Communication equipment optimization method and device based on power management

Abstract

The present invention relates to a communication equipment optimization method and device based on power management. The method: performs delayed power domain transformation and minimum mean square error power balancing processing on the power consumption data of the communication equipment to obtain power consumption characteristic data; inputs a heterogeneous device management neural network to perform power management prediction to obtain equipment power consumption prediction parameters; deploys a virtual power manager and performs power response control on the communication base station and the backup power supply to obtain power regulation parameters; constructs a device power consumption scheduling matrix, and uses a block iteration method to perform balanced optimization on the device power consumption scheduling matrix to obtain a multi-device power consumption scheduling scheme; performs power allocation control according to the multi-device power consumption scheduling scheme, generates an initial power allocation strategy, and dynamically adjusts the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy. The present invention improves the accuracy and adaptability of power allocation.

Description

Technical Field

[0001] The present invention relates to the technical field of power management, and in particular to a communication equipment optimization method and device based on power management. Background Art

[0002] Traditional communication equipment power management methods mainly rely on fixed power allocation strategies and simple threshold control, which are difficult to adapt to dynamic power consumption changes under complex load environments, resulting in low energy utilization efficiency of equipment and high operating costs.

[0003] In practical applications, due to the obvious time-varying and uncertain nature of the load state and power consumption characteristics of communication equipment, traditional power management methods often cannot accurately capture the power consumption variation patterns of equipment, making it difficult to achieve precise power allocation and scheduling control. In addition, existing power management systems lack effective support for multi-device collaborative optimization and cannot fully utilize the power consumption complementarity characteristics between devices, affecting the energy efficiency of the overall system. Summary of the invention

[0004] The main purpose of the present invention is to provide a communication equipment optimization method and device based on power management, which improves the accuracy and adaptability of power allocation.

[0005] To achieve the above object, the present invention provides a communication device optimization method based on power management, comprising the following steps:

[0006] Performing delayed power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data;

[0007] Inputting the power consumption characteristic data into a heterogeneous device management neural network for power management prediction to obtain device power consumption prediction parameters;

[0008] Deploy a virtual power manager based on the device power consumption prediction parameters, and perform power response control on the communication base station and the backup power supply through the virtual power manager to obtain power regulation parameters;

[0009] Building a device power consumption scheduling matrix based on the power supply adjustment parameters, and optimizing the device power consumption scheduling matrix in a block-by-block iterative manner to obtain a multi-device power consumption scheduling solution;

[0010] Power allocation control is performed according to the multi-device power consumption scheduling scheme to generate an initial power allocation strategy, and the initial power allocation strategy is dynamically adjusted through a three-layer Bayesian network model to obtain a target power allocation strategy.

[0011] The present invention also provides a communication equipment optimization device based on power management, comprising:

[0012] A transformation module, used to perform delayed power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data;

[0013] A prediction module, used for inputting the power consumption characteristic data into a heterogeneous device management neural network to perform power management prediction and obtain device power consumption prediction parameters;

[0014] A control module, configured to deploy a virtual power manager based on the device power consumption prediction parameter, and to perform power response control on the communication base station and the backup power supply through the virtual power manager to obtain a power regulation parameter;

[0015] A construction module is used to construct a device power consumption scheduling matrix based on the power supply adjustment parameters, and to perform balanced optimization on the device power consumption scheduling matrix in a block iterative manner to obtain a multi-device power consumption scheduling solution;

[0016] The adjustment module is used to perform power allocation control according to the multi-device power consumption scheduling scheme, generate an initial power allocation strategy, and dynamically adjust the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy.

[0017] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0019] In summary, the technical solution provided by the present invention can accurately extract the power consumption characteristics of the device through delayed power consumption domain transformation and minimum mean square error power balancing processing, adopt heterogeneous device management neural network to perform power management prediction, and combine multi-layer feature extraction and timing analysis to improve the accuracy of power consumption prediction; based on the virtual power manager to perform power response control, realize the coordinated management of communication base stations and backup power supplies, and enhance the power supply stability of the system; optimize the power consumption scheduling matrix through block iteration, optimize the power consumption scheduling strategy of multiple devices, and reduce the overall energy consumption of the system; use a three-layer Bayesian network model to dynamically adjust the power allocation strategy, thereby improving the accuracy and adaptability of power allocation; use multi-layer probability distribution estimation and variational inference technology to enhance the system's modeling ability and prediction accuracy for power consumption fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the steps of a communication device optimization method based on power management in one embodiment of the present invention;

[0021] Figure 2 It is a structural block diagram of a communication equipment optimization device based on power management in one embodiment of the present invention;

[0022] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0023] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] Reference Figure 1 This embodiment provides a communication device optimization method based on power management, comprising the following steps:

[0026] S1, performing delay power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data;

[0027] Among them, comprehensive data collection is performed on the operating status of the communication equipment, including relevant parameters such as operating current, operating voltage, operating temperature, load state, operating time and standby time, which together constitute the original power consumption data set of the equipment. The delayed power consumption domain transformation is performed on the original power consumption data, and the original power consumption data set is transformed into the delayed power consumption feature space through certain mapping rules to better extract the time dependence and change characteristics in the power consumption data. Through the delayed power consumption domain transformation, the delayed power consumption feature matrix is ​​obtained to describe the change law of the power consumption characteristics of the equipment at different time points, and capture the dynamic characteristics of the power consumption of the equipment over time and operating conditions. Based on the delayed power consumption feature matrix, the first power balancing matrix with the minimum mean square error of the device level is constructed, and the noise and deviation in the power consumption characteristics are eliminated by minimizing the mean square error, so as to obtain a more accurate and stable description of the power consumption characteristics. On this basis, the power consumption characteristics of the first power balancing matrix are divided, and the matrix is ​​split into several power consumption characteristic sub-matrix groups, and the power consumption characteristics are refined and analyzed to facilitate subsequent processing and calculation. Through the power consumption characteristic sub-matrix group, the characteristic differences of the equipment under different power consumption states are captured, and the precision and flexibility of optimization are improved. Iterative calculation and matrix dimension reduction are performed based on the power consumption characteristic sub-matrix group to obtain the second power consumption balancing matrix. Iterative calculation is to gradually approach the optimal power balance state, while matrix dimension reduction is to reduce the amount of data and complexity, making subsequent analysis and processing more efficient. Discrete Fourier transform is performed on the second power balance matrix to convert power consumption data from the time domain to the frequency domain. Frequency domain analysis helps to reveal the periodic change characteristics of device power consumption and the impact of different frequency components on overall power consumption. The frequency domain power consumption characteristics obtained by discrete Fourier transform help to extract the power consumption fluctuation parameters and load change parameters of the device under different load conditions, describe the power consumption changes of the device under different load conditions, and the response characteristics of the device power consumption when the load fluctuates. Based on the frequency domain power consumption characteristics, the power management characteristics are fused to obtain the power consumption feature vector of the device. The power consumption feature vector combines multiple power consumption feature parameters and can comprehensively describe the power consumption behavior characteristics of the device under different states. The power consumption feature vector is subjected to feature dimensionality reduction and normalization processing. The purpose of feature dimensionality reduction is to remove redundant information and reduce the complexity of data, while the purpose of normalization processing is to adjust the eigenvalues ​​to a relatively uniform scale to facilitate subsequent analysis and modeling. After feature dimensionality reduction and normalization processing, power consumption feature data is obtained.

[0028] S2, inputting the power consumption characteristic data into the heterogeneous device management neural network for power management prediction to obtain device power consumption prediction parameters;

[0029] Specifically, the power consumption feature data is input into the device feature encoding layer of the heterogeneous device management neural network. The feature encoding layer consists of four heterogeneous feature extraction units, each of which includes three layers of one-dimensional convolutional networks and two layers of fully connected networks. Through the three layers of one-dimensional convolutional networks, the device feature encoding layer can extract the local pattern and change trend of the power consumption feature data, and the two layers of fully connected networks integrate these extracted local features to form a higher-level feature expression to obtain the feature encoding of the device. The device feature encoding is input into the timing analysis layer of the heterogeneous device management neural network for processing to capture the time dependency and dynamic change characteristics of the device power consumption characteristics. The timing analysis layer contains two main structures: a bidirectional gated recurrent unit and a self-attention mechanism. Among them, the bidirectional gated recurrent unit is set with 128 hidden units, which can share information between the forward and backward directions and better capture the characteristics of the device power consumption changing over time. The self-attention mechanism contains 8 attention heads. By calculating the correlation between features at different time steps, the weight of the features is automatically adjusted, so as to more flexibly focus on the important feature changes in the power consumption data and obtain the timing feature map. The timing feature map is input into the multimodal fusion layer of the heterogeneous device management neural network for processing. The multimodal fusion layer fuses multiple types of features to comprehensively consider the impact of different modal features on power consumption prediction. The multimodal fusion layer contains three parallel feature conversion branches, each of which includes two layers of fully connected networks and one normalization layer. These feature conversion branches can convert and standardize the input time series feature map in different ways to obtain a more balanced and unified fusion feature vector. The multi-branch parallel processing method enables the device management neural network to model features from different angles, improving the effectiveness and diversity of feature fusion. The fused feature vector is input to the power generator of the heterogeneous device management neural network for processing. The power generator adopts a deep residual network structure. The deep residual network contains six residual blocks, each of which consists of two layers of convolution and a jump connection. The advantage of this structure is that it can effectively alleviate the gradient vanishing problem during deep training of neural networks. Through jump connections, the input features can be directly passed to subsequent layers, realizing effective retention and gradual optimization of features. Through multi-layer convolution processing of residual blocks, the power generator can deeply model the fused features, extract more representative features, obtain power consumption prediction sequences, and reflect the power consumption change trend of the device at different time points. The power consumption prediction sequence is input into the discriminator network of the heterogeneous device management neural network for processing. The discriminator network consists of four layers of progressive discriminant units, each of which contains a convolutional layer and a pooling layer. This structure can gradually compress the dimensions of the input features and extract important information from the features, thereby evaluating the credibility of the power consumption prediction sequence. Through feature extraction in the convolutional layer and feature compression in the pooling layer, the discriminator network can effectively distinguish the rationality of the power consumption prediction sequence and give an output of prediction credibility.Based on the prediction credibility of the discriminator network, the power consumption prediction sequence is input into the parameter generation network of the heterogeneous device management neural network for parameter generation. The parameter generation network contains three fully connected layers and two normalization layers. After processing by these layers, the various parameters of the device power consumption are finally normalized and output through the softmax activation function to obtain the power consumption prediction parameters of the device, which reflects the power consumption requirements of the device under different conditions.

[0030] S3, deploying a virtual power manager based on the device power consumption prediction parameters, and performing power response control on the communication base station and the backup power supply through the virtual power manager to obtain power regulation parameters;

[0031] It should be noted that the power response model of the virtual power manager is constructed based on the equipment power consumption prediction parameters. The equipment power consumption prediction parameters are divided into a base station power consumption response matrix and a backup power supply response matrix. According to the different characteristics of the base station and the backup power supply, the corresponding power supply response strategies are designed respectively. The power control calculation is performed on the base station power consumption response matrix, the power allocation weight and the power supply response threshold of the base station are determined, and the power control strategy of the base station is obtained. The power control calculation process comprehensively considers the power demand changes of the communication base station under different load conditions and the maximum load capacity of the equipment, so as to ensure that the base station minimizes power consumption while meeting the communication needs. The determination of the allocation weight and the response threshold enables the base station to flexibly adjust the power consumption under various working conditions, thereby avoiding energy waste and improving power utilization efficiency. At the same time, dynamic programming calculation is performed on the backup power supply response matrix to solve the optimal switching timing of the backup power supply. Comprehensively considering factors such as the remaining power of the backup power supply, the current load demand and the future power demand, the optimal backup power supply switching time is found to achieve efficient utilization of the backup power supply. The resulting backup power supply power control strategy ensures that the backup power supply can intervene in time when the system needs it and remain on standby when it is not needed, thereby extending the service life of the backup power supply and improving the power guarantee capability. Power response optimization is performed based on the base station power control strategy and the backup power supply power control strategy. The power management between the base station and the backup power supply is coordinated to ensure that the overall power consumption of the system is maintained at the optimal level while meeting the communication requirements. Through optimization calculation, the power response optimization parameters are obtained to guide the coordinated work of the base station and the backup power supply to achieve power management optimization of the entire communication system. Based on the power response optimization parameters, a control strategy model of the virtual power manager is constructed. The model integrates the power response characteristics of the base station and the backup power supply and globally controls the power allocation of the entire system. The control strategy model is deployed to the communication base station and the backup power supply through the real-time scheduling interface of the virtual power manager. The deployment process generates a power response execution instruction, which includes control information such as power allocation, response time, and switching rules for the base station and the backup power supply to ensure that the cooperation between the two can achieve the best in actual operation. The communication base station and the backup power supply make real-time adjustments based on the received power response execution instructions to make the power consumption response of the system more flexible and efficient. In the process of executing the power response execution instruction, the virtual power manager collects the real-time power consumption fluctuation data of the communication base station and the switching status data of the backup power supply. The power consumption fluctuation data can reflect the current load changes of the base station, while the backup power supply switching status data shows the working status and switching frequency of the backup power supply. By analyzing these real-time data, the virtual power manager can dynamically correct the power response optimization parameters, and promptly correct the errors caused by changes in the external environment or fluctuations in the internal state of the system, ensuring that power management is always in the optimal state.By dynamically correcting the power response optimization parameters, the final power regulation parameters are obtained, which are used to continuously adjust and optimize the power configuration of the communication base station and backup power supply, so that the power consumption of the entire communication equipment is always maintained at the optimal level under different load and environmental conditions.

[0032] S4, constructing a device power consumption scheduling matrix based on the power supply adjustment parameters, and optimizing the device power consumption scheduling matrix in a balanced manner using a block iteration method to obtain a multi-device power consumption scheduling solution;

[0033] Specifically, the power adjustment parameters are divided into time series. The power consumption data is classified into the prediction stage and the real-time stage to obtain the power consumption scheduling data of the two stages. The requirements of the device in the future power consumption prediction and the dynamic response in the actual operation are considered respectively to ensure that the scheduling scheme can not only foresee the future power consumption requirements of the device, but also respond quickly to the power consumption changes in real time. The device power consumption scheduling matrix is ​​constructed based on the two-stage power consumption scheduling data. The device power consumption scheduling matrix contains the device power allocation coefficient and the load state parameter, which is used to describe the power demand and load change of each device in different time periods. By constructing the scheduling matrix, the power consumption of the device is systematically described. The device power consumption scheduling matrix can uniformly represent the power demand of different devices in different time periods. The device power consumption scheduling matrix is ​​processed in blocks. The scheduling matrix is ​​divided into multiple sub-scheduling matrices according to the time dimension and the device type. Each sub-scheduling matrix corresponds to a power consumption allocation strategy for a specific scheduling period, and a scheduling sub-matrix group is obtained. Block processing can effectively reduce the computational complexity, making the power scheduling of each device more flexible and detailed. The scheduling sub-matrix group is iteratively optimized and calculated, and the power allocation coefficient of each sub-scheduling matrix is ​​gradually adjusted through multiple iterations until a local optimal state is reached. Iterative optimization calculation ensures that each sub-scheduling matrix can maximize power consumption, reduce unnecessary energy waste, and keep the device power allocation at the optimal level while meeting the load demand. The target scheduling sub-matrix is ​​obtained by adjusting the power allocation coefficient of each sub-scheduling matrix. The target scheduling sub-matrix is ​​reorganized according to the time series to construct the device scheduling timing plan. The device scheduling timing plan is an overall scheduling strategy that describes the power consumption allocation requirements of different devices in different time periods. In order to improve the balance of power consumption management, the device scheduling timing plan is processed for power consumption balance to eliminate the large power consumption differences between devices, so that the power consumption distribution in the system is uniform. Device power consumption control instructions are generated based on the balanced scheduling plan. In order to meet the different control requirements of the prediction stage and the real-time stage, these device power consumption control instructions are classified to obtain a staged control instruction set. The control instructions in the prediction stage are used to pre-configure the future power consumption of the device, while the control instructions in the real-time stage are used to dynamically adjust the device power during actual operation to cope with immediate changes in the load. The instruction set is organized in time and prioritized to obtain a multi-device power consumption scheduling plan. Timing organization is to ensure that different control instructions are executed at the correct time, while priority sorting is to ensure that the most important tasks are executed first when multiple tasks compete, so as to ensure the stability and efficiency of the system.

[0034] S5, executes power allocation control according to the multi-device power consumption scheduling scheme, generates an initial power allocation strategy, and dynamically adjusts the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy.

[0035] Among them, power allocation control is performed based on the multi-device power consumption scheduling scheme to generate an initial power allocation strategy. The initial power allocation strategy is based on the device power demand and load status, and effectively allocates the available power in the system by scheduling the power consumption of the device. The initial power allocation strategy is input into the first layer of the three-layer Bayesian network model, namely the power-aware network. The structure of the power-aware network includes four parallel power feature extraction units, each of which consists of three convolutional layers and two fully connected layers. These feature extraction units extract the power features implicit in the initial power allocation strategy, extract spatial relationship features through the convolutional layer, and then integrate features through the fully connected layer to obtain a power feature vector containing multiple power information. The power feature vector is input into the second layer of the three-layer Bayesian network model, namely the prior distribution network. The prior distribution network contains three probability distribution estimation modules, each of which consists of a variational inference layer and a Gaussian mixture model, and estimates the power distribution parameters based on the input power feature vector. The variational inference layer is used to estimate the prior probability of power distribution and establish the probabilistic relationship between the power requirements of different devices, while the Gaussian mixture model refines the prior distribution to ensure the modeling of complex power distribution relationships. This process enables the model to infer the probability distribution of power requirements of each device at different times based on the power feature vector and generate power distribution parameters. The power distribution parameters are input into the third layer of the three-layer Bayesian network model, namely the posterior inference network for dynamic update. The posterior inference network of the third layer contains the Markov chain Monte Carlo sampler and the variational posterior inference, which work together to dynamically sample and update the power distribution parameters. The Markov chain Monte Carlo sampler samples the power distribution parameters by simulating random processes to obtain more possible situations about power requirements, while the variational posterior inference uses these sampling results, combined with prior information and observation data, to perform a posteriori estimation of the power distribution parameters to obtain the most likely power requirements. Through the dynamic update process, the power adjustment parameters are obtained to reflect the power change requirements of the equipment under different environments and load conditions, so that the power allocation can be more adapted to the changes in actual conditions. Based on the power adjustment parameters, the initial power allocation strategy is updated. The parameter update process is carried out through the Bayesian update rule, which is a method for updating probability distribution based on prior information and new observation data to calculate the optimal power allocation value. Through the Bayesian update rule, the feature extraction results of the power-aware network, the probability estimation of the prior distribution network, and the dynamic adjustment results of the posterior inference network are comprehensively considered to finally obtain the optimal power allocation strategy, that is, the target power allocation strategy.

[0036] The power feature vector is processed by data grouping. The original power feature vector is classified according to the power change characteristics, load fluctuation characteristics and equipment response characteristics, and divided into three groups of 128-dimensional feature vectors to obtain the power feature data group. The purpose of grouping the feature vector is to refine the expression of the power feature so as to adopt a targeted probability modeling method for power consumption data with different characteristics. The power feature data group is input into the first probability distribution estimation module in the second layer prior distribution network, which includes a variational inference layer and a Gaussian mixture model. The power feature data is input into the variational inference layer, which consists of three layers of feedforward neural network, each layer contains 256 neurons, and uses the ReLU activation function to effectively extract the complex nonlinear relationship in the power feature. By maximizing the lower limit of variational evidence, the variational inference layer can approximate the data features and obtain the first prior probability distribution. By means of variational inference, the power feature is modeled without relying on the precise probability distribution function, thereby improving the ability to understand the data and the accuracy of prediction. The first prior probability distribution obtained is modeled by the Gaussian mixture model of the first probability distribution estimation module. The Gaussian mixture model uses 8 mixture components, and the mean vector and covariance matrix of each component are iteratively calculated through the expectation maximization algorithm to obtain the first power distribution parameters. The expectation maximization algorithm gradually optimizes the parameters of each Gaussian component by continuously alternating between the expectation step and the maximization step, so that the Gaussian mixture model can fit the data distribution well and obtain a more accurate description of the power characteristics. The power characteristic data group is input into the second probability distribution estimation module in the second layer of the prior distribution network. The variational inference layer of this module constructs a load fluctuation probability model through a four-layer variational autoencoder. The variational autoencoder is a generative model that can learn data latent variables. Its encoding dimension is 64, while the latent variable dimension is 32. By using the variational autoencoder, the complexity of the data is effectively reduced and the potential features in the load fluctuation are captured. In order to improve the modeling accuracy of the power characteristics of the model, six groups of Gaussian mixture components are used to model the latent variables to obtain the second power distribution parameters. The combination of the variational autoencoder and the Gaussian mixture model helps to improve the modeling ability of the load fluctuation characteristics, and can more accurately predict the change trend of power consumption under complex load changes. The power feature data set is input into the third probability distribution estimation module in the second layer of prior distribution network, which establishes a device response probability model through a variational inference layer containing two layers of stacked variational recurrent neural networks. Each layer of variational recurrent neural network contains 128 LSTM units, which can capture the timing dependency characteristics in the device response and provide a reliable basis for the dynamic adjustment of power scheduling. When performing variational inference on the power feature, four groups of Gaussian mixture components are used to model it to obtain the third power distribution parameters.The variational recurrent neural network can effectively cope with the power demand changes caused by environmental changes and load fluctuations in actual use of the equipment by learning the response characteristics of the equipment in different time periods, making power scheduling more flexible and accurate. The first power distribution parameter, the second power distribution parameter and the third power distribution parameter are weighted and fused to form an overall power distribution description.

[0037] In one example, delay power consumption domain transformation and minimum mean square error power equalization processing are performed on power consumption data of a communication device to obtain power consumption characteristic data, including:

[0038] Collect data on the operating current, operating voltage, operating temperature, load status, operating time, and standby time of the communication equipment to obtain the original power consumption data set;

[0039] Performing delay power consumption domain transformation on the original power consumption data set, mapping the original power consumption data set to the delay power consumption feature space, and obtaining a delay power consumption feature matrix;

[0040] Constructing a first power balancing matrix with minimum mean square error at the device level based on the delay power consumption characteristic matrix, and dividing the first power balancing matrix by power consumption characteristics to obtain a power consumption characteristic sub-matrix group;

[0041] Perform iterative calculation and matrix dimension reduction based on the power consumption characteristic sub-matrix group to obtain a second power consumption balancing matrix;

[0042] Performing discrete Fourier transform on the second power consumption balancing matrix to obtain frequency domain power consumption characteristics, and extracting power consumption fluctuation parameters and load change parameters of the device under different load states according to the frequency domain power consumption characteristics to obtain dynamic power consumption characteristics;

[0043] Power management features are fused based on dynamic power consumption features to obtain device power consumption feature vectors, and feature dimension reduction and normalization processing are performed on the device power consumption feature vectors to obtain power consumption feature data.

[0044] In this example, data is collected on the operating status of the device, including the device's operating current, voltage, operating temperature, load status, operating time, standby time and other related parameters, describing the device's energy consumption from multiple aspects. and voltage Used to calculate the power consumption of the device at different time points , the formula is: ;

[0045] in, Is the device at time The operating voltage at the moment, is the operating current at the corresponding moment, is the instantaneous power consumption of the device. ) and standby time ) to analyze the power consumption behavior of the device under various working states. Using the original power consumption data set, perform delay power consumption domain transformation on it to obtain the delay power consumption feature matrix. Map the original data to a space that better reflects its dynamic characteristics, revealing the timing dependence and delay characteristics of power consumption. The delay power consumption domain transformation is achieved through the sliding window method and the time delay embedding method. Assume that the original power consumption data set is , by mapping the power consumption data to the delay feature space, the delay power consumption feature matrix is ​​obtained It is expressed as: ;

[0046] in, is the length of the sliding window. Each row in represents a feature vector containing multiple time delay embeddings. In this way, the variation relationship of power consumption at different time points is revealed. , construct the first power balancing matrix with minimum mean square error at device level The goal of minimum mean square error is to minimize the square error between each data point and the mean in the matrix, so as to eliminate noise and abnormal fluctuations and improve the stability and accuracy of power consumption characteristics. Assume that the elements in the matrix are ,but The calculation formula is: ;

[0047] Through the above process, the first power balancing matrix which eliminates the influence of random noise is obtained. , which can more accurately reflect the real power consumption characteristics of the device. The first power balancing matrix is ​​divided into power consumption characteristics to obtain several power consumption characteristic sub-matrix groups. The sub-matrix is ​​divided based on different time periods or load states of the device, thereby refining the description of the power consumption characteristics. Assume that the sub-matrix group is , each sub-matrix corresponds to a specific subset of power consumption characteristics. Based on the power consumption characteristic sub-matrix group, iterative calculation and matrix dimension reduction are performed to obtain the second power consumption balancing matrix . Matrix dimension reduction is achieved through methods such as principal component analysis, thereby removing redundant information and making the data more compact and easier to process. Perform discrete Fourier transform to obtain frequency domain power consumption characteristics Through Fourier transform, the power consumption characteristics are transformed from the time domain to the frequency domain, so as to more intuitively analyze the periodicity and frequency components of power consumption. Assuming the matrix One of the time series is , then its discrete Fourier transform is: ;

[0048] in, is the frequency domain signal, is the length of the time series. By analyzing the frequency domain signal, the power consumption fluctuation parameters and load change parameters of the device under different load states are extracted to obtain the dynamic power consumption characteristics, which describe the power consumption changes of the device under different frequency components. Based on the dynamic power consumption characteristics, the power management characteristics are fused to obtain the power consumption feature vector of the device. . Perform feature dimensionality reduction and normalization on the device power consumption feature vector. The purpose of feature dimensionality reduction is to remove redundant features to make the data more compact and easier to calculate, while normalization is to adjust the feature values ​​to the same scale so that the features can be compared more fairly in subsequent analysis. Assume that the feature vector is , then the normalized formula is: ;

[0049] in, is the normalized eigenvalue, and are the minimum and maximum values ​​in the feature vector respectively. After feature dimension reduction and normalization, the power consumption feature data is finally obtained.

[0050] In one example, power consumption feature data is input into a heterogeneous device management neural network for power management prediction, and device power consumption prediction parameters are obtained, including:

[0051] The power consumption feature data is input into the device feature encoding layer of the heterogeneous device management neural network. The device feature encoding layer includes four heterogeneous feature extraction units. Each heterogeneous feature extraction unit includes three layers of one-dimensional convolutional networks and two layers of fully connected networks to obtain device feature encoding.

[0052] The device feature encoding is processed through the timing analysis layer of the heterogeneous device management neural network. The timing analysis layer includes a bidirectional gated recurrent unit and a self-attention mechanism. The bidirectional gated recurrent unit is set with 128 hidden units, and the self-attention mechanism includes 8 attention heads to obtain a timing feature map.

[0053] The time series feature map is input into the multimodal fusion layer of the heterogeneous device management neural network. The multimodal fusion layer contains three parallel feature conversion branches. Each feature conversion branch contains two layers of fully connected networks and one normalization layer to obtain a fused feature vector.

[0054] The fused feature vector is input into the power consumption generator of the heterogeneous device management neural network for processing. The power consumption generator adopts a deep residual network structure. The deep residual network structure contains six residual blocks. Each residual block contains two layers of convolution and one skip connection to obtain a power consumption prediction sequence.

[0055] The power consumption prediction sequence is input into the discriminator network of the heterogeneous device management neural network. The discriminator network includes four layers of progressive discriminant units. Each progressive discriminant unit includes a convolution layer and a pooling layer to obtain the prediction credibility.

[0056] Based on the prediction credibility, the power consumption prediction sequence is input into the parameter generation network of the heterogeneous device management neural network for parameter generation. The parameter generation network contains three fully connected layers and two normalization layers, and outputs the device power consumption prediction parameters through the softmax activation function.

[0057] In this example, the power consumption feature data is input into the device feature encoding layer of the heterogeneous device management neural network. The device feature encoding layer consists of four heterogeneous feature extraction units, each of which contains three layers of one-dimensional convolutional networks and two layers of fully connected networks. Assume that the input power consumption feature data is ,in Represents the time dimension, Represents the number of features. These features are convolved through a one-dimensional convolutional network to extract local features in the data. The calculation of one-dimensional convolution is expressed as: ;

[0058] in, It is The weight matrix of the layer convolution kernel, * represents the convolution operation, is the bias term, is an activation function (such as ReLU function). Through the processing of the three-layer convolutional network, the high-level representation of power consumption characteristics is extracted layer by layer, capturing the feature change pattern in different time windows. The two-layer fully connected network performs a global feature combination on the convolutional features to form a more compact feature representation and obtain the device feature encoding vector ,in is the encoding dimension. The device feature encoding is input into the timing analysis layer of the heterogeneous device management neural network. The timing analysis layer consists of a bidirectional gated recurrent unit (Bi-GRU) and a self-attention mechanism. The bidirectional gated recurrent unit can effectively capture the timing dependency of power consumption features, especially in the case of complex time dependencies in device power consumption features. Assume that the input feature encoding is , then the forward and backward calculation formulas of the bidirectional gated recurrent unit are:

[0059] ;

[0060] ;

[0061] in, and Represents the time step The forward and backward hidden states of the bidirectional gated recurrent unit are connected to obtain a complete temporal feature vector by transferring information in both the forward and backward directions. The self-attention mechanism is used to strengthen the modeling of feature sequences. The self-attention mechanism can assign different weights to each time step by calculating the correlation between different time steps in the feature sequence, thereby emphasizing the time step features that are most important for the overall prediction. The calculation process of self-attention includes querying the matrix , key matrix Sum Matrix Generation of:

[0062] ;

[0063] in, are the projection matrices of query, key and value respectively. By calculating the attention weight : ;

[0064] in, is the dimension of the key, used for scaling to avoid large values. Then calculate the time series feature map: ;

[0065] After being processed by the self-attention mechanism, a feature map containing global temporal correlation is obtained. , which can better represent the dependence and importance of the device power consumption characteristics in the time dimension. Map the timing characteristics The input is fed into the multimodal fusion layer of the heterogeneous device management neural network. The multimodal fusion layer contains three parallel feature transformation branches, each of which consists of two fully connected layers and a normalization layer. Each feature transformation branch is responsible for extracting features from different aspects and normalizing them to ensure the scale consistency of the features. Assume that the input feature is , then the calculation of the fully connected layer is: ;

[0066] in, and are the weights and biases of the fully connected layer, is the activation function. The normalization layer normalizes the features to zero mean and unit variance by subtracting the mean and dividing by the standard deviation to obtain a more stable fusion feature. Through three parallel feature conversion branches, feature representations in different modes are obtained, and finally these features are fused to obtain a fusion feature vector . The fused feature vector The input is processed into the power generator of the heterogeneous device management neural network. The power generator adopts a deep residual network (ResNet) structure, which contains six residual blocks. Each residual block contains two layers of convolution and one skip connection. Assume that the input feature is , then the output of the residual block is: ;

[0067] in, and is the weight matrix of the convolutional layer, * represents the convolution operation, and is the bias term. The existence of jump connections enables the network to better retain input features, thereby alleviating the gradient vanishing problem in deep networks. After being processed by multiple layers of residual blocks, the power consumption generator generates a more accurate power consumption prediction sequence. , which represents the power consumption prediction value of the device in different time periods. The input is fed into the discriminator network of the heterogeneous device management neural network for credibility assessment. The discriminator network contains four layers of progressive discriminant units, each of which contains a convolutional layer and a pooling layer. The convolutional layer is used to extract local features in the power consumption prediction sequence, while the pooling layer is used to reduce the dimension of the features and enhance the generalization ability of the model. Assume that the input feature is , then the calculation of the convolutional layer is: ;

[0068] Based on the credibility of the prediction, the power consumption prediction sequence is input into the parameter generation network of the heterogeneous device management neural network for parameter generation. The parameter generation network contains three fully connected layers and two normalization layers. After the fully connected layers and normalization processing, the softmax activation function is used to output the power consumption prediction parameters of the device. The softmax function converts the output value into a probability distribution form, which represents the power consumption distribution of the device in different time periods. The calculation formula is: ;

[0069] in, For the The probability of an output node, is the linear combination output of the fully connected layer. In this way, the power consumption prediction parameters of the device are finally obtained to guide the power management and power allocation of the device.

[0070] In one example, a virtual power manager is deployed based on the device power consumption prediction parameters, and the virtual power manager performs power response control on the communication base station and the backup power supply to obtain power adjustment parameters, including:

[0071] A power response model of a virtual power manager is constructed based on the device power consumption prediction parameters, and the device power consumption prediction parameters are divided into a base station power consumption response matrix and a backup power supply response matrix;

[0072] Perform power control calculation on the base station power consumption response matrix, determine the power allocation weight and power supply response threshold of the base station, and obtain the base station power control strategy;

[0073] Perform dynamic programming calculation on the backup power response matrix to solve the optimal switching timing of the backup power supply and obtain the backup power supply power control strategy;

[0074] Power supply response optimization is performed based on the base station power control strategy and the backup power supply power control strategy to obtain power supply response optimization parameters;

[0075] Build a control strategy model for the virtual power manager based on the power response optimization parameters, and deploy the control strategy model to the communication base station and backup power supply through the real-time scheduling interface to generate power response execution instructions;

[0076] Respond to and execute the power response execution instruction, collect the real-time power consumption fluctuation data of the communication base station and the switching status data of the backup power supply, and dynamically correct the power response optimization parameters based on the real-time power consumption fluctuation data and the switching status data to obtain the power regulation parameters.

[0077] In this example, the device power consumption prediction parameters are used to accurately describe the power consumption requirements and response characteristics of the communication base station and the backup power supply in different time periods. Assume that the device power consumption prediction parameters are ,in represents the time step, Represents the number of devices. By dividing these parameters into the base station power consumption response matrix and Backup Power Response Matrix , respectively managing the power consumption requirements of the communication base station and the backup power supply. Perform power control calculation on the base station power consumption response matrix to determine the power allocation weight and power supply response threshold of the base station, and obtain the base station power control strategy. The purpose of power control calculation is to ensure that the power demand of the base station in different time periods can be reasonably met while minimizing unnecessary power consumption. Assume that the base station power consumption response matrix is ,in represents the number of base stations, for each time step , determine the base station Power allocation weight and power supply response threshold The power allocation weight is used to determine the power priority of each base station, and the power response threshold is used to determine whether power scheduling is required when the power demand exceeds a certain limit. The power allocation weight is calculated using the following formula: ;

[0078] in, Indicates base station At time step The power consumption requirement, Indicates base station At time step The power allocation weights are set to ensure that the sum of the weights of all base stations is 1. Based on these weights, the power response threshold is determined to implement the power control strategy: ;

[0079] in, is an adjustment coefficient, is the maximum available power capacity of the base station. By determining these power allocation weights and response thresholds, a reasonable power control strategy is formulated for the base station to meet the power demand and reduce energy consumption under different load conditions. At the same time, for the backup power response matrix, dynamic programming calculation is performed to solve the optimal switching timing of the backup power supply and obtain the power control strategy of the backup power supply. The goal of dynamic programming is to find when to start and shut down the backup power supply under given power consumption prediction conditions to maximize system stability and minimize energy consumption. Assume that the backup power response matrix is ,in Indicates the number of backup power supplies and defines a state transfer function ,in represents the system at time step status, Indicates that at time step The goal of dynamic programming is to find an optimal strategy for the operation of the backup power supply (such as starting or shutting down). , so that the following objective function is maximized: ;

[0080] in, Indicates in status Next action The benefits brought are expressed as the weighted sum of power savings and reliability improvements. Through iterative calculation and maximization of the benefit function, the optimal switching timing of the backup power supply is obtained to determine how to switch the backup power supply at different time points to ensure system stability and minimize energy consumption. The power response is optimized based on the base station power control strategy and the backup power supply power control strategy to obtain the power response optimization parameters. Coordinate the power allocation between the base station and the backup power supply at the system level to ensure that the entire power management strategy achieves the optimal energy efficiency while meeting the system power consumption requirements. Assume that the power control strategies of the base station and the backup power supply are and , power supply response optimization parameters Determined by the following formula: ;

[0081] in, and is the optimization weight coefficient, which is used to balance the adjustment between the base station and the backup power supply. These optimization parameters ensure that the power scheduling between different devices reaches the optimal state. Based on the power response optimization parameters, a control strategy model of the virtual power manager is constructed, and the control strategy model is deployed to the communication base station and the backup power supply through the real-time scheduling interface to generate power response execution instructions. The control strategy model includes fine control of power allocation, response time, load changes, etc. of the base station and the backup power supply to ensure that the system can respond flexibly and achieve optimal energy efficiency under complex and changing conditions. Through the real-time scheduling interface, the power response execution instructions are quickly sent to each device to ensure that each device can be adjusted in time according to the current power consumption requirements. In the process of executing the power response execution instructions, the virtual power manager collects the real-time power consumption fluctuation data of the communication base station and the switching status data of the backup power supply in real time. Assume that the real-time power consumption fluctuation data is , the switching status data of the backup power supply is , these data reflect the actual operating status of the system when executing the control strategy. By analyzing these real-time data, the virtual power manager dynamically corrects the power response optimization parameters to cope with changes in power consumption requirements caused by external environmental changes, load fluctuations, etc., thereby further optimizing power management. The dynamic correction of the power response optimization parameters is achieved through the following formula: ;

[0082] in, is the correction factor, is the original power consumption prediction parameter, It is the power consumption data collected in real time. Through dynamic correction, the power management strategy is adjusted in real time according to the change of actual power consumption to ensure that the power allocation is always in the optimal state. Through dynamic correction of the power response optimization parameters, the power adjustment parameters are obtained to further guide the power allocation of communication base stations and backup power supplies to achieve global optimal power management. For example, in the management process of a communication base station, if the power consumption of the base station is predicted to be 50W at a certain moment, and the actual measured power consumption is 60W, the power manager will adapt to this change by adjusting the power allocation weight and response threshold to ensure that the backup power supply can intervene in time to meet the power consumption requirements.

[0083] In one example, a device power consumption scheduling matrix is ​​constructed based on power regulation parameters, and the device power consumption scheduling matrix is ​​balanced and optimized using a block iteration method to obtain a multi-device power consumption scheduling solution, including:

[0084] Perform time series division on the power adjustment parameters, and classify the power consumption data into the prediction stage and the real-time stage to obtain the two-stage power consumption scheduling data;

[0085] Based on the two-stage power consumption scheduling data, a device power consumption scheduling matrix is ​​constructed, where the device power consumption scheduling matrix includes device power allocation coefficients and load state parameters;

[0086] The device power consumption scheduling matrix is ​​divided into blocks and divided into multiple sub-scheduling matrices according to the time dimension and device type. Each sub-scheduling matrix corresponds to a power consumption allocation strategy for a scheduling period, and a scheduling sub-matrix group is obtained;

[0087] Perform iterative optimization calculation based on the scheduling sub-matrix group, and adjust the power allocation coefficient of each sub-scheduling matrix to obtain a target scheduling sub-matrix;

[0088] The target scheduling sub-matrix is ​​reorganized according to the time series to construct the equipment scheduling timing plan, and the power consumption of the equipment scheduling timing plan is balanced to obtain a balanced scheduling plan;

[0089] The device power consumption control instructions are generated based on the balanced scheduling scheme, and the instructions are classified according to the control requirements of the prediction stage and the real-time stage to obtain a staged control instruction set. The staged control instruction set is then organized in time sequence and prioritized to obtain a multi-device power consumption scheduling scheme.

[0090] In this example, the power adjustment parameters are time-series partitioned to divide the power consumption data of the device into two different parts: the prediction stage and the real-time stage. The data in these two stages represent the system's prediction of future power consumption requirements and the real-time power consumption of the device during actual operation. Assume that the power adjustment parameter matrix is ,in represents the time step, Represents the number of devices. Through time series division, two sub-matrices are obtained, namely the prediction stage matrix and real-time phase matrix ,in . These two matrices represent the power consumption in the prediction stage and the actual power consumption in the real-time stage respectively. The device power consumption scheduling matrix is ​​constructed based on the two-stage power consumption scheduling data. The device power consumption scheduling matrix is ​​used to describe the power allocation and load status of the device in different time periods. The power consumption scheduling matrix contains not only the power allocation coefficient of the device, but also the load status parameters. Assume that the power consumption scheduling matrix is , where each row contains the power allocation coefficient and the load state parameter. Power allocation coefficient Used to indicate equipment At time step The power allocation weight, and the load state parameter It is used to describe the device at the time step After constructing the device power consumption scheduling matrix, in order to improve the flexibility and computational efficiency of the scheduling strategy, the device power consumption scheduling matrix is ​​divided into blocks and divided into multiple sub-scheduling matrices according to the time dimension and device type. Assume that there are a total of The power consumption scheduling matrix is ​​divided into multiple sub-scheduling matrices according to time and device type. Represents the power consumption allocation strategy for a specific time period and device type. It helps to optimize independently for different equipment types and different time periods, thereby improving scheduling efficiency. Based on the sub-scheduling matrix group, iterative optimization calculation is performed to adjust the power allocation coefficient in each sub-scheduling matrix to achieve the best power utilization effect. In the iterative optimization, the power allocation coefficient is gradually adjusted to minimize power waste and meet the power requirements of the equipment. For each sub-scheduling matrix , power allocation coefficient The update is achieved through the gradient descent method, the formula is as follows: ;

[0091] in, Indicates The device at the iteration At time step The power distribution coefficient, is the learning rate, is the loss function, which is used to measure the energy efficiency and power utilization of the current power allocation scheme. Through iterative optimization, the power allocation is adjusted so that each sub-scheduling matrix can achieve the optimal power allocation under different load conditions. The optimized target scheduling sub-matrix group is obtained. . Reorganize the target scheduling submatrix according to the time series to construct the device scheduling timing plan. The device scheduling timing plan is used to describe the power consumption allocation strategy of each device in the entire time period. By performing power consumption balancing on these scheduling plans, the power consumption distribution between devices is made more uniform, improving the overall energy efficiency of the system. Assume that the final device scheduling timing plan is , each element in the matrix represents the balanced power allocation value of different devices at different time steps. In order to apply this balanced scheduling scheme to actual power management, device power consumption control instructions are generated based on the balanced scheduling scheme. The power consumption control instructions are classified according to the control requirements of the prediction stage and the real-time stage to obtain a staged control instruction set. The control instructions in the prediction stage are used for the pre-allocation of device power to ensure that the power consumption requirements can be met in time during future load changes; the control instructions in the real-time stage are used to dynamically adjust the actual power consumption to cope with unpredictable load fluctuations. The key to ensuring that the control instructions can be executed in a timely and reasonable manner is to organize the staged control instruction set in time and prioritize it. Assume that the staged control instruction set is ,in represents the control instruction set of the prediction stage, Represents the control instruction set of the real-time stage. These instruction sets are prioritized according to the importance of each device and the urgency of the real-time load. For example, assuming the priority coefficient is , the priority of control instructions is determined by the following formula: ;

[0092] in, Indicates the device At time step The control instruction priority, Indicates the importance weight of the device, is the load status parameter of the device. By sorting by priority, it is ensured that when multiple devices are running in parallel, important devices can get power support first and meet their load requirements in real time. In the final multi-device power consumption scheduling scheme, the power allocation and control strategy of each device are optimized according to its actual load and energy efficiency target. Through this scheduling scheme, the power management of each device is realized to achieve the overall optimal power consumption scheduling.

[0093] In one example, power allocation control is performed according to a multi-device power consumption scheduling scheme to generate an initial power allocation strategy, and the initial power allocation strategy is dynamically adjusted through a three-layer Bayesian network model to obtain a target power allocation strategy, including:

[0094] Perform power allocation control according to the multi-device power consumption scheduling scheme and generate an initial power allocation strategy;

[0095] The initial power allocation strategy is input into the first power-aware network of the three-layer Bayesian network model. The first power-aware network contains four parallel power feature extraction units. Each power feature extraction unit contains three convolutional layers and two fully connected layers to obtain a power feature vector.

[0096] The power feature vector is input into the second prior distribution network of the three-layer Bayesian network model, the second prior distribution network includes three probability distribution estimation modules, each probability distribution estimation module includes a variational inference layer and a Gaussian mixture model, and the power distribution parameters are obtained;

[0097] The power distribution parameters are dynamically updated through the third layer of the posterior inference network of the three-layer Bayesian network model. The third layer of the posterior inference network includes a Markov chain Monte Carlo sampler and a variational posterior inference device to obtain the power adjustment parameters.

[0098] The initial power allocation strategy is updated based on the power adjustment parameters, and the optimal power allocation value is calculated through the Bayesian update rule to obtain the target power allocation strategy.

[0099] In this example, the initial power allocation of each device is determined based on the power consumption scheduling scheme of multiple devices. In a communication network, each device (such as a base station or a backup power supply) has different power consumption requirements in different time periods. The generation of the initial power allocation strategy aims to reasonably allocate limited power resources to meet the needs of the devices as much as possible. Assume that the power consumption scheduling scheme of multiple devices is ,in represents the time step, Represents the number of devices. The initial power allocation strategy is obtained by allocating and calculating the power consumption requirements of each device in different time periods. , where each element Represents the time step Time Equipment The initial power allocation strategy is calculated by the following formula: ;

[0100] in, Indicates the device At time step The power consumption requirement is Represents the power allocation ratio of the device at this time step. Through this step, the total power allocation ratio is guaranteed to be 1, thereby achieving reasonable scheduling of electricity. The initial power allocation strategy is input into the first layer of the three-layer Bayesian network model, namely the power-aware network, for power feature extraction. The power-aware network consists of four parallel power feature extraction units, each of which contains three convolutional layers and two fully connected layers. The function of the convolutional layer is to extract the local features of the power consumption data and capture the change pattern of power in the time dimension through convolution operations. Assume that the input initial power allocation strategy is , after the The output features after layer convolution are: ;

[0101] in, is the weight of the convolution kernel, * represents the convolution operation, is the bias term, is an activation function (such as ReLU) used to introduce nonlinearity. This method enables the network to extract the deep representation of power features layer by layer, and the convolutional features are fused and compressed through two fully connected layers to obtain the power feature vector ,in Represents the dimension of the feature vector. The power feature vector is input into the second layer of the three-layer Bayesian network model, namely the prior distribution network. The prior distribution network contains three probability distribution estimation modules, each of which contains a variational inference layer and a Gaussian mixture model. The role of the variational inference layer is to model the distribution of power features through variational inference and obtain a probabilistic description of each feature. Assume that the power feature vector is , then the output of the variational inference layer is the prior probability distribution , the formula is: ;

[0102] in, and are weight matrices and bias terms respectively, and the output is converted into probability form through the softmax function. The prior distribution is modeled by the Gaussian mixture model. The function of the Gaussian mixture model is to fit the overall distribution of the data with multiple Gaussian distributions to obtain a more accurate description of the power distribution. Assume that the Gaussian mixture model contains The power distribution parameters are expressed as: ;

[0103] in, Indicates The weights of the Gaussian components, and denote the mean and covariance matrices respectively, is a Gaussian distribution. Through this process, the power distribution parameter is obtained , describing the prior probability distribution of the device power characteristics. The power distribution parameters are dynamically updated through the third layer of the three-layer Bayesian network model, namely the posterior inference network. The posterior inference network contains a Markov chain Monte Carlo (MCMC) sampler and a variational posterior inference. The MCMC sampler is used to sample from the prior distribution of power to approximately obtain the posterior distribution of the device. Assume that the prior distribution is , the sample set of posterior distribution is obtained through MCMC sampling ,in Represents the number of samples. The variational posterior inference engine uses these sampled samples to update the posterior parameters of the power distribution and obtain the dynamically updated power adjustment parameters Based on the obtained power adjustment parameters, the initial power allocation strategy is updated to generate a more reasonable power allocation strategy. The parameter update process follows the Bayesian update rule, which updates the estimated value of power allocation by combining the prior distribution and the observed data to obtain the posterior distribution. Assume that the initial power allocation strategy is , the power adjustment parameter is , then the updated power allocation value is:

[0104] ;

[0105] in, is the learning rate, which controls the update amplitude. In this way, the target power allocation strategy is obtained , compared with the initial power allocation strategy, it is more in line with the actual needs of the equipment and the energy efficiency goals of the system.

[0106] In one example, the power feature vector is input into the second prior distribution network of the three-layer Bayesian network model, the second prior distribution network includes three probability distribution estimation modules, each probability distribution estimation module includes a variational inference layer and a Gaussian mixture model, and the power distribution parameters are obtained, including:

[0107] Performing data grouping processing on the power feature vectors, dividing them into three groups of 128-dimensional feature vectors according to power variation characteristics, load fluctuation characteristics, and device response characteristics, to obtain a power feature data group;

[0108] The power feature data set is input into the variational inference layer of the first probability distribution estimation module in the second layer prior distribution network. The variational inference layer includes three layers of feedforward neural network, each layer includes 256 neurons and ReLU activation function, and the first prior probability distribution is obtained by the variational evidence lower bound maximization method.

[0109] The first prior probability distribution is modeled by a Gaussian mixture model of a first probability distribution estimation module, the Gaussian mixture model includes 8 mixture components, and an expectation maximization algorithm is used to iteratively calculate the mean vector and covariance matrix of each component to obtain first power distribution parameters;

[0110] The power feature data group is input into the second probability distribution estimation module in the second layer prior distribution network. The second probability distribution estimation module constructs a load fluctuation probability model through a variational inference layer including four layers of variational autoencoders. The encoding dimension of the variational autoencoder is 64, the latent variable dimension is 32, and six groups of Gaussian mixture components are used for modeling to obtain the second power distribution parameters.

[0111] The power characteristic data group is input into the third probability distribution estimation module in the second layer prior distribution network. The third probability distribution estimation module establishes a device response probability model through a variational inference layer including two layers of stacked variational recurrent neural networks. Each layer of the variational recurrent neural network includes 128 LSTM units and adopts four groups of Gaussian mixture components for modeling to obtain the third power distribution parameters.

[0112] The first power distribution parameter, the second power distribution parameter and the third power distribution parameter are subjected to parameter weighted fusion to obtain a power distribution parameter.

[0113] In this example, data grouping is performed on the power feature vector, and the features are divided into three groups of 128-dimensional feature vectors according to power change features, load fluctuation features, and device response features, thus obtaining a complete power feature data group. The original power feature vector is set to ,in represents the number of samples, Indicates the feature dimension of each sample. In this feature data, according to the attribute of the feature, it is divided into three groups of feature vectors representing power change characteristics, load fluctuation characteristics and equipment response characteristics, assuming that the feature dimension of each group is 128. After the division process, the power feature data group is obtained , where the dimension of each feature group is 128. The power feature data set is input into the second layer of prior distribution network and processed by the variational inference layer in the first probability distribution estimation module. The variational inference layer contains three layers of feedforward neural network, each layer consists of 256 neurons and uses ReLU activation function to enhance the nonlinear expression ability of the feature. Assume that the input feature is , through the processing of three-layer feedforward neural network, the output of each layer is expressed as: ;

[0114] in, For the The weight matrix of the layer feed-forward neural network, is the bias term, is the ReLU activation function, which is used to introduce nonlinear characteristics. By maximizing the variational evidence lower bound, the potential variables are approximated and the first prior probability distribution is obtained. . Its goal is to make the approximate distribution as close to the true distribution as possible by minimizing the KL divergence between the true posterior distribution and the approximate posterior distribution. The first prior probability distribution is modeled by the Gaussian mixture model in the first probability distribution estimation module. The Gaussian mixture model contains 8 mixture components to simulate complex multi-peak distribution characteristics. Assume that the probability density function of the first Gaussian mixture model is: ;

[0115] in, Indicates The weights of the mixture components satisfy , For the The mean vector of the Gaussian components, is the covariance matrix. These parameters are iteratively updated through the expectation maximization algorithm to obtain the mean and covariance of each component and obtain the first power distribution parameter . The EM algorithm calculates the expected value of each data point belonging to each Gaussian component in the expectation step, and then updates the mean and covariance according to these expected values ​​in the maximization step until convergence. The power feature data set is input into the second probability distribution estimation module in the second layer prior distribution network, which constructs a load fluctuation probability model through a variational inference layer containing four layers of variational autoencoders. The variational autoencoder is a deep generative model that can effectively model the probability distribution of complex data. Assume that the input feature is , after being processed by the four-layer encoder, the encoding dimension is compressed to 64, and the dimension of the latent variable is 32. The encoding process is expressed as:

[0116] ;

[0117] in, is the approximate posterior distribution, is the parameter of the encoder. In order to simulate the complex characteristics of load fluctuations, six groups of Gaussian mixture components are used to model the latent variables, and their probability density function is: ;

[0118] By combining variational inference and EM algorithm, the second power distribution parameter is obtained , these parameters can accurately describe the probability distribution of load fluctuation characteristics. The device response characteristics in the power characteristic data set are input into the third probability distribution estimation module in the second layer prior distribution network. The third probability distribution estimation module uses a variational inference layer containing two layers of stacked variational recurrent neural networks to establish a probability model of device response. Variational recurrent neural networks can effectively capture time series characteristics and are suitable for modeling device response characteristics. Assume that the input feature is , processed by a variational recurrent neural network with 128 LSTM units in each layer, the potential characteristics of the device response are obtained , its approximate posterior distribution is: ;

[0119] in, are the parameters of VRNN. Four sets of Gaussian mixture components are used to model it: ;

[0120] After parameter estimation, the third power distribution parameter is obtained , these parameters are used to describe the probability characteristics of the device response. The first power distribution parameter, the second power distribution parameter and the third power distribution parameter are weighted and fused to obtain the overall power distribution parameter. Assume that the weighting coefficients are , then the fused power distribution parameters are expressed as: ;

[0121] Among them, the selection of weighting coefficients depends on the importance of each feature group to the overall power consumption prediction. Through weighted fusion, the three characteristics of power change, load fluctuation and equipment response are comprehensively considered to obtain a more comprehensive power distribution model.

[0122] Reference Figure 2 This embodiment provides a communication device optimization device based on power management, including:

[0123] Transformation module 1, used for performing delayed power consumption domain transformation and minimum mean square error power equalization processing on power consumption data of the communication device to obtain power consumption characteristic data;

[0124] Prediction module 2, used for inputting power consumption characteristic data into heterogeneous device management neural network for power management prediction, and obtaining device power consumption prediction parameters;

[0125] Control module 3, used to deploy a virtual power manager based on the equipment power consumption prediction parameters, and perform power response control on the communication base station and the backup power supply through the virtual power manager to obtain power regulation parameters;

[0126] Construction module 4 is used to construct a device power consumption scheduling matrix based on the power supply adjustment parameters, and to optimize the device power consumption scheduling matrix in a block-by-block iterative manner to obtain a multi-device power consumption scheduling solution;

[0127] The adjustment module 5 is used to perform power allocation control according to the multi-device power consumption scheduling scheme, generate an initial power allocation strategy, and dynamically adjust the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy.

[0128] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0129] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0130] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0131] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0132] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0133] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0134] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A communication device optimization method based on power management, characterized in that: The following steps are involved: Performing delayed power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data; Inputting the power consumption characteristic data into a heterogeneous device management neural network for power management prediction to obtain device power consumption prediction parameters; Deploy a virtual power manager based on the device power consumption prediction parameters, and perform power response control on the communication base station and the backup power supply through the virtual power manager to obtain power regulation parameters; Based on the power adjustment parameters, a device power consumption scheduling matrix is ​​constructed, and the device power consumption scheduling matrix is ​​balanced and optimized by block iteration to obtain a multi-device power consumption scheduling scheme; specifically, the following steps are performed: time series division is performed on the power adjustment parameters, and the power consumption data is classified according to the prediction stage and the real-time stage to obtain two-stage power consumption scheduling data; based on the two-stage power consumption scheduling data, a device power consumption scheduling matrix is ​​constructed, and the device power consumption scheduling matrix includes a device power allocation coefficient and a load state parameter; the device power consumption scheduling matrix is ​​processed in blocks, and is divided into a plurality of sub-scheduling matrices according to the time dimension and the device type, and each sub-scheduling matrix corresponds to a scheduling The power consumption allocation strategy of each time period is obtained to obtain a scheduling sub-matrix group; iterative optimization calculation is performed based on the scheduling sub-matrix group, and the power allocation coefficient of each sub-scheduling matrix is ​​adjusted to obtain a target scheduling sub-matrix; the target scheduling sub-matrix is ​​reorganized according to the time series to construct a device scheduling timing plan, and the device scheduling timing plan is subjected to power consumption balancing processing to obtain a balanced scheduling plan; based on the balanced scheduling plan, device power consumption control instructions are generated, and the instructions are classified according to the control requirements of the prediction stage and the real-time stage to obtain a staged control instruction set, and the staged control instruction set is organized in time sequence and prioritized to obtain a multi-device power consumption scheduling plan; Power allocation control is performed according to the multi-device power consumption scheduling scheme to generate an initial power allocation strategy, and the initial power allocation strategy is dynamically adjusted through a three-layer Bayesian network model to obtain a target power allocation strategy.

2. The communication device optimization method based on power management according to claim 1, characterized in that: The performing of delayed power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data includes: Collect data on the operating current, operating voltage, operating temperature, load status, operating time, and standby time of the communication equipment to obtain the original power consumption data set; Performing a delay power consumption domain transformation on the original power consumption data set, mapping the original power consumption data set to a delay power consumption feature space, and obtaining a delay power consumption feature matrix; Constructing a first power balancing matrix with minimum mean square error at the device level based on the delay power consumption characteristic matrix, and dividing the first power balancing matrix by power consumption characteristics to obtain a power consumption characteristic sub-matrix group; Perform iterative calculation and matrix dimension reduction based on the power consumption characteristic sub-matrix group to obtain a second power consumption balancing matrix; Performing a discrete Fourier transform on the second power consumption balancing matrix to obtain a frequency domain power consumption feature, and extracting a power consumption fluctuation parameter and a load change parameter of the device under different load states according to the frequency domain power consumption feature to obtain a dynamic power consumption feature; Power management feature fusion is performed based on the dynamic power consumption feature to obtain a device power consumption feature vector, and feature dimension reduction and normalization processing are performed on the device power consumption feature vector to obtain power consumption feature data.

3. The communication device optimization method based on power management according to claim 2, characterized in that: The step of inputting the power consumption characteristic data into a heterogeneous device management neural network for power management prediction to obtain device power consumption prediction parameters includes: Inputting the power consumption feature data into a device feature encoding layer of a heterogeneous device management neural network, wherein the device feature encoding layer comprises four heterogeneous feature extraction units, each of which comprises three layers of one-dimensional convolutional networks and two layers of fully connected networks, to obtain device feature encoding; The device feature encoding is processed by the timing analysis layer of the heterogeneous device management neural network, the timing analysis layer includes a bidirectional gated recurrent unit and a self-attention mechanism, the bidirectional gated recurrent unit is set with 128 hidden units, and the self-attention mechanism includes 8 attention heads to obtain a timing feature map; Inputting the time series feature map into the multimodal fusion layer of the heterogeneous device management neural network, the multimodal fusion layer comprising three parallel feature conversion branches, each feature conversion branch comprising two layers of fully connected networks and one layer of normalization layer, to obtain a fused feature vector; Inputting the fused feature vector into the power consumption generator of the heterogeneous device management neural network for processing, the power consumption generator adopts a deep residual network structure, the deep residual network structure includes six residual blocks, each residual block includes two layers of convolution and one skip connection, and obtains a power consumption prediction sequence; Inputting the power consumption prediction sequence into a discriminator network of the heterogeneous device management neural network, wherein the discriminator network comprises four layers of progressive discriminant units, each of which comprises a convolutional layer and a pooling layer, to obtain prediction credibility; Based on the prediction credibility, the power consumption prediction sequence is input into the parameter generation network of the heterogeneous device management neural network for parameter generation. The parameter generation network includes three fully connected layers and two normalization layers, and outputs the device power consumption prediction parameters through a softmax activation function.

4. The communication device optimization method based on power management according to claim 3, characterized in that: The deploying a virtual power manager based on the device power consumption prediction parameter, and performing power response control on the communication base station and the backup power supply through the virtual power manager to obtain the power adjustment parameter, includes: Building a power response model of a virtual power manager based on the device power consumption prediction parameters, and dividing the device power consumption prediction parameters into a base station power consumption response matrix and a backup power supply response matrix; Performing power control calculation on the base station power consumption response matrix, determining the power allocation weight and power supply response threshold of the base station, and obtaining the base station power control strategy; Performing dynamic programming calculation on the backup power response matrix to solve the optimal switching timing of the backup power supply and obtain a backup power control strategy; Perform power supply response optimization based on the base station power control strategy and the backup power supply power control strategy to obtain power supply response optimization parameters; Building a control strategy model of a virtual power manager based on the power response optimization parameters, and deploying the control strategy model to a communication base station and a backup power supply through a real-time scheduling interface to generate a power response execution instruction; Respond to and execute the power response execution instruction, collect the real-time power consumption fluctuation data of the communication base station and the switching status data of the backup power supply, and dynamically correct the power response optimization parameters based on the real-time power consumption fluctuation data and the switching status data to obtain the power regulation parameters.

5. The communication device optimization method based on power management according to claim 1, characterized in that: The performing power allocation control according to the multi-device power consumption scheduling scheme, generating an initial power allocation strategy, and dynamically adjusting the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy, includes: Execute power allocation control according to the multi-device power consumption scheduling scheme to generate an initial power allocation strategy; Inputting the initial power allocation strategy into the first power-aware network layer of the three-layer Bayesian network model, wherein the first power-aware network layer comprises four parallel power feature extraction units, each of which comprises three convolutional layers and two fully connected layers, to obtain a power feature vector; Inputting the power feature vector into the second layer prior distribution network of the three-layer Bayesian network model, wherein the second layer prior distribution network comprises three probability distribution estimation modules, each probability distribution estimation module comprises a variational inference layer and a Gaussian mixture model, to obtain power distribution parameters; The power distribution parameters are dynamically updated through the third layer of the posterior inference network of the three-layer Bayesian network model, wherein the third layer of the posterior inference network includes a Markov chain Monte Carlo sampler and a variational posterior inference device to obtain a power adjustment parameter; The initial power allocation strategy is updated based on the power adjustment parameters, and the optimal power allocation value is calculated by the Bayesian update rule to obtain the target power allocation strategy.

6. The communication device optimization method based on power management according to claim 5, characterized in that: The power feature vector is input into the second layer prior distribution network of the three-layer Bayesian network model, the second layer prior distribution network includes three probability distribution estimation modules, each probability distribution estimation module includes a variational inference layer and a Gaussian mixture model, and the power distribution parameters are obtained, including: Performing data grouping processing on the power characteristic vectors, dividing them into three groups of 128-dimensional characteristic vectors according to power variation characteristics, load fluctuation characteristics and device response characteristics, to obtain power characteristic data groups; Input the power feature data set into the variational inference layer of the first probability distribution estimation module in the second layer prior distribution network, wherein the variational inference layer comprises a three-layer feedforward neural network, each layer comprises 256 neurons and a ReLU activation function, and obtains a first prior probability distribution by a variational evidence lower bound maximization method; The first prior probability distribution is modeled by a Gaussian mixture model of a first probability distribution estimation module, wherein the Gaussian mixture model includes 8 mixture components, and an expectation-maximization algorithm is used to iteratively calculate a mean vector and a covariance matrix of each component to obtain first power distribution parameters; The power characteristic data group is input into the second probability distribution estimation module in the second layer prior distribution network, the second probability distribution estimation module constructs a load fluctuation probability model through a variational inference layer including four layers of variational autoencoders, the encoding dimension of the variational autoencoder is 64, the latent variable dimension is 32, and six groups of Gaussian mixture components are used for modeling to obtain second power distribution parameters; Inputting the power characteristic data group into the third probability distribution estimation module in the second layer prior distribution network, the third probability distribution estimation module establishes a device response probability model through a variational inference layer including two layers of stacked variational recurrent neural networks, each layer of the variational recurrent neural network includes 128 LSTM units, and adopts four groups of Gaussian mixture components for modeling to obtain third power distribution parameters; The first power distribution parameter, the second power distribution parameter and the third power distribution parameter are subjected to parameter weighted fusion to obtain a power distribution parameter.

7. A communication equipment optimization device based on power management, characterized in that: For implementing the steps of the method according to any one of claims 1 to 6, the device comprises: A transformation module, used to perform delayed power consumption domain transformation and minimum mean square error power equalization processing on the power consumption data of the communication device to obtain power consumption characteristic data; A prediction module, used for inputting the power consumption characteristic data into a heterogeneous device management neural network to perform power management prediction and obtain device power consumption prediction parameters; A control module, configured to deploy a virtual power manager based on the device power consumption prediction parameter, and to perform power response control on the communication base station and the backup power supply through the virtual power manager to obtain a power regulation parameter; A construction module is used to construct a device power consumption scheduling matrix based on the power supply adjustment parameters, and to optimize the device power consumption scheduling matrix in a block-by-block iterative manner to obtain a multi-device power consumption scheduling scheme; specifically, the module comprises: performing time series division on the power supply adjustment parameters, and classifying the power consumption data according to the prediction stage and the real-time stage to obtain two-stage power consumption scheduling data; constructing a device power consumption scheduling matrix based on the two-stage power consumption scheduling data, wherein the device power consumption scheduling matrix includes a device power allocation coefficient and a load state parameter; performing block processing on the device power consumption scheduling matrix, and dividing it into a plurality of sub-scheduling matrices according to the time dimension and the device type, wherein each sub-scheduling matrix corresponds to a The power consumption allocation strategy of each scheduling period is obtained to obtain a scheduling sub-matrix group; iterative optimization calculation is performed based on the scheduling sub-matrix group, and the power allocation coefficient of each sub-scheduling matrix is ​​adjusted to obtain a target scheduling sub-matrix; the target scheduling sub-matrix is ​​reorganized according to the time series to construct a device scheduling timing plan, and the power consumption balancing processing of the device scheduling timing plan is performed to obtain a balanced scheduling plan; based on the balanced scheduling plan, device power consumption control instructions are generated, and the instructions are classified according to the control requirements of the prediction stage and the real-time stage to obtain a staged control instruction set, and the staged control instruction set is organized in time sequence and prioritized to obtain a multi-device power consumption scheduling plan; The adjustment module is used to perform power allocation control according to the multi-device power consumption scheduling scheme, generate an initial power allocation strategy, and dynamically adjust the initial power allocation strategy through a three-layer Bayesian network model to obtain a target power allocation strategy.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images