Adaptive power supply method, device and equipment for POE power supply

Abstract

The present invention relates to an adaptive power supply method, device and equipment for a POE power supply. The method: collect the first voltage and current data of each port of the POE power supply system, establish an initial dynamic load model; input a neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculate the voltage prediction value at the next moment based on the model parameters; construct n linear state models, and input an interactive multi-model analyzer to obtain the current optimal operation model; perform Fourier series expansion on the output control quantity of the PD controller to obtain a set of Fourier coefficients; perform opportunity constraint optimization calculation, and obtain the optimal control sequence by solving the optimization objective function; output a control signal to the power supply unit of the POE power supply system, perform dynamic optimization, and obtain a target dynamic load model. The present invention can update the dynamic load model in real time, ensuring the accuracy and adaptability of power supply control.

Description

Technical Field

[0001] The present invention relates to the technical field of POE power supply, and in particular to an adaptive power supply method, device and equipment of a POE power supply. Background Art

[0002] The traditional POE power supply system adopts a fixed power output mode, which cannot accurately supply power according to the dynamic power demand of the load equipment, resulting in low power supply efficiency, and it is also unable to effectively deal with sudden load changes.

[0003] In practical applications, the POE power supply system faces a variety of complex power supply scenarios. Different types of terminal devices have different power demand characteristics, and the power demand of the same device in different working states also varies significantly. Traditional POE power supply control methods are difficult to accurately capture and predict these dynamically changing load characteristics, thus affecting the stability and reliability of power supply. Summary of the invention

[0004] The main purpose of the present invention is to provide an adaptive power supply method, device and equipment for a POE power supply. The present invention can update the dynamic load model in real time, thereby ensuring the accuracy and adaptability of power supply control.

[0005] To achieve the above object, the present invention provides an adaptive power supply method for a POE power supply, comprising the following steps:

[0006] Collect the first voltage and current data of each port of the POE power supply system, calculate the real-time power value, and establish an initial dynamic load model including the power change curve and load characteristics;

[0007] Inputting the historical data sequence of the initial dynamic load model into a neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters;

[0008] Constructing n linear state models according to the characteristic data of the initial dynamic load model, and inputting the n linear state models into an interactive multi-model analyzer, calculating the probability value of each linear state model, and obtaining a current optimal operation model;

[0009] Based on the voltage prediction value and the optimal operation model, performing Fourier series expansion on the output control quantity of the PD controller to obtain a set of Fourier coefficients;

[0010] Performing a chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving an optimization objective function;

[0011] Based on the optimal control sequence, a control signal is output to the power supply unit of the POE power supply system, and the second voltage and current data are collected and fed back to the initial dynamic load model for dynamic optimization to obtain a target dynamic load model.

[0012] The present invention also provides an adaptive power supply device for a POE power supply, comprising:

[0013] The acquisition module is used to collect the first voltage and current data of each port of the POE power supply system, calculate the real-time power value, and establish an initial dynamic load model including the power change curve and load characteristics;

[0014] A calculation module, used for inputting the historical data sequence of the initial dynamic load model into a neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters;

[0015] A construction module, used to construct n linear state models according to the characteristic data of the initial dynamic load model, and input the n linear state models into an interactive multi-model analyzer, calculate the probability value of each linear state model, and obtain the current optimal operation model;

[0016] An expansion module, used for performing Fourier series expansion on the output control quantity of the PD controller based on the voltage prediction value and the optimal operation model to obtain a set of Fourier coefficients;

[0017] A solution module, used for performing chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving an optimization objective function;

[0018] A dynamic optimization module is used to output a control signal to the power supply unit of the POE power supply system based on the optimal control sequence, collect the second voltage and current data and feed it back to the initial dynamic load model for dynamic optimization to obtain a target dynamic load model.

[0019] 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.

[0020] 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.

[0021] In summary, the technical solution provided by the present invention establishes an initial dynamic load model and introduces a neural network Gaussian process analyzer. The method of the present invention accurately models and predicts the load characteristics, effectively reduces the voltage fluctuation of the power supply system, and improves the power supply stability; adopts an interactive multi-model analysis method, and the present invention can accurately identify the dynamic characteristics of the system under different load conditions, and realizes accurate modeling and control of complex load conditions; combined with the PD controller and Fourier series expansion technology, the method of the present invention can reduce the control complexity while effectively suppressing the periodic disturbance of the system; by introducing the chance constrained optimization algorithm, the method of the present invention can achieve optimal control of the power supply process while ensuring the power supply reliability; adopts a dynamic optimization feedback mechanism, and the method of the present invention can update the dynamic load model in real time, ensuring the accuracy and adaptability of the power supply control. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the steps of an adaptive power supply method of a POE power supply in one embodiment of the present invention;

[0023] Figure 2 It is a structural block diagram of an adaptive power supply device of a POE power supply in one embodiment of the present invention;

[0024] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0025] 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

[0026] 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.

[0027] Reference Figure 1 This embodiment provides a POE power supply adaptive power supply method, including the following steps:

[0028] S1, collecting the first voltage and current data of each port of the POE power supply system, calculating the real-time power value, and establishing an initial dynamic load model including a power change curve and load characteristics;

[0029] Among them, a sampling time interval is set for each port of the POE power supply system, and the voltage and current values ​​of the port are collected in real time at each sampling time point to form a sampling data set containing the original voltage and current data. According to the collected voltage and current values, the real-time power value of each port is calculated using the product operation formula (that is, power is equal to the product of voltage and current). The real-time power values ​​are sorted and stored according to the time series to form a power time series, reflecting the trend of power change over time. The power time series is curve fitted, and the power time series is converted into a mathematical expression to obtain a set of coefficient matrices that can describe the law of power change, including the basic mode and characteristics of power change. At the same time, the load characteristics are analyzed based on the original sampling data set. The characteristic information (such as the distribution and fluctuation characteristics of the voltage and current values) in the original sampling data is extracted and a load characteristic vector is constructed to reflect the actual working state of the system and the basic characteristics of the load. The load characteristic vector is subjected to principal component analysis to obtain a load characteristic matrix, which describes the main dynamic characteristics of the load and simplifies the complexity of subsequent modeling. The power change curve coefficient matrix and the load characteristic matrix are subjected to dynamic model construction and state equation modeling to obtain the initial state space equation group. State space equations are a common modeling method for dynamic systems, which can describe the relationship between the input, output and internal state of the system in the form of state variables. The construction of the initial state space equations aims to combine power changes and load characteristics to form a system description that can not only reflect the dynamics of power changes but also capture the laws of load characteristics. Power supply constraints are set for the initial state space equations, including the upper and lower limits of the power supply voltage and current, as well as system stability, power factor and other requirements. The power supply constraints are added to the initial state space equations to form a constrained state space equation. In order to enable the model to adapt to the real-time changing load demand, online parameter identification technology is introduced. During the operation of the power supply system, the real-time collected data will be continuously updated. These data will be dynamically combined with the constrained state space equations, and the parameter values ​​of the model will be continuously adjusted through the online parameter identification method. Dynamic adjustment can ensure that the model can always accurately reflect the dynamic characteristics of the current load, so as to achieve a rapid response to load changes during operation and obtain the initial dynamic load model.

[0030] S2, inputting the historical data sequence of the initial dynamic load model into the neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters;

[0031] Specifically, the historical data sequence in the initial dynamic load model is divided into a training data set and a validation data set. The training data set is used to learn model parameters, while the validation data set is used to evaluate the generalization ability of the model. The training data set is standardized, and the mean and variance of the data are adjusted to convert it into a standardized feature vector with zero mean and unit variance, and a standardized label is generated at the same time. Based on the standardized feature vector, a Gaussian kernel function is constructed to capture the nonlinear relationship in the data. The Gaussian kernel function maps the original input to a high-dimensional space by defining the similarity between samples, thereby enhancing the model's learning ability for complex patterns. After constructing the Gaussian kernel function, its covariance matrix is ​​calculated, and the matrix is ​​eigendecomposed to obtain the eigenvalue matrix and eigenvector matrix. The eigenvalue matrix describes the main direction of the data distribution and its range of variation, while the eigenvector matrix provides a specific projection method on these main directions. The eigenvalue matrix and the eigenvector matrix are used to calculate the Bayesian prior distribution to reflect the initial assumptions of the model parameters. The prior distribution provides a preliminary range of model parameters by combining domain knowledge and data distribution. Based on the prior distribution, the standardized labels are introduced to construct the corresponding likelihood function to describe the matching degree of the model under the given data. By combining the prior distribution with the likelihood function, the maximum a posteriori probability estimation method is used to calculate the posterior distribution of the model parameters. The posterior distribution reflects the optimization results of the model parameters under the condition of the observed data. The posterior distribution of the model parameters is input into the neural network Gaussian process model and cross-validated with the validation data set. Cross-validation can effectively evaluate the performance of the model on unseen data and calculate the model validation error by dividing the validation data set into multiple folds for training and validation respectively. Based on the validation error, the parameters of the neural network Gaussian process model are optimized, and the optimal model parameters are obtained by continuously adjusting the structure and hyperparameters of the model. The optimal model parameters represent the best configuration of the Gaussian process model when processing a specific data pattern, and significantly reduce the validation error. The state vector at the current moment is input into the neural network Gaussian process model, and forward calculation is performed based on the optimal model parameters to obtain the voltage prediction value at the next moment. The forward calculation process includes the mapping of the state vector and the Gaussian kernel, the dynamic update in the feature space, and the final output prediction result. Through this series of calculations, the voltage prediction value obtained can accurately reflect the operating status of the system at the next moment.

[0032] S3, constructing n linear state models according to the characteristic data of the initial dynamic load model, and inputting the n linear state models into the interactive multi-model analyzer, calculating the probability value of each linear state model, and obtaining the current optimal operation model;

[0033] It should be noted that the characteristic data of the initial dynamic load model are classified according to different load conditions. By analyzing the dynamic characteristics of the load, such as power fluctuation range, working cycle, load response rate, etc., the load characteristic data are divided into n subsets with similar characteristics. These load characteristic subsets are used as inputs to accurately reflect the operating state of the system under different working conditions. The state space equation is constructed for each load characteristic subset. The state space equation describes the dynamic behavior of the system in the form of state variables, including state transfer equations and observation equations. The former reflects the law of change of the system state over time, and the latter describes the relationship between the output signal and the state. By extracting the dynamic characteristics in the load subset, n groups of state space expressions are constructed. At the same time, based on these expressions, the state transfer matrix and measurement matrix are generated to obtain n linear state models. These models are simplified linear system descriptions in form, but can effectively capture the main dynamic behaviors under different load conditions. The n linear state models are input into the interactive multi-model analyzer for dynamic probability calculation. The core of the interactive multi-model analyzer is to estimate the initial probability values ​​of multiple models using mixed probability density functions. These initial probability values ​​reflect the possibility of each linear state model under the current working condition. After obtaining the initial probability value, the likelihood function value of each linear state model is calculated by the maximum likelihood estimator in combination with the voltage and current data collected in real time, that is, the current measurement value. The higher the likelihood function value of the model, the higher the degree of matching with the current system state. In order to improve the accuracy of model selection, the likelihood function value and the initial probability value are updated in a Bayesian manner. The posterior probability values ​​of each linear state model are calculated by the Bayesian formula. These posterior probability values ​​combine historical information (initial probability) and current measurement data (likelihood function value) to more accurately reflect the applicability of each model under the current working conditions. In order to ensure the standardization of the probability value and facilitate comparison, the posterior probability value is normalized to obtain the corrected probability value. The normalized corrected probability value can be directly used for model selection, and the model with the highest probability value is regarded as the optimal operating model of the current system. According to the calculation results of the corrected probability value, the linear state model with the largest probability is selected as the current optimal operating model.

[0034] S4, based on the voltage prediction value and the optimal operation model, the output control quantity of the PD controller is expanded by Fourier series to obtain a set of Fourier coefficients;

[0035] Specifically, the difference between the voltage prediction value and the standard voltage value is calculated to obtain a voltage error sequence, which reflects the degree of deviation between the current voltage and the target voltage. According to the classical PD control principle, the control equation of the PD controller is established based on the voltage error sequence. The controller output sequence is generated by weighted combination of the proportional term and the differential term of the error to describe the dynamic response of the PD controller at different time points. By determining the period boundary of the controller output sequence and performing a discrete Fourier transform on it, the fundamental frequency and main harmonic order of the control sequence are extracted. The fundamental frequency and harmonic order determine the accuracy and computational complexity of the Fourier series expansion. After the fundamental frequency and harmonic order are determined, the controller output sequence is decomposed into the Fourier series form to obtain a set of initial Fourier coefficients, which reflects the frequency characteristics of the control sequence. The initial Fourier coefficients are corrected and calculated according to the state equation of the optimal operation model. The optimal operation model describes the dynamic characteristics of the system under the current working conditions in the form of a state equation, and this model is used to deeply adjust the initial Fourier coefficients. By substituting the initial Fourier coefficients into the state equation of the optimal operation model and correcting them according to the actual dynamic behavior of the system, a correction coefficient matrix is ​​obtained. The correction coefficient matrix is ​​weightedly combined with the initial Fourier coefficients to form weighted Fourier coefficients. Different frequency components are assigned different weights according to the dynamic requirements of the system to ensure that the final coefficients are more in line with the frequency domain characteristics of the actual operation of the system. The weighted Fourier coefficients are verified for orthogonality to ensure that the Fourier coefficients are mathematically independent of each other, thereby avoiding the accumulation of redundancy or deviation. The weighted Fourier coefficients are processed by the Schmidt orthogonalization method to generate a set of orthogonalized coefficients that meet the orthogonality requirements. These orthogonalized coefficients are mutually orthogonal in the frequency domain, which can better describe the frequency characteristics of the control quantity and improve the stability and accuracy of the Fourier expansion. The orthogonalized coefficients are combined to form a set of Fourier coefficients.

[0036] S5, performing chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining the optimal control sequence by solving the optimization objective function;

[0037] Among them, according to the characteristics of the voltage prediction value, a set of voltage constraint conditions is constructed. The constraints include the upper and lower limits of the voltage, the short-term fluctuation threshold and the stability requirements. On this basis, in order to adapt to the uncertainty and randomness of the system, the confidence level is introduced, and the probability of satisfying each constraint is set to form a set of probabilistic constraints. At the same time, according to the frequency domain characteristics of the control signal described by the Fourier coefficient set, a set of control quantity constraint conditions is constructed. The control constraints include the amplitude limit of the power output, the harmonic distortion range and the dynamic response rate of the system. The control quantity constraint conditions and the probabilistic constraints of the voltage work together to define the feasible solution space of the power supply system. The probabilistic constraints are transformed. Through the analysis of the probability distribution characteristics, the probabilistic constraints are transformed into a set of deterministic constraint equations, thereby simplifying the problem to an optimization problem based on deterministic conditions. The optimization objective function is constructed based on the deterministic constraint equations. The optimization objective function takes the maximization of power supply efficiency, the minimization of power fluctuation or the minimization of power supply equipment loss as the goal. By comprehensively considering the constraints and system performance, a mathematical expression containing multi-objective trade-offs is formed. The objective function is decomposed into convex optimization, and the original problem is transformed into a dual problem by introducing the Lagrange multiplier method, and expressed in the form of a dual function. The Lagrange multiplier method can effectively handle complex constraint problems and reduce the difficulty of solving them by transforming the original problem into a dual problem that is easier to solve. After constructing the dual function, the KKT (Karush-Kuhn-Tucker) conditional equation group is constructed using the dual function. The KKT condition is a necessary and sufficient condition for convex optimization problems, and its form integrates the relationship between constraint conditions, objective function gradients, and Lagrange multipliers. The KKT conditional equation group is iteratively solved by the gradient descent method to obtain the optimal solution. The advantage of the gradient descent method is that it can converge quickly and adapt to the computational requirements of high-dimensional optimization problems, and is suitable for the optimization solution of complex power supply systems. The optimal solution is substituted into the optimization objective function, and a control sequence is generated by reverse iterative calculation. The reverse iterative process gradually adjusts the control signal to meet the constraint conditions and optimization objectives, ensuring that the generated control sequence has high accuracy and efficiency. The control sequence is dynamically optimized by decomposing the problem into multiple sub-problems, optimizing layer by layer and reversely synthesizing the final solution, so as to ensure that the generated optimal control sequence not only meets all constraints but also reaches the optimal state globally, and finally obtains the optimal control sequence that meets the constraints.

[0038] S6, outputting a control signal to the power supply unit of the POE power supply system based on the optimal control sequence, collecting the second voltage and current data and feeding it back to the initial dynamic load model for dynamic optimization, and obtaining a target dynamic load model.

[0039] Specifically, the optimal control sequence is converted into an analog control signal through a digital-to-analog converter, and the signal is output to the power supply unit of the POE power supply system to adjust the operating parameters of the power supply unit, thereby exerting precise control on the system. The discrete optimal control sequence is converted into a continuous control quantity, so that the power supply system can operate in a manner that adapts to the current load demand, thereby realizing dynamic power supply. At the output port of the power supply unit, the voltage and current are sampled in real time by a sensor to obtain second voltage and current data. Using these data, the real-time power is calculated and organized into a power data sequence in the form of a time series. The real-time sampled data reflects the actual dynamic behavior of the system after applying the optimal control sequence. The power data sequence is subjected to sliding window processing to extract the dynamic power change curve. The role of sliding window processing is to eliminate the influence of short-term fluctuations and highlight the long-term trend and dynamic characteristics of power changes. The dynamic power change curve is data fused with the initial dynamic load model so as to comprehensively describe the system characteristics by combining the real-time operation data with the historical model characteristics. The data fusion process involves statistical analysis and feature comparison of the dynamic power change curve with the power feature data of the initial dynamic load model to generate an updated power feature matrix. Based on the updated power characteristic matrix, the state space equations are reconstructed to obtain a new state equation coefficient matrix. The new state equation coefficient matrix is ​​optimized online. Online optimization is achieved through the recursive least squares method, which is based on real-time data and gradually converges to the global optimal solution by recursively updating parameters to obtain the optimized coefficient matrix. The optimized coefficient matrix is ​​substituted into the initial dynamic load model for parameter update, and finally the target dynamic load model is formed.

[0040] In one example, first voltage and current data of each port of the POE power supply system are collected, a real-time power value is calculated, and an initial dynamic load model including a power change curve and load characteristics is established, including:

[0041] A sampling time interval is set for each port of the POE power supply system, and the port voltage value and current value are collected at each sampling time point to obtain an original sampling data set, and a product operation is performed based on the voltage value and the current value in the original sampling data set to obtain the real-time power value of each port;

[0042] Sort and store the real-time power values ​​according to the time series, and construct a power time series containing multiple sampling points;

[0043] Based on the power time series, curve fitting is performed to obtain the power change curve coefficient matrix. Based on the original sampling data set, the load characteristic vector is constructed, and the principal component analysis is performed on the load characteristic vector to obtain the load characteristic matrix.

[0044] The power variation curve coefficient matrix and the load characteristic matrix are dynamically modeled and state equations are modeled to obtain the initial state space equations.

[0045] Power supply constraints are set for the initial state space equations to obtain a constrained state space equation group, and the constrained state space equation group is combined with real-time sampling data for online parameter identification to obtain an initial dynamic load model.

[0046] In this example, a sampling time interval is set at each port of the POE power supply system to ensure the synchronization and timeliness of data collection. At each sampling time point, the voltage value of the port is collected by a high-precision voltage sensor and current sensor. and current value , and obtain an original sampling data set containing multiple sampling points . Perform product operations based on the voltage and current values ​​in the original sampled data set , calculate the real-time power value of each port ,in represents the sampling time, is the power at each moment. All real-time power values ​​are sorted and stored in time series to form a power time series The power time series reflects the change of system power over time. In order to extract the trend characteristics of power change, the power time series is curve fitted. Assuming that the fitting adopts polynomial form, the power change curve is expressed as ,in are the fitted polynomial coefficients, is the order of the polynomial, and the fitting result is expressed as the coefficient matrix of the power change curve At the same time, the load feature vector is constructed based on the original sampling data set ,in and are the mean values ​​of voltage and current, respectively. is the standard deviation of power, and are the maximum and minimum values ​​of the power, respectively. These eigenvectors can capture the statistical characteristics of the load, but due to the high feature dimension and the presence of redundant information, principal component analysis is performed to extract the main features. Perform eigenvalue decomposition, select the principal component with the largest contribution, and obtain the load characteristic matrix , where each column represents the projection of a main feature. The coefficient matrix of the power change curve and the load characteristic matrix Combined with the state equation, the dynamic model of the system is constructed. The dynamic behavior of the system is expressed in the form of state equations, and the state variables are defined. Indicates the current state of the system, such as ,in is the first-order derivative of power, indicating the speed of power change, is the second-order derivative of power, indicating the acceleration of power change. The state equation is expressed as:

[0047] ;

[0048] in, is the state transfer matrix, which represents the dynamic relationship between states; is the control matrix, describing the control input The impact on the system state. For example, if the control input is a voltage change ,but The output equation is expressed as:

[0049] ;

[0050] in is the output variable of the system, such as real-time power value, and They are the output matrix and the direct transmission matrix. In order to ensure that the model meets the actual operation requirements of the power supply system, power supply constraints are added to the initial state space equations. For example, the voltage must satisfy , the power needs to meet These constraints are introduced into the model to obtain a constrained state space equation set. , combined with online parameter identification technology to dynamically update the model. The coefficient matrix of the model is updated by recursive least squares method. The core idea of ​​the recursive least squares method is to update the parameter estimates with the current measured data and minimize the prediction error. To adjust the parameters, the model can more accurately describe the dynamic behavior of the system. Substitute the model coefficients optimized by the above steps into the initial dynamic load model, complete the parameter update, and obtain the target dynamic load model.

[0051] In one example, the historical data sequence of the initial dynamic load model is input into the neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and the voltage prediction value at the next moment is calculated based on the model parameters, including:

[0052] The historical data sequence of the initial dynamic load model is divided to obtain a training data set and a validation data set, and the training data set is standardized to obtain a standardized feature vector and a standardized label;

[0053] A Gaussian kernel function is constructed based on the standardized eigenvector, and the covariance matrix of the Gaussian kernel function is eigen-decomposed to obtain the eigenvalue matrix and the eigenvector matrix;

[0054] The Bayesian prior distribution is calculated for the eigenvalue matrix and the eigenvector matrix to obtain the prior probability distribution of the model parameters, and the likelihood function is constructed based on the prior probability distribution and the standardized labels, and the posterior distribution of the model parameters is obtained by the maximum a posteriori probability calculation;

[0055] The posterior distribution of the model parameters is input into the neural network Gaussian process model, and the validation data set is cross-validated to obtain the model validation error. Based on the validation error, the parameters of the neural network Gaussian process model are optimized to obtain the optimal model parameters.

[0056] The state vector at the current moment is input into the neural network Gaussian process model, and forward calculation is performed based on the optimal model parameters to obtain the voltage prediction value at the next moment.

[0057] In this example, the historical data sequence of the initial dynamic load model is preprocessed. The historical data sequence contains the input characteristics of the load and the corresponding target output, such as voltage , Current ,power The data is divided into a training data set and a validation data set. The training data set is used for model learning, and the validation data set is used to evaluate model performance. When dividing the data, the time series ratio segmentation method is adopted, for example, 80% is used as the training set and 20% is used as the validation set. In order to improve the robustness and convergence speed of the model, the training data set is standardized and the eigenvector is calculated. The mean and standard deviation , and for each feature according to the formula After standardization, the standardized feature vector is obtained. and standardized labels , they satisfy the zero mean and unit variance distribution, which is suitable for the subsequent training of Gaussian process models. Based on the standardized feature vector , construct the Gaussian kernel function , the kernel function is used to measure the and The Gaussian kernel function is:

[0058] ;

[0059] in is the bandwidth parameter of the kernel function, which is used to control the scale of similarity in the feature space. Based on the Gaussian kernel function, the covariance matrix is ​​constructed ,in . Covariance matrix is symmetric and semi-positive, and can capture the correlation between data. Perform eigendecomposition to obtain the eigenvalue matrix and the eigenvector matrix ,in represents the eigenvalues ​​of the covariance matrix, is the corresponding eigenvector. Using the eigenvalue matrix and the eigenvector matrix , perform Bayesian derivation of the prior probability distribution of the model parameters. Assume that the prior distribution of the model parameters is a multivariate normal distribution , where w is the parameter vector to be estimated, the mean of the prior distribution is 0, and the covariance is . Combined with standardized labels , through the likelihood function To express the relationship between data and parameters, the likelihood function is expressed as:

[0060] ;

[0061] in is the variance of the observation noise. By combining the prior distribution and the likelihood function through Bayes’ theorem, we get the posterior distribution:

[0062] ;

[0063] The posterior distribution is optimized by the maximum a posteriori probability estimation to obtain the posterior expectation of the model parameters. The parameters in the posterior distribution are input into the neural network Gaussian process model. This model combines the nonlinear fitting ability of the neural network and the probabilistic inference characteristics of the Gaussian process to more accurately capture the dynamic behavior of complex systems. In the process of training the model, the validation data set is used for cross-validation. The validation data is divided into Each time you use Use one copy as training data and the remaining copy as validation data to calculate the validation error of the model. ,in is the actual value, is the predicted value, is the number of validation samples. Based on the validation error, the model parameters are optimized and the model parameters are updated by the gradient descent method to finally obtain the optimal model parameters. After the training is completed, the optimal model parameters and the current state vector are used , input the neural network Gaussian process model for forward calculation to predict the voltage at the next moment The prediction formula is expressed as:

[0064] ;

[0065] in are the optimized model parameters, It is the predicted noise. Through this process, the system can predict the future voltage change trend in real time based on historical data.

[0066] In one example, n linear state models are constructed according to characteristic data of the initial dynamic load model, and the n linear state models are input into the interactive multi-model analyzer to calculate the probability value of each linear state model to obtain the current optimal operation model, including:

[0067] According to different load conditions, the characteristic data of the initial dynamic load model are classified to obtain n load characteristic subsets;

[0068] Constructing state space equations for each load feature subset to obtain n groups of state space expressions, and constructing state transfer matrices and measurement matrices based on the state space expressions to obtain n linear state models;

[0069] Interactively calculate n linear state models, obtain the initial probability value of each linear state model based on the mixed probability density function, and input the initial probability value and the current measurement value into the maximum likelihood estimator to calculate the likelihood function value of each linear state model;

[0070] Based on the likelihood function value and the initial probability value, a Bayesian update calculation is performed to obtain the posterior probability value of each linear state model, and the posterior probability value is normalized to obtain the revised probability value of each linear state model;

[0071] According to the corrected probability value, the linear state model with the largest probability is selected as the current optimal operating model.

[0072] In this example, key features such as voltage are extracted from the feature data of the initial dynamic load model. , Current ,power , load fluctuation range and system response rate, etc. These characteristics are expressed in vector form ,in Represents the dynamic response rate of the load. By clustering these features, for example, using - Mean clustering or density-based spatial clustering, dividing feature data into Each load characteristic subset Represents a specific load condition, such as light load, heavy load or sudden load state. After completing the feature classification, each load feature subset Construct state space equations. The state space equations describe the dynamic behavior of the system with state variables, including state transfer equations and observation equations. Assume that the state vector of the system is , the control input is The observed output is , then the state transfer equation is expressed as:

[0073] ;

[0074] in is the state transfer matrix, which indicates the law of load state changing over time; is the control matrix, describing the effect of the input on the system state; is the process noise, assuming ,in is the covariance matrix of process noise. The observation equation is expressed as:

[0075] ;

[0076] in It is the measurement matrix, which describes the mapping relationship from state to observation; is the direct transfer matrix, reflecting the direct impact of the input on the observed output; is the measurement noise, assuming ,in is the covariance matrix of the measurement noise. Fit and obtain the corresponding state space expression and matrix , thus constructing In order to realize the interactive calculation of multiple linear state models, the probability value of each model is initialized based on the mixed probability density function. Assume that the state distribution of the current system is a mixed Gaussian distribution, which is expressed as:

[0077] ;

[0078] in It is a model The initial probability value satisfies , and are the mean and covariance of the states respectively. Combined with the current observation data , the initial probability value The measured values ​​are input into the maximum likelihood estimator to calculate the likelihood function value for each model:

[0079] ;

[0080] in Representation Model The degree of adaptation to the current observation value. Based on the likelihood function value and the initial probability value, the posterior probability value of each linear state model is calculated using the Bayesian update formula:

[0081] ;

[0082] in It is a model The posterior probability value. In order to facilitate comparison and subsequent processing, the posterior probability value is normalized to obtain the corrected probability value ,in According to the modified probability value , select the model with the highest probability as the optimal operating model of the current system.

[0083] In one example, based on the voltage prediction value and the optimal operation model, the output control quantity of the PD controller is expanded by Fourier series to obtain a set of Fourier coefficients, including:

[0084] The difference between the predicted voltage value and the standard voltage value is calculated to obtain a voltage error sequence, and the control equation of the PD controller is constructed based on the voltage error sequence to obtain a controller output sequence;

[0085] The controller output sequence is subjected to period boundary determination, the fundamental frequency and harmonic order are obtained by discrete Fourier transform, and the controller output sequence is expanded into Fourier series form according to the determined fundamental frequency and harmonic order to obtain the initial Fourier coefficients;

[0086] The initial Fourier coefficients are corrected and calculated according to the state equation of the optimal operation model to obtain a correction coefficient matrix, and the correction coefficient matrix is ​​weightedly combined with the initial Fourier coefficients to obtain weighted Fourier coefficients;

[0087] The orthogonality of the weighted Fourier coefficients is verified, the orthogonalized coefficients are obtained through Schmidt orthogonalization, and the orthogonalized coefficients are combined to form a Fourier coefficient set.

[0088] In this example, the voltage prediction value is And standard voltage value Perform difference calculation to obtain the voltage error sequence , the formula is:

[0089] ;

[0090] in represents the sampling time, is the voltage error value at that moment. The voltage error sequence reflects the deviation between the real-time state and the target state of the power supply system, and provides an adjustment basis for the controller. , construct the control equation of the PD (proportional-derivative) controller to generate the output sequence of the controller. The general form of the PD control equation is:

[0091] ;

[0092] in is the output sequence of the controller, and are proportional gain and differential gain respectively, is the time derivative of the voltage error, reflecting the dynamic characteristics of the error change. Through discretization, the time derivative is approximated by finite differences as follows:

[0093] ;

[0094] in is the sampling time interval. After substituting into the control equation, the output sequence of the controller is gradually calculated. , which is the direct reference signal for the power supply system to adjust the voltage output. After that, the frequency domain analysis is performed to reveal its main periodic characteristics and harmonic components. , that is, find the repetition period of the output signal. The period boundary is determined by analyzing the autocorrelation function of the output sequence, and the delay corresponding to its maximum value is the period length of the signal. After determining the period, the controller output sequence is transformed into the frequency domain using discrete Fourier transform, and the formula is:

[0095] ;

[0096] in is the Fourier coefficient of the frequency domain signal, is the frequency index, is an imaginary unit, is the cycle length. Through Fourier transform, we can get the fundamental frequency of the controller output sequence and main harmonic orders , which reflects the main frequency components of the output signal. Based on the above Fourier coefficients, the controller output sequence Expanded into Fourier series form, the formula is:

[0097] ;

[0098] in and are the real and imaginary parts of the Fourier coefficients respectively. The initial Fourier coefficients are directly obtained from In order to optimize the initial Fourier coefficients, the state equation of the optimal operation model is combined to modify the calculation. Assume that the state equation of the optimal operation model is:

[0099] ;

[0100] in is the state variable of the system, is the state transition matrix, is the control matrix, is the input signal (generated by Fourier expansion). By solving the state equation, we get the system response matrix to the Fourier coefficients , and accordingly the initial Fourier coefficients are corrected, the formula is:

[0101] ;

[0102] in is the modified coefficient matrix, is the initial Fourier coefficient matrix, Represents the state response matrix. The correction coefficient matrix is ​​weighted combined with the initial Fourier coefficients, and the formula is:

[0103] ;

[0104] in and is a weighting coefficient used to balance the relationship between the original characteristics and the model correction. Orthogonality verification is performed to ensure the independence of each harmonic component. The weighted Fourier coefficients are processed into orthogonalized coefficients using the Schmidt orthogonalization method. The orthogonalization process is:

[0105] ;

[0106] in After orthogonalization, Fourier coefficients, represents the inner product operation, Represents the second norm of the vector. Through this process, a set of mutually orthogonal Fourier coefficients are obtained. The orthogonalized coefficients are combined to form a Fourier coefficient set for optimal control of the power supply system.

[0107] In one example, a chance-constrained optimization calculation is performed based on the voltage prediction value and the Fourier coefficient set, and an optimal control sequence is obtained by solving the optimization objective function, including:

[0108] A voltage constraint condition set is constructed based on the voltage prediction value and a confidence level is set to obtain a probabilistic constraint condition;

[0109] A set of control quantity constraint conditions is constructed according to a set of Fourier coefficients to obtain control constraint conditions, and the probability constraint conditions and the control constraint conditions are transformed to obtain a set of deterministic constraint equations;

[0110] The optimization objective function is constructed based on the deterministic constraint equations, and the optimization objective function is decomposed by convex optimization, and the dual function is obtained by the Lagrange multiplier method;

[0111] Based on the dual function, the KKT conditional equations are constructed and the optimal solution is obtained by the gradient descent method;

[0112] The optimal solution is substituted into the optimization objective function, the control sequence is obtained through reverse iterative calculation, and the control sequence is dynamically optimized to obtain the optimal control sequence that meets the constraints.

[0113] In this example, the voltage prediction value is Construct a set of voltage constraints. Assume that the standard range of voltage is , and the voltage prediction value There is a certain degree of uncertainty, and the prediction error is described by a random variable ,Right now ,in is the expected value of the prediction represents the noise of Gaussian distribution. In order to introduce probability constraints, the confidence level is set (e.g. 95% or ), requiring the probability of voltage being within the standard range to reach The mathematical expression is:

[0114] ;

[0115] Through the distribution characteristics of random variables, the above probability constraints are transformed into deterministic constraints:

[0116] ;

[0117] in is the quantile function of the standard normal distribution, which represents the amplitude of the confidence interval. This constraint transforms the random characteristics of the voltage into the relationship between the expected value and the noise amplitude, forming a deterministic expression of the probability constraint. At the same time, based on the Fourier coefficient set Construct a set of control quantity constraints. Assume that the control quantity The amplitude needs to be limited to the range , the control quantity of Fourier series expansion is expressed as:

[0118] ;

[0119] The control constraints are written as:

[0120] ;

[0121] The above probability constraints and control constraints are integrated and transformed into a set of deterministic constraint equations, which can be expressed as ,in is a linear or nonlinear function of the control variable, is the constraint boundary. Based on the deterministic constraint equations, the optimization objective function is constructed. The optimization objective function It is defined as minimizing the total energy consumption of the system while ensuring the output voltage stability and the smoothness of the control signal. The objective function is in the form of:

[0122] ;

[0123] in is the smoothness weight coefficient, which represents the penalty term for the rate of change of the control variable. By performing convex optimization decomposition on the objective function, the optimization problem is transformed into a Lagrangian dual problem. The Lagrangian multiplier is introduced. and constraint functions , construct the Lagrangian function:

[0124] ;

[0125] in is a non-negative Lagrange multiplier. Solve the dual function for the Lagrangian function , that is, for the original variable Minimize:

[0126] ;

[0127] The Karush-Kuhn-Tucker (KKT) conditional equations are constructed based on the dual function. The KKT condition includes the following three parts: The gradient of the objective function must be zero, that is, ; Constraints Must satisfy; complementarity conditions between Lagrange multipliers and constraints . Use the gradient descent method to iteratively solve the KKT conditional equations and update the Fourier coefficient set and Lagrange multipliers , until it converges to the optimal solution. Substitute the optimization objective function and generate the control sequence through reverse iterative calculation In order to optimize the global performance of the control sequence, dynamic programming technology is introduced to decompose the problem into multiple stages. In each stage, sub-problems are solved and the optimal solution is synthesized in reverse. The recursive relationship of dynamic programming is:

[0128] ;

[0129] in It is a stage The optimal value function of is a single-stage objective function, is the next state.

[0130] In one example, a control signal is output to a power supply unit of a POE power supply system based on an optimal control sequence, and second voltage and current data are collected and fed back to an initial dynamic load model for dynamic optimization to obtain a target dynamic load model, including:

[0131] The optimal control sequence is converted into a control signal through a digital-to-analog converter, and the control signal is output to the POE power supply unit to obtain a control quantity;

[0132] Sampling the voltage and current of the output port of the POE power supply unit to obtain second voltage and current data, and calculating the real-time power based on the second voltage and current data to obtain a power data sequence;

[0133] Perform sliding window processing on the power data sequence to obtain a dynamic power change curve, and fuse the dynamic power change curve with the initial dynamic load model to obtain an updated power feature matrix;

[0134] Reconstruct the state space equations based on the updated power characteristic matrix to obtain a new state equation coefficient matrix;

[0135] The new state equation coefficient matrix is ​​optimized and calculated online, and the optimized coefficient matrix is ​​obtained by recursive least square method. The optimized coefficient matrix is ​​substituted into the initial dynamic load model for parameter update to obtain the target dynamic load model.

[0136] In this example, the optimal control sequence Convert the digital control signal into an actual control signal. Convert to analog control voltage . Analog control signal It is the direct input quantity of the power supply unit, acting on the power module of the POE power supply unit, adjusting the voltage and current of the output port to form a control quantity. After the power supply unit responds to the control signal, it collects the actual voltage of the output port. and current , forming the second voltage and current data sequence Through these data, real-time power Calculated by the formula:

[0137] ;

[0138] in is the output voltage, is the output current. The power data sequence obtained is Describe the energy output characteristics of the power supply system under the control sequence. In order to analyze the dynamic characteristics of power data, the power data sequence Sliding window processing is performed. The sliding window technology is used to extract the local characteristics of the data. Suppose the window length is , the sliding window moves point by point in the sequence, and the mean and fluctuation range within a window are calculated each time, which is defined as:

[0139] ;

[0140] in is the average power in the window, is the standard deviation, reflecting the local fluctuation. Sliding window processing generates dynamic power change curve , reflecting the changing trend and fluctuation characteristics of power in time series. The dynamic power change curve is integrated with the data of the initial dynamic load model to update the power characteristic matrix The power feature matrix includes the mean, fluctuation range, maximum and minimum values ​​of real-time power, for example:

[0141] ;

[0142] in and are the maximum and minimum values ​​in the sliding window respectively. Based on the updated power feature matrix , reconstruct the state space equations. Assume that the state vector of the system is , the input is the control quantity , the output is power , the state equation and observation equation are:

[0143] ;

[0144] ;

[0145] in is the state transition matrix, describing the change of state over time; is the input matrix, which represents the influence of the control input on the state; is the output matrix, mapping states to power outputs; is the direct transfer matrix, describing the part of the control input that directly affects the power; and They are process noise and measurement noise respectively, assuming they follow Gaussian distribution In order to make the state space equation consistent with the real-time data, the new state equation coefficient matrix Perform online optimization calculations. Use the recursive least squares method to update the coefficient matrix. The update formula of the recursive least squares method is:

[0146] ;

[0147] ;

[0148] ;

[0149] in is a vector of parameter estimates containing The expanded form of is the regression vector, containing the state and input; is the error covariance matrix; is the forgetting factor, which is used to control the influence of historical data. Through the recursive least squares method, the state equation coefficient matrix is ​​dynamically adjusted to keep it consistent with the real-time data. Substitute the initial dynamic load model, complete the parameter update, and obtain the target dynamic load model.

[0150] Reference Figure 2 This embodiment provides an adaptive power supply device for a POE power supply, comprising:

[0151] The acquisition module 1 is used to collect the first voltage and current data of each port of the POE power supply system, calculate the real-time power value, and establish an initial dynamic load model including the power change curve and load characteristics;

[0152] Calculation module 2, used for inputting the historical data sequence of the initial dynamic load model into the neural network Gaussian process analyzer for Bayesian reasoning training, obtaining model parameters, and calculating the voltage prediction value at the next moment based on the model parameters;

[0153] A construction module 3 is used to construct n linear state models according to the characteristic data of the initial dynamic load model, and input the n linear state models into the interactive multi-model analyzer to calculate the probability value of each linear state model to obtain the current optimal operation model;

[0154] An expansion module 4 is used to perform Fourier series expansion on the output control quantity of the PD controller based on the voltage prediction value and the optimal operation model to obtain a Fourier coefficient set;

[0155] A solution module 5 is used for performing a chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving the optimization objective function;

[0156] The dynamic optimization module 6 is used to output a control signal to the power supply unit of the POE power supply system based on the optimal control sequence, collect the second voltage and current data and feed it back to the initial dynamic load model for dynamic optimization to obtain a target dynamic load model.

[0157] 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.

[0158] 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 3 As 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.

[0159] 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.

[0160] 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.

[0161] 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.

[0162] 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.

[0163] 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 POE power supply adaptive power supply method, characterized in that: The following steps are involved: The first voltage and current data of each port of the POE power supply system are collected, and the real-time power value is calculated, and an initial dynamic load model including a power change curve and load characteristics is established at the same time; specifically including: setting a sampling time interval for each port of the POE power supply system, collecting the port voltage value and current value at each sampling time point to obtain an original sampling data set, and performing a product operation based on the voltage value and the current value in the original sampling data set to obtain the real-time power value of each port; sorting and storing the real-time power values ​​according to a time series, and constructing a power time series including multiple sampling points; performing curve fitting based on the power time series to obtain a power change curve coefficient matrix, constructing a load feature vector based on the original sampling data set, and performing principal component analysis on the load feature vector to obtain a load feature matrix; performing dynamic model construction and state equation modeling on the power change curve coefficient matrix and the load feature matrix to obtain an initial state space equation group; setting power supply constraints on the initial state space equation group to obtain a constrained state space equation group, and performing online parameter identification on the constrained state space equation group and the real-time sampling data to obtain an initial dynamic load model; Inputting the historical data sequence of the initial dynamic load model into a neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters; Constructing n linear state models according to the characteristic data of the initial dynamic load model, and inputting the n linear state models into an interactive multi-model analyzer, calculating the probability value of each linear state model, and obtaining a current optimal operation model; Based on the voltage prediction value and the optimal operation model, performing Fourier series expansion on the output control quantity of the PD controller to obtain a set of Fourier coefficients; Performing a chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving an optimization objective function; Based on the optimal control sequence, a control signal is output to the power supply unit of the POE power supply system, and the second voltage and current data are collected and fed back to the initial dynamic load model for dynamic optimization to obtain a target dynamic load model.

2. The adaptive power supply method of POE power supply according to claim 1, characterized in that: The inputting the historical data sequence of the initial dynamic load model into the neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters, includes: Dividing the historical data sequence of the initial dynamic load model to obtain a training data set and a verification data set, and standardizing the training data set to obtain a standardized feature vector and a standardized label; Constructing a Gaussian kernel function based on the standardized eigenvector, and performing eigendecomposition on the covariance matrix of the Gaussian kernel function to obtain an eigenvalue matrix and an eigenvector matrix; Performing Bayesian prior distribution calculation on the eigenvalue matrix and the eigenvector matrix to obtain the prior probability distribution of the model parameters, constructing a likelihood function based on the prior probability distribution and the standardized labels, and obtaining the posterior distribution of the model parameters by maximum a posteriori probability calculation; Inputting the posterior distribution of the model parameters into the neural network Gaussian process model, cross-validating the validation data set to obtain a model validation error, and optimizing the parameters of the neural network Gaussian process model based on the validation error to obtain optimal model parameters; The state vector at the current moment is input into the neural network Gaussian process model, and forward calculation is performed based on the optimal model parameters to obtain the voltage prediction value at the next moment.

3. The adaptive power supply method of POE power supply according to claim 2, characterized in that: The method comprises: constructing n linear state models according to the characteristic data of the initial dynamic load model, inputting the n linear state models into an interactive multi-model analyzer, calculating the probability value of each linear state model, and obtaining the current optimal operation model, including: Classifying the characteristic data of the initial dynamic load model according to different load conditions to obtain n load characteristic subsets; Constructing a state space equation for each of the load feature subsets to obtain n groups of state space expressions, and constructing a state transfer matrix and a measurement matrix based on the state space expressions to obtain n linear state models; Interactively calculating the n linear state models, obtaining an initial probability value of each linear state model based on a mixed probability density function, and inputting the initial probability value and a current measurement value into a maximum likelihood estimator to calculate a likelihood function value of each linear state model; Performing a Bayesian update calculation based on the likelihood function value and the initial probability value to obtain a posterior probability value of each linear state model, and normalizing the posterior probability value to obtain a revised probability value of each linear state model; According to the modified probability value, the linear state model with the largest probability is selected as the current optimal operation model.

4. The adaptive power supply method of POE power supply according to claim 3, characterized in that: Based on the voltage prediction value and the optimal operation model, the output control quantity of the PD controller is expanded by Fourier series to obtain a set of Fourier coefficients, including: Calculating the difference between the voltage prediction value and the standard voltage value to obtain a voltage error sequence, and constructing a control equation of a PD controller based on the voltage error sequence to obtain a controller output sequence; Determine the period boundary of the controller output sequence, obtain the fundamental frequency and harmonic order by discrete Fourier transform, and expand the controller output sequence into Fourier series form according to the determined fundamental frequency and harmonic order to obtain initial Fourier coefficients; Performing correction calculation on the initial Fourier coefficients according to the state equation of the optimal operation model to obtain a correction coefficient matrix, and performing weighted combination of the correction coefficient matrix and the initial Fourier coefficients to obtain weighted Fourier coefficients; The orthogonality of the weighted Fourier coefficients is verified, orthogonalized coefficients are obtained by Schmidt orthogonalization processing, and the orthogonalized coefficients are combined to form a Fourier coefficient set.

5. The adaptive power supply method of POE power supply according to claim 4, characterized in that: The performing of chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving an optimization objective function, comprises: Constructing a voltage constraint condition set based on the voltage prediction value and setting a confidence level to obtain a probability constraint condition; Constructing a set of control quantity constraint conditions according to the set of Fourier coefficients to obtain control constraint conditions, and transforming the probability constraint conditions and the control constraint conditions to obtain a set of deterministic constraint equations; Constructing an optimization objective function based on the deterministic constraint equation group, performing convex optimization decomposition on the optimization objective function, and obtaining a dual function through a Lagrange multiplier method; Constructing a KKT conditional equation group based on the dual function, and solving it by gradient descent method to obtain the optimal solution; The optimal solution is substituted into the optimization objective function, a control sequence is obtained through reverse iterative calculation, and the control sequence is dynamically optimized to obtain an optimal control sequence that meets the constraint conditions.

6. The adaptive power supply method of POE power supply according to claim 5, characterized in that: The method outputting a control signal based on the optimal control sequence to the power supply unit of the POE power supply system, collecting the second voltage and current data and feeding it back to the initial dynamic load model for dynamic optimization, and obtaining a target dynamic load model includes: Converting the optimal control sequence into a control signal through a digital-to-analog converter, and outputting the control signal to a POE power supply unit to obtain a control amount; Sampling the voltage and current of the output port of the POE power supply unit to obtain second voltage and current data, and calculating the real-time power based on the second voltage and current data to obtain a power data sequence; Performing sliding window processing on the power data sequence to obtain a dynamic power change curve, and performing data fusion on the dynamic power change curve and the initial dynamic load model to obtain an updated power characteristic matrix; Reconstruct the state space equations based on the updated power characteristic matrix to obtain a new state equation coefficient matrix; The new state equation coefficient matrix is ​​optimized and calculated online, and an optimized coefficient matrix is ​​obtained by recursive least square method, and the optimized coefficient matrix is ​​substituted into the initial dynamic load model for parameter update to obtain a target dynamic load model.

7. An adaptive power supply device for a POE power supply, characterized in that: The steps for implementing the adaptive power supply method of the POE power supply according to any one of claims 1 to 6, wherein the adaptive power supply device of the POE power supply comprises: The acquisition module is used to collect the first voltage and current data of each port of the POE power supply system, calculate the real-time power value, and establish an initial dynamic load model including the power change curve and load characteristics; A calculation module, used for inputting the historical data sequence of the initial dynamic load model into a neural network Gaussian process analyzer for Bayesian reasoning training to obtain model parameters, and calculating the voltage prediction value at the next moment based on the model parameters; A construction module, used to construct n linear state models according to the characteristic data of the initial dynamic load model, and input the n linear state models into an interactive multi-model analyzer, calculate the probability value of each linear state model, and obtain the current optimal operation model; An expansion module, used for performing Fourier series expansion on the output control quantity of the PD controller based on the voltage prediction value and the optimal operation model to obtain a set of Fourier coefficients; A solution module, used for performing chance-constrained optimization calculation based on the voltage prediction value and the Fourier coefficient set, and obtaining an optimal control sequence by solving an optimization objective function; A dynamic optimization module is used to output a control signal to the power supply unit of the POE power supply system based on the optimal control sequence, collect the second voltage and current data and feed it back to the initial dynamic load model for dynamic optimization to obtain a target dynamic load model.

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