalent series resistance of C3 is 0.12, the state of charge is 0.67, and the temperature is 47.1; construct a capacitance state vector Lt+1(0.13, 0.64, 48.7, 0.15, 0.61, 50.3, 0.12, 0.67, 47.1); Take [Lt, Ut] as the input and Lt+1 as the output to construct a training sample, and summarize multiple sampling records to form a training set, and train the neural network as a capacitance state prediction model.

[0021] Step S400: Extract the content features of the user-uploaded information in real time, construct a state matrix, structural context, and operation level label, and input them into the prediction model to obtain the future power value; Among them, Step S400 includes: Step S401: Obtain the real-time information content uploaded by the user side, extract the content type of the real-time information content, construct a content feature vector of the real-time information content, and input it into the operation level prediction model corresponding to the content type to obtain the operation level label of the real-time information content; Step S402: Extract the task computational intensity and memory bandwidth requirements of the real-time information content through the static code analysis tool built into the terminal system, and perform normalization processing to form the structural context vector of the real-time information content; Step S403: Preset the number of sampling periods for analysis and optimization as n, obtain the sampling period set of the real-time information content, and construct the operating state matrix of the real-time information content; Step S404: Input the operating state matrix, structural context vector, and operating level label of the real-time information content into the prediction model of future power to obtain the predicted value of future power; For example, the user uploads an image recognition task, and the real-time content information is as follows: Image size: 3.5MB; Resolution: 2560×1440; Number of channels: 3 (RGB); Model type: EfficientNet-B3; Task type: image_classification; Construct the feature vector (0.35, 0.43, 0.75, 0.8, 0.33), and input it into the "image recognition class" operating level prediction model, output: operating level L4; Static code analysis result: Normalized computational intensity: 0.92; Normalized memory bandwidth requirement: 0.85; Construct the structural context vector J=(0.92, 0.85); Assume that the operating state data in the first 5 cycles are collected as Table 2

[0022] Table 2 Construct the operating state matrix; Input the operating state matrix, structural context vector, and operating level label into the future power prediction model, and the predicted value of future power obtained is 84.6W.

[0023] Step S500: Collect the capacitance state vector of the target sampling period, combine multiple candidate action vectors to predict the future capacitance state, calculate the cost score, and sort the candidate action vectors according to the cost score to obtain the candidate action vector set; Among them, Step S500 includes: Step S501: Obtain the time point corresponding to the nth sampling period for analysis and optimization, and collect the real-time state variables of each capacitance unit of the terminal system corresponding to the time point to obtain the real-time capacitance state vector; Step S502: Preset several candidate action vectors, input the real-time capacitance state vector and each candidate action vector into the prediction model of capacitance state to predict the next capacitance state vector; Step S503: Extract the equivalent series resistance as R and the state of charge as H in a certain predicted next capacitor state vector, and calculate the cost score according to the following formula: ; where P represents the cost score, R0 represents the standard value of the equivalent series resistance, and r1 and r2 represent the weights of the equivalent series resistance and the state of charge respectively; Step S504: Aggregate the cost scores corresponding to all candidate action vectors, arrange all candidate action vectors in ascending order of the cost score to obtain a set of candidate action vectors; For example, at sampling time T5, collect the following state parameters of each capacitor unit: for C1, the equivalent series resistance is 0.12, the state of charge is 0.66, and the temperature is 48.5; for C2, the equivalent series resistance is 0.14, the state of charge is 0.64, and the temperature is 49.8; for C3, the equivalent series resistance is 0.13, the state of charge is 0.67, and the temperature is 47.2; construct the capacitor state vector as Lt5(0.12, 0.66, 48.5, 0.14, 0.64, 49.8, 0.13, 0.67, 47.2); Preset candidate action vectors U1 = (1, 0, 1); U2 = (1, 1, 0); U3 = (0, 1, 1); U4 = (1, 1, 1); After splicing each candidate action vector with Lt5 and inputting it into the capacitor state prediction model to predict the next-cycle capacitor state vector, the next capacitor 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 capacitor 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 capacitor 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 capacitor state vector corresponding to U4 is (0.13, 0.66, 49.1, 0.15, 0.63, 50.2, 0.14, 0.66, 47.8); 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 obtained is (U1, U4, U2, U3).

[0024] Step S600: Input the real-time capacitance state, power value, predicted power, and each candidate action vector into the digital twin system for simulation to obtain prediction metrics. Screen the characteristic actions based on the threshold, determine the final adjustment instruction, and adjust the capacitance unit according to the final adjustment instruction. Among them, step S600 includes: Step S601: Combine the real-time capacitance state vector, real-time current power value, and predicted value of future power corresponding to the nth sampling period with each candidate action vector in the candidate action vector set respectively, input them into the digital twin system, perform simulation operations, and obtain the predicted number of voltage dips, predicted response delay, and predicted capacitance temperature corresponding to each candidate action vector. Step S602: Preset a response delay threshold and a capacitance temperature threshold. If a certain candidate action vector meets all the following conditions: the predicted number of voltage dips is 0; the predicted response delay is less than the response delay threshold; the predicted capacitance temperature is less than the capacitance temperature threshold; then set the candidate action vector as the characteristic action vector. Step S603: Extract the order of each characteristic action vector in the candidate action vector set, select the characteristic action vector with the smallest order as the final adjustment instruction, and input it into the terminal system to adjust the capacitance unit. For example, combine Lt5, the current power value, the predicted value of future power with each candidate action vector, input them into the digital twin system for simulation, and obtain the following prediction metrics as Table 3;

[0025] Table 3 The preset response delay threshold is 0.20 s, the capacitance temperature threshold is 50.0, U1 meets all the adjustments, and the final adjustment instruction is U1=(1, 0, 1).

[0026] To better implement the above method, a full-pole ear capacitance regulation system for an AI server is also proposed. The system includes a predicted operation level module, a future power module, a capacitance state module, a real-time prediction module, a candidate action vector module, and a real-time adjustment module; Predicted operation level module: Receive the information content uploaded by the user in the AI server terminal system, set the sampling period, collect the server power parameters and record the capacitance adjustment parameters, extract the characteristic parameters of the information content and normalize them to construct a characteristic vector, calculate the operation score, divide the operation level, and use the characteristic vector and the operation level to construct a training sample and train a random forest model to predict the operation level; Among them, the predicted operation level module includes a calculation operation score unit and an establishment operation level prediction model unit: Calculation and operation scoring unit: In the terminal system of the AI server, receive the information content uploaded by the user terminal, preset the sampling period, monitor the power parameters of the server during the sampling period, record the capacitance adjustment parameters, generate a sampling record for executing the information content, collect a certain piece of information content uploaded by the user terminal, extract the content type of the information content, determine the characteristic parameters corresponding to the content type, calculate the characteristic values of each characteristic parameter in the query content, and perform normalization processing to construct a content feature vector. In the historical sampling record set of the information content, collect each power response data of a certain historical sampling record and perform normalization processing to calculate the operation score; Establishment of operation level prediction model unit: Preset several operation levels and set the average operation score range corresponding to each operation level. Automatically divide each piece of information content into the corresponding operation level. In the information content set of a certain content type, obtain the content feature vector and operation level of each piece of information content, 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, where the content feature vector is used as the input and the operation level label is used as the output to establish the operation level prediction model corresponding to the content type.

[0027] Future power module: Set the training period, extract the information content feature vector and input it into the operation level prediction model, generate a structural context vector in combination with the code analysis tool, construct an operation state matrix, calculate the future power, form a training sample triple, construct a dual-channel neural network fusion based on LSTM and GRU to extract features, and predict the future power through a mixture density network, and optimize the model using maximum likelihood estimation; Among them, the future power module includes a training sample triple construction unit and a future power prediction model establishment unit: Construct a training sample triple unit: Select several consecutive days as the training period. During the training period, collect a certain piece of information content uploaded by the user terminal, extract the content type of the information content, construct the content feature vector of the information content, and input it into the running level prediction model corresponding to the content type to obtain the running level label of the information content. Through the static code analysis tool built into the terminal system, extract the task computing intensity and memory bandwidth requirements of the information content, and perform normalization processing to form the structural context vector of the information content. Obtain the sampling record set for executing the information content, collect the average current, power change rate, and load fluctuation coefficient in each sampling record, and form the running state vector of the sampling record. Arrange all the running state vectors in the order of sampling time to form the running state matrix of the information content. When the information content is executed, preset the number of future sampling periods, collect the power values corresponding to each future sampling period, calculate the future power, and combine the running state matrix, structural context vector, running level label, and future power of the information content to construct a training sample triple; Establish a future power prediction model unit: Based on the training sample triple, construct a dual-channel neural network model for future power prediction. Use the running state matrix in the training sample triple as the first channel, and extract the state vector representing the running time series fluctuation characteristics through the LSTM network. Use the structural context vector and running level label in the training sample triple as the second channel, and 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. Adopt the mixture density network MDN structure to output the future power. Obtain all the training sample triples within the training period, use the 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, and perform model training to obtain the prediction model of the future power.

[0028] Capacitor state module: 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 on states to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, construct a digital twin system for capacitor behavior simulation, and output multi-dimensional operation indicators; Real-time prediction module: Real-time extract the feature of the information content uploaded by the user, construct a state matrix, a structural context, and a running level label, and input them into the prediction model to obtain the future power value; Candidate action vector module: Collect the capacitor state vector of the target sampling period, combine multiple candidate action vectors to predict the future capacitor state, calculate the cost score, and sort the candidate action vectors according to the cost score to obtain a set of candidate action vectors; Real-time adjustment module: Input the real-time capacitance state, power value, predicted power, and each candidate action vector into the digital twin system for simulation to obtain predicted metrics. Screen the characteristic actions according to the threshold, determine the final adjustment instruction, and adjust the capacitance unit according to the final adjustment instruction.

[0029] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A method for regulating all-tab capacitors of an AI server, characterized in that, The method includes: Step S100: Receive the information content uploaded by the user in the AI server terminal system, set the sampling period, collect the server power parameters and record the capacitance adjustment parameters, extract the feature parameters of the information content and normalize them to construct a feature vector, calculate the operation score, divide the operation level, and use the feature vector and the operation level to construct a training sample and train a random forest model to predict the operation level; Step S200: Set the training period, extract the information content feature vector and input it into the operation level prediction model, generate a structure context vector in combination with the code analysis tool, construct an operation state matrix, calculate the future power, form a training sample triple, construct a dual-channel neural network fusion based on LSTM and GRU to extract features, and predict the future power through a mixture density network, and optimize the model using maximum likelihood estimation; 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 on states to form an action vector, generate a capacitor state training set and train a capacitor state prediction model, construct a digital twin system for capacitor behavior simulation, and output multi-dimensional operation indicators; Step S400: Extract the features of the user-uploaded information content in real time, construct a state matrix, a structure context, and an operation level label, and input them into the prediction model to obtain the future power value; Step S500: Collect the capacitor state vector of the target sampling period, combine multiple candidate action vectors to predict the future capacitor state, calculate the cost score, sort the candidate action vectors according to the cost score, and 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 for simulation, obtain the prediction index, screen the feature actions according to the threshold, determine the final adjustment instruction, and adjust the capacitor unit according to the final adjustment instruction.

2. The all-tab capacitor regulation method 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, receive the information content uploaded by the user terminal, preset the sampling period, monitor the power parameters of the server during the sampling period, record the capacitance adjustment parameters, and generate a sampling record for executing the information content; Step S102: Collect a certain piece of information content uploaded by the user terminal, extract the content type of the information content, determine the feature parameters corresponding to the content type, calculate the feature values of each feature parameter in the query content, and perform normalization processing to construct a content feature vector as A=(a1,a2,...ab), where a1,a2,...ab respectively represent the normalized values of the 1st, 2nd,...bth feature parameters; Step S103: In the historical sampling record set of the information content, collect each power response data of a certain historical sampling record, and perform normalization processing, and calculate the operation score according to the following formula: ； Among them, B represents the running score, and D c represents the normalized value of the c-th power response data, and E c represents the weight value of the c-th power response data. Summarize all historical sampling records and calculate the average value of the running score; Step S104: Preset several operation levels, and set the average operation score range corresponding to each operation level, and automatically divide each piece of information content into the corresponding operation level; Step S105: In the information content set of a certain content type, obtain the content feature vector and running level of each piece of information content. Use the running level as a label and jointly construct a training sample set with the content feature vector. Input the training sample set into a random forest model for training, where the content feature vector is used as the input and the running level label is used as the output, to establish a running level prediction model corresponding to the content type.

3. A full-tab capacitor regulation method for an AI server according to claim 2, characterized in that The said step S200 includes the following steps: Step S201: Select a continuous number of days as the training period. During the training period, collect a piece of information content uploaded by the user terminal, extract the content type of the information content, construct the content feature vector of the information content, and input it into the running level prediction model corresponding to the content type to obtain the running level label of the information content; Step S202: Through the static code analysis tool built into the terminal system, extract the task computational intensity and memory bandwidth requirements of the information content, and perform normalization processing to form the structural context vector of the information content; Step S203: Obtain the sampling record set for executing the information content, collect the average current, power change rate, and load fluctuation coefficient in each sampling record, form the running state vector of the sampling record, and arrange all the running state vectors in the order of sampling time to form the running state matrix of the information content; Step S204: When the execution of the information content is completed, preset the number of future sampling periods, collect the power values corresponding to each future sampling period, and calculate the future power according to the following formula: ； Among them, F represents the future power, and G e represents the power value corresponding to the e-th future sampling period, f represents the number of future sampling periods, and the operation state matrix, structural context vector, operation level label, and future power of the information content are combined to construct a training sample triple as S = [Y, (J, X), F], where Y represents the operation state matrix, J represents the structural context vector, X represents the operation level label, and F represents the future power; Step S205: Based on the training sample triple, construct a dual-channel neural network model for future power prediction. Use the running state matrix in the training sample triple as the first channel, and extract the state vector representing the running time series fluctuation characteristics through the LSTM network. Use the structural context vector and running level label in the training sample triple as the second channel, and 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, and use the mixture density network MDN structure to output the future power; Step S206: Obtain all the training sample triples within the training period, use the 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, and perform model training to obtain the prediction model of the future power.

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

5. A full-tab capacitor regulation method for an AI server according to claim 4, characterized in that The step S400 includes the following steps: Step S401: Obtain the real-time information content uploaded by the user side, extract the content type of the real-time information content, construct the content feature vector of the real-time information content, and input it into the operation level prediction model corresponding to the content type to obtain the operation level label of the real-time information content; Step S402: Extract the task computational intensity and memory bandwidth requirements of the real-time information content through the static code analysis tool built into the terminal system, and perform normalization processing to form the structural context vector of the real-time information content; Step S403: Preset the number of sampling periods for analysis and optimization as n, obtain the sampling period set of the real-time information content, and construct the operation state matrix of the real-time information content; Step S404: Input the operation state matrix, structural context vector, and operation level label of the real-time information content into the prediction model of future power to obtain the predicted value of future power.

6. The all-terminal ear capacitance regulation method for an AI server according to claim 5, characterized in that, The step S500 includes the following steps: Step S501: Obtain the time point corresponding to the nth sampling period for analysis and optimization, and collect the real-time state variables of each capacitance unit of the terminal system corresponding to the time point to obtain the real-time capacitance state vector; Step S502: Preset a number of candidate action vectors, input the real-time capacitance state vector and each candidate action vector into the prediction model of the capacitance state to predict the next capacitance state vector; Step S503: Extract the equivalent series resistance as R and the state of charge as H in a certain 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 the equivalent series resistance, and r1 and r2 represent the weights of the equivalent series resistance and the state of charge respectively; Step S504: Summarize the cost scores corresponding to all candidate action vectors, arrange all candidate action vectors in ascending order of cost score to obtain a set of candidate action vectors.

7. A full-tab capacitor regulation method for an AI server according to claim 6, characterized in that, The step S600 includes the following steps: Step S601: Combine the real-time capacitance state vector, real-time current power value, and predicted value of future power corresponding to the time point of the nth sampling period with each candidate action vector in the set of candidate action vectors respectively, input them into the digital twin system, perform a simulation operation, and obtain the predicted number of voltage dips, predicted response delay, and predicted capacitance temperature corresponding to each candidate action vector; Step S602: Preset a response delay threshold and a capacitor temperature threshold. If a certain candidate action vector satisfies all of the following conditions: the predicted number of voltage dips is 0; the predicted response delay is less than the response delay threshold; the predicted capacitor temperature is less than the capacitor temperature threshold; then set the candidate action vector as the characteristic action vector; Step S603: Extract the rank of each characteristic action vector in the candidate action vector set, select the characteristic action vector with the smallest rank as the final adjustment instruction, and input it into the terminal system to adjust the capacitor unit.

8. A full-tab capacitor regulation system for an AI server, which is used to implement the full-tab capacitor regulation method for an AI server according to any one of claims 1-7, characterized in that, The system includes a predicted operation 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 predicted operation level module: receives the information content uploaded by the user in the AI server terminal system, sets the sampling period, collects the server power parameters and records the capacitor adjustment parameters, extracts the characteristic parameters of the information content and normalizes them to construct a characteristic vector, calculates the operation score, divides the operation level, constructs a training sample using the characteristic vector and the operation level, and trains a random forest model to predict the operation level; The future power module: sets the training period, extracts the information content characteristic vector and inputs it into the operation level prediction model, generates a structural context vector in combination with the code analysis tool, constructs an operation state matrix, calculates the future power, forms a training sample triple, constructs a dual-channel neural network fusion based on LSTM and GRU to extract features, and predicts the future power through a mixture density network, and optimizes the model using maximum likelihood estimation; 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 on states to form an action vector, generates a capacitor state training set and trains a capacitor state prediction model, constructs a digital twin system for capacitor behavior simulation, and outputs multi-dimensional operation indicators; The real-time prediction module: extracts the characteristics of the information content uploaded by the user in real time, constructs a state matrix, a structural context, and an operation level label, and inputs them into the prediction model to obtain the future power value; The candidate action vector module: collects the capacitor state vector of the target sampling period, combines multiple candidate action vectors to predict the future capacitor state, calculates the cost score, sorts the candidate action vectors according to the cost score, and obtains 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 for simulation, obtains the predicted indicators, screens the characteristic actions according to the threshold, determines the final adjustment instruction, and adjusts the capacitor unit according to the final adjustment instruction.

9. The all-pole ear capacitance regulation system for an AI server according to claim 8, wherein The predicted operation level module includes an operation score calculation unit and an operation level prediction model establishment unit: The computing operation scoring unit: In the terminal system of the AI server, it 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 piece of 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 values of each characteristic parameter in the query content, and performs normalization processing to construct a content feature vector. In the historical sampling record set of the information content, it collects each power response data of a certain historical sampling record and performs normalization processing to calculate the operation score; The unit for establishing an operation level prediction model: Presets several operation levels and sets the average operation score range corresponding to each operation level, automatically divides each piece of information content into the corresponding operation level. In the set of information content of a certain content type, it obtains the content feature vector and operation level of each piece of information content, uses the operation level as a label, and jointly constructs a training sample set with the content feature vector. Inputs the training sample set into a random forest model for training, where the content feature vector is used as the input and the operation level label is used as the output to establish the operation level prediction model corresponding to the content type.

10. A full-tab capacitor regulation system for an 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 unit for constructing a training sample triple: Selects several consecutive days as the training period. During the training period, it collects a certain piece of information content uploaded by the user terminal, extracts the content type of the information content, constructs the content feature vector of the information content, and inputs it into the operation level prediction model corresponding to the content type to obtain the operation level label of the information content. Through the built-in static code analysis tool of the terminal system, it extracts the task computing intensity and memory bandwidth requirements of the information content and performs normalization processing to form the structural context vector of the information content. It obtains the sampling record set for executing the information content, collects the average current, power change rate, and load fluctuation coefficient in each sampling record, and forms the operation state vector of the sampling record. Arranges all the operation state vectors in the order of sampling time to form the operation state matrix of the information content. When the information content is executed, it presets the number of future sampling periods, collects the power values corresponding to each future sampling period, calculates the future power, and combines the operation state matrix, structural context vector, operation level label, and future power of the information content to construct a training sample triple; The unit for establishing the future power prediction model: Based on the training sample triples, construct a dual-channel neural network model for future power prediction. Take the operation state matrix in the training sample triples as the first channel, and extract the state vector representing the operation time series fluctuation characteristics through the LSTM network. Take the structural context vector and the operation level label in the training sample triples as the second channel, and 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. Adopt the mixture density network (MDN) structure to output the future power. Obtain all the training sample triples within the training period, use the 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, conduct model training, and obtain the prediction model of the future power.

Images