A method and system for intelligently controlling the operation status of a DC power supply

Abstract

The present invention relates to a method and system for intelligently controlling the operating state of a direct current power supply. The method includes: modeling multiple components of a direct current power supply system, obtaining a system dynamic model and performing loss characteristic analysis to obtain system total loss distribution data; calculating an optimal operating parameter combination based on the system total loss distribution data; creating a first power supply control strategy combining feedforward compensation and output voltage feedback; inputting the real-time operating data of the direct current power supply system into a multi-branch graph convolutional neural network for feature extraction, and obtaining an intelligent power supply control model through an adaptive learning algorithm; dynamically adjusting state transfer parameters based on the prediction results of the first power supply control strategy and the intelligent power supply control model to obtain a second power supply control strategy. The implementation of the present invention can comprehensively consider multiple factors such as the dynamic characteristics of the system, loss distribution, and load changes, and can adaptively adjust the control strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply operation status monitoring, and in particular to a method and system for intelligently controlling the operation status of a direct current power supply. Background Art

[0002] Traditional DC power supply control methods often have difficulty coping with complex and changeable operating environments and load conditions, resulting in problems such as low system efficiency, slow dynamic response, and insufficient reliability. Especially when faced with large-scale distributed energy access and highly dynamic load changes, existing control strategies are difficult to accurately analyze and optimize system losses, thus affecting the performance and economy of the overall system. In addition, when dealing with high-dimensional, nonlinear, and time-varying DC power supply systems, how to extract stable and reliable features and achieve robust control conversion between different operating domains has become an urgent problem to be solved. Summary of the invention

[0003] The main purpose of the present invention is to provide a method and system for intelligently controlling the operating status of a DC power supply, so as to comprehensively consider the dynamic characteristics of the system, loss distribution, load changes and other factors, and adaptively adjust the control strategy.

[0004] To achieve the above object, the present invention provides a method for intelligently controlling the operating state of a DC power supply, comprising the following steps:

[0005] Modeling multiple components of a DC power supply system to obtain a system dynamic model, and based on the system dynamic model, performing loss characteristic analysis on the multiple components to obtain system total loss distribution data;

[0006] Based on the total loss distribution data of the system, analyzing the impact of operating parameters on the loss of the DC power supply system, and calculating the optimal operating parameter combination;

[0007] According to the optimal operating parameter combination, a first power supply control strategy combining feedforward compensation and output voltage feedback is created, and an initial control amount is generated;

[0008] Inputting the real-time operation data of the DC power supply system into a multi-branch graph convolutional neural network for feature extraction, constructing domain-invariant features, updating the initial control quantity through an adaptive learning algorithm, and obtaining an intelligent power supply control model;

[0009] Based on the prediction results of the first power control strategy and the intelligent power control model, the state space of the DC power supply system is analyzed, the state duration and transfer conditions are determined, and the state transfer parameters are dynamically adjusted to obtain the second power control strategy.

[0010] The present invention also provides a DC power supply operating state intelligent control device, comprising:

[0011] A modeling module, used to model multiple components of a DC power supply system to obtain a system dynamic model, and based on the system dynamic model, perform loss characteristic analysis on the multiple components to obtain system total loss distribution data;

[0012] A calculation module, used for analyzing the influence of operating parameters on the loss of the DC power supply system based on the total loss distribution data of the system, and calculating the optimal combination of operating parameters;

[0013] A creation module, used for creating a first power supply control strategy combining feedforward compensation and output voltage feedback according to the optimal operating parameter combination, and generating an initial control amount;

[0014] An extraction module, used for inputting the real-time operation data of the DC power supply system into a multi-branch graph convolutional neural network for feature extraction, constructing domain-invariant features, updating the initial control amount through an adaptive learning algorithm, and obtaining an intelligent power supply control model;

[0015] An adjustment module is used to analyze the state space of the DC power supply system based on the prediction results of the first power supply control strategy and the intelligent power supply control model, determine the state duration and transfer conditions, dynamically adjust the state transfer parameters, and obtain the second power supply control strategy.

[0016] 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.

[0017] 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.

[0018] In summary, the technical solution provided by the present invention can obtain accurate system dynamic models and loss distribution data by modeling multiple components of the DC power supply system in detail and combining loss characteristic analysis. By calculating the optimal operating parameter combination, the system loss can be minimized under different load conditions. It can dynamically adapt to load changes and significantly improve the energy efficiency of the system. The control strategy combined with feedforward compensation and output voltage feedback can not only improve the dynamic response speed of the system, but also effectively suppress the influence of load disturbance on the output voltage, thereby improving the stability and reliability of the system. The multi-branch graph convolutional neural network is used for feature extraction, which can effectively capture the spatial, frequency and timing characteristics of the system. The modeling and prediction capabilities of complex nonlinear systems are improved. Through adaptive instance normalization and adaptive learning algorithms, the system can generate domain invariant features, which improves the generalization ability and robustness of the control model under different operating environments. Based on the state space analysis of the control error sequence and the dynamically adjusted state transition parameters, the accurate division and smooth switching of the system operating state are realized. This multi-mode control strategy can better adapt to different operating stages of the system and improve the accuracy and flexibility of control. By adjusting control parameters with real-time feedback, the system can continuously optimize its performance and adapt to environmental changes and load fluctuations, thereby achieving long-term stable and efficient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the steps of a method for intelligently controlling the operating state of a DC power supply in one embodiment of the present invention;

[0020] Figure 2 It is a structural block diagram of an intelligent control device for the operation status of a DC power supply in one embodiment of the present invention;

[0021] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0022] 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

[0023] 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.

[0024] Reference Figure 1 , this embodiment provides a method for intelligently controlling the operation state of a DC power supply, comprising the following steps:

[0025] S1, modeling multiple components of the DC power supply system to obtain a system dynamic model, and based on the system dynamic model, analyzing the loss characteristics of the multiple components to obtain system total loss distribution data;

[0026] Among them, the multiple components of the DC power supply system include a DC converter, an energy input terminal, and a load output terminal. The circuit topology analysis of the DC converter is carried out, and a dynamic model of the DC converter is formed by establishing a state equation including the on-state and off-state of the switch device. The dynamic model can reflect the circuit behavior of the DC converter under various operating states, including the conversion of the two main states of on and off. Through modeling, the working characteristics of the switch device in different states and the impact on the overall performance of the system can be analyzed more accurately. The voltage and current characteristics of the energy input terminal are analyzed, and a mathematical expression considering the input impedance and input voltage fluctuation is established to obtain an equivalent model of the energy input terminal. The model reflects the impact of input voltage fluctuation on the stability of the DC power supply system. At the same time, the load characteristics of the load output terminal are analyzed. By establishing a mathematical description including load impedance and power demand, an equivalent model of the load output terminal is obtained. The model can reflect the operating state of the system under different load conditions, especially when the load changes, how the power demand of the load terminal affects the overall working performance of the system. The dynamic model of the DC converter, the equivalent model of the energy input terminal, and the equivalent model of the load output terminal are coupled and analyzed to establish the state space equation of the system. Through coupling analysis, the overall dynamic model of the DC power supply system is obtained, which can fully describe the interaction between the various components in the system and their behavioral changes under different states. Based on the system dynamic model, the loss characteristics of multiple components of the DC power supply system are analyzed, the energy loss of each component under different operating conditions is evaluated, and the total loss distribution data of the system is obtained.

[0027] Furthermore, based on the dynamic model of the DC converter, the conduction loss and switching loss of the switching device are analyzed, and a power loss equation including the switching frequency and duty cycle is established to form a DC converter loss model. This model reflects the loss of the switching device under different working conditions, and thus reveals the overall energy loss behavior of the DC converter. According to the energy input equivalent model, the impact of input voltage fluctuations on system efficiency is analyzed, and an input loss function is established to reflect how input voltage changes affect the energy efficiency performance of the entire system. A loss model for the energy input end is obtained to describe the change law of system loss under various input voltage conditions. Based on the load output equivalent model, the impact of load changes on system efficiency is analyzed, and a relationship function between load power and system loss is established to obtain a load output loss model to reveal the loss distribution of the system under different load conditions. The DC converter loss model, the energy input loss model, and the load output loss model are superimposed to construct a loss model for the entire system, which comprehensively reflects the energy loss of each part of the entire DC power supply system under different operating conditions. The overall loss model of the system is numerically calculated under various operating conditions to generate system total loss distribution data.

[0028] S2, based on the total system loss distribution data, analyze the impact of operating parameters on the loss of the DC power supply system and calculate the optimal operating parameter combination;

[0029] Specifically, the parameter sensitivity analysis of the overall loss model of the system is carried out to determine the key operating parameter set including switching frequency, duty cycle and output voltage. The optimization objective function is constructed based on the key operating parameter set, which is used to guide the subsequent optimization process to ensure that the optimal parameter combination that can minimize the system loss is found. Based on the optimization objective function, the initial parameters of the particle swarm optimization algorithm are set, including the number of particles, the number of iterations, the inertia weight and the learning factor. These parameters determine the initial state of the algorithm and its ability to search the solution space. According to these parameters, the initialized particle swarm is obtained, which contains multiple candidate solutions, each of which corresponds to a different combination of operating parameters. The initialized particle swarm is substituted into the overall loss model of the system to calculate the fitness value of each particle under different load conditions. The fitness value reflects the performance of each candidate solution under specific load conditions. By calculating the fitness value, the initial optimal solution is obtained, which represents the best operating parameter combination under the current particle swarm. Based on the initial optimal solution, the speed and position of the particle swarm are updated. By updating the position of the particles in each iteration, the particle swarm gradually approaches the global optimal solution. In each iteration, a local optimal solution and a global optimal solution are generated. The local optimal solution is the best solution in a certain iteration, while the global optimal solution is the solution with the best overall performance in all iterations. In order to ensure the stability and accuracy of the optimization results, the convergence analysis of the global optimal solution of each iteration is performed. By analyzing the convergence of the global optimal solution, it is determined whether the preset termination condition is met. If the judgment result shows that the algorithm has not converged or reached the global optimum, iterative optimization will continue, and the particle swarm will be updated until the termination condition is reached. When the iterative process meets the termination condition, the final optimal operating parameter combination is output.

[0030] S3, creating a first power supply control strategy combining feedforward compensation and output voltage feedback according to the optimal operating parameter combination, and generating an initial control quantity;

[0031] It should be noted that based on the optimal operating parameter combination and the system dynamic model, the system control equation including the switching frequency and duty cycle is constructed to obtain the system control model. The control model can reflect the dynamic behavior of the DC power supply system, and determine the operating state of the system by combining parameters such as the switching frequency and duty cycle. In order to simplify the control complexity of the system, the system control model is linearized so that it can be adjusted using a linear control strategy within a certain operating range. A proportional integral (PI) controller is designed as the feedback controller of the output voltage. By calculating the appropriate proportional integral parameters, the feedback control component used to adjust the system output voltage is obtained. This component can help the system quickly recover to a stable state when disturbed. At the same time, a feedforward compensator is constructed based on the overall loss model of the system and the optimal operating parameter combination. The compensator is used to predict the impact of load disturbances on the system output voltage and perform feedforward compensation according to the load changes. By calculating the impact of load disturbances, the feedforward compensator generates a feedforward control component, which can adjust the output in advance when the system is subject to load changes, thereby reducing the impact of load fluctuations on system stability. The feedback control component and the feedforward control component are weighted superimposed to form a composite control quantity. The composite control quantity combines the advantages of feedforward compensation and feedback control, and can better cope with various disturbances inside and outside the system. In order to ensure the effectiveness of the composite control quantity, a stability analysis is performed on it, and the closed-loop pole position of the system is calculated to determine the stability of the system. Through this analysis, the stability performance of the system under the current control strategy is clarified, and the system stability judgment result is generated. According to the stability judgment result, if it is found that the composite control quantity fails to meet the system stability requirements, the parameters are adjusted and optimized to correct the composite control quantity. Through this process, it is ensured that the control strategy maintains the stability and efficiency of the system in practical applications. The corrected composite control quantity is substituted into the system control equation to generate the initial control quantity, which is used as the output of the first power supply control strategy. This strategy can effectively combine feedforward compensation and feedback regulation to achieve precise control of the DC power supply system, while ensuring the stable operation of the system, maximizing the system's operating efficiency and response speed.

[0032] S4, input the real-time operation data of the DC power supply system into a multi-branch graph convolutional neural network for feature extraction, construct domain-invariant features, update the initial control quantity through an adaptive learning algorithm, and obtain an intelligent power supply control model;

[0033] Specifically, a multi-scale wavelet transform is performed on the real-time operation data of the DC power supply system to decompose the time-frequency domain characteristics of the original signal to obtain a time-frequency domain feature matrix, which contains the characteristics of the operation data at different time scales and frequencies. A multi-branch graph convolutional neural network is constructed, and the network design includes three branches: spatial branch, frequency branch and time series branch. Each branch contains a graph attention layer, a dynamic graph convolution layer and a jump connection structure. This design can help the network capture more complex features. In the spatial branch, the time-frequency domain feature matrix is ​​input, and the node features are weighted through the spatial graph attention mechanism to generate spatial attention features. To enhance the feature extraction capability, the attention mechanism adopts a multi-head attention structure with 8 heads and a dimension of 64 for each head. A dynamic graph convolution operation is performed on the spatial attention feature to dynamically update the adjacency matrix of the graph and generate adaptive spatial features. The convolution layer contains 3 layers and uses a gated linear unit as an activation function to improve the nonlinear processing capability. In the frequency branch, the time-frequency domain feature matrix is ​​input, and the importance of different frequency components is sorted through the frequency graph attention mechanism to extract the frequency attention feature. In order to further explore the features of the frequency level, the wavelet graph convolution operation is used to combine the multi-scale wavelet coefficients to obtain multi-resolution frequency features. This process can effectively retain the multi-level detailed features of the frequency information, making the system more sensitive and accurate in the frequency dimension. For the time series branch, the time-frequency domain feature matrix captures the long-term and short-term dependencies through the time series graph attention mechanism to generate time series attention features. The time series graph attention mechanism adopts the Transformer structure, which contains 4 encoder layers, each with 8 attention heads. Through this structure, the complex long-term and short-term dependency problems in the time series are effectively handled, and the model's ability to capture time features is enhanced. The causal graph convolution operation is performed on the time series attention features, and the cause-and-effect relationship in the time series data is maintained through the dilated convolution (the dilated rate is [1, 2, 4, 8]), and the time series dependency features are generated to reflect the correlation between the system at different time nodes. The adaptive spatial features, multi-resolution frequency features, and time series dependency features are cross-branch feature fusion to obtain a comprehensive feature representation. In order to make the comprehensive feature stable in different data domains, it is adaptively normalized to obtain domain-invariant features, which help the model maintain consistent performance under different working conditions and is not affected by data distribution differences. Based on the domain-invariant features, an adaptive learning algorithm is designed to dynamically update the initial control quantity. Through this algorithm, the model can adjust the control strategy according to real-time data, and finally obtain an intelligent power supply control model to achieve intelligent control and optimization of the operating status of the DC power supply system.

[0034] In this embodiment, the feature dimensions of adaptive spatial features, multi-resolution frequency features, and time-dependent features are aligned. A fully connected layer is used to map each feature to the same dimensional space to ensure that the data structures of different feature sources are consistent in dimension and form a unified dimensional feature. Based on the unified dimensional feature, the importance of each branch feature is evaluated by constructing an attention-guided feature aggregation module. This module can dynamically allocate the influence of each feature by calculating the attention weights of each feature branch. The attention mechanism calculates the weight of each feature according to its contribution to the task, and weights each feature to obtain a weighted feature representation. By reasonably allocating feature weights, more useful information can be used preferentially in different scenarios to improve the overall performance of the model. Nonlinear feature conversion and dimensionality reduction operations are performed on the weighted feature representation to extract deep information from the features and reduce redundant features. Nonlinear conversion uses activation functions such as ReLU or GELU to enhance the expressiveness of the model, while dimensionality reduction simplifies the feature representation through operations such as PCA (principal component analysis) or pooling to obtain preliminary fusion features. In order to ensure the integrity of information and the transfer of gradients, residual connections are performed on the preliminary fused features so that the original features and processed features can be effectively combined. At the same time, layer normalization is used to ensure that the data distribution between different feature dimensions is standardized and a comprehensive feature representation is generated. An adaptive instance normalization layer is constructed based on the comprehensive feature representation. The normalization parameters, including scaling factors and offsets, are dynamically calculated based on the features of each input sample. This layer adaptively adjusts the scaling factors and offsets based on the distribution of the comprehensive feature representation so that the normalized features can maintain stable performance in different samples. By learning the normalization parameter set, it is ensured that each sample feature can match its own characteristics without being affected by other samples. The comprehensive feature representation and the normalization parameter set are input into the adaptive instance normalization layer for feature normalization to ensure that the features remain consistent under different working conditions or different data distributions, and domain-invariant features are obtained.

[0035] S5, based on the prediction results of the first power control strategy and the intelligent power control model, analyze the state space of the DC power system, determine the state duration and transfer conditions, dynamically adjust the state transfer parameters, and obtain the second power control strategy.

[0036] Among them, the prediction results of the first power control strategy and the intelligent power control model are compared, and the control deviation between the two is calculated to obtain the control error sequence of the system, which reflects the difference between the actual behavior of the system under the current control strategy and the model prediction. The mean and standard deviation of the control error sequence are calculated to evaluate the steady-state error of the system, that is, the stability performance of the system under different operating states. Based on the steady-state error index, the threshold of the system state transition is set, and different operating state intervals are clearly divided. The division of the state space helps to understand the characteristics of the system under various operating conditions and ensure that the control strategy can make corresponding adjustments for different states. The system response characteristics under each divided state are analyzed, the duration of each state and the frequency of state transition are calculated, and the characteristics of the state transition are obtained, which reflects the behavior law of the system when it transitions from one operating state to another, and helps to determine how the system should adjust the state switching conditions under different operating conditions. According to the state transition characteristics, an adaptive state retention time algorithm is executed. The algorithm can dynamically adjust the state switching conditions of the system to ensure that the system can transition stably and efficiently under different operating conditions, avoiding unnecessary energy loss and system instability caused by frequent state switching. Through the adaptive algorithm, the state transition rules can be optimized to ensure the smoothness and response speed of the system operation. Based on the optimized state transition rules, a state switching controller is constructed to perform smooth transitions between different operating states. The function of the state controller is to smoothly transition between different operating states of the system to avoid sudden state jumps that cause system instability or reduced efficiency. Through the function of the state switching controller, a multi-mode control strategy is implemented, that is, the control scheme is dynamically adjusted according to the actual operating state to ensure that the system can maintain the best operating state under various conditions. The multi-mode control strategy is integrated with the first power supply control strategy to generate a comprehensive control output. Closed-loop correction is performed on the comprehensive control output, and the control parameters are adjusted through real-time feedback to ensure that the deviation of the system in actual operation is corrected in time, and finally the second power supply control strategy is generated.

[0037] Furthermore, the control error sequence is processed by sliding window, and the entire control error sequence is divided into multiple continuous subsequences, each of which represents the error fluctuation in a certain period of time. The mean and standard deviation of each subsequence are calculated respectively to obtain the error index reflecting the steady-state characteristics of the system, revealing the error distribution of the system in different time periods. The system steady-state error index is input into the kernel density estimation algorithm to calculate the probability density function of the error distribution. Kernel density estimation is a non-parametric statistical method that can smooth the error distribution and help analyze the error trend more intuitively. The maximum inter-class variance method is used to determine the initial threshold. This method can effectively find the preliminary classification boundary of the error sequence by maximizing the variance between different categories, and generate a candidate threshold set, which contains a series of candidate thresholds that may be used to divide different operating states. Each threshold in the candidate threshold set is evaluated, and the ratio of the intra-class variance to the inter-class variance is calculated. The intra-class variance represents the degree of discreteness of data within the same category, while the inter-class variance reflects the degree of separation between different categories. By comparing the ratio of the intra-class and inter-class variances, the advantages and disadvantages of each threshold are evaluated, and the threshold evaluation index is obtained. According to this indicator, the optimal threshold is selected as the final state transfer threshold. The optimal threshold distinguishes different operating states to the greatest extent, ensuring the accuracy and rationality of state division. Based on the determined final threshold, the control error sequence is segmented and the error sequence is divided into multiple different operating states, each of which represents the steady state or transition state of the system under specific conditions. The adjacent states of the preliminary state division results are merged. Through the merging process, frequent state switching caused by error fluctuations is avoided and the stability of the system is improved. The minimum state duration is set to ensure that each state has sufficient duration in actual operation to avoid energy efficiency reduction or control instability caused by frequent state switching of the system. The state space division result of the system is obtained.

[0038] In one example, multiple components of a DC power supply system are modeled to obtain a system dynamic model, and based on the system dynamic model, loss characteristics of the multiple components are analyzed to obtain total system loss distribution data, including:

[0039] The multiple components of a DC power system include a DC converter, energy input, and load output;

[0040] Conduct circuit topology analysis on the DC converter, establish state equations including the on-state and off-state of the switch device, and obtain the dynamic model of the DC converter;

[0041] Analyze the voltage and current characteristics of the energy input terminal, establish a mathematical expression that takes into account input impedance and input voltage fluctuations, and obtain an equivalent model of the energy input terminal;

[0042] Analyze the load characteristics of the load output terminal, establish a mathematical description including load impedance and power demand, and obtain an equivalent model of the load output terminal;

[0043] The dynamic model of the DC converter, the equivalent model of the energy input end and the equivalent model of the load output end are coupled and analyzed to establish the system state space equation and obtain the system dynamic model;

[0044] Based on the system dynamic model, the loss characteristics of multiple components are analyzed to obtain the total system loss distribution data.

[0045] In this example, during the operation of the DC power system, multiple components work together, including the DC converter, energy input, and load output. Perform circuit topology analysis on the DC converter. The DC converter is one of the core components of the entire system. Its function is to convert the input DC power into an output voltage that meets the load requirements. In order to accurately describe the dynamic behavior of the DC converter, a state equation containing the on-state and off-state of the switching device is established to reflect the voltage and current changes generated by the switching device during the on-and-off process. Set the current and voltage For time The state equation of the converter is expressed as:

[0046] ;

[0047] in, represents the converter inductance, is the input voltage, is the output voltage, is the rate of change of current. This equation changes in different switching states: when the switch is turned on, current flows through the inductor and energy is stored; when it is turned off, the energy in the inductor is released to the load. Through this equation, a dynamic model of the DC converter is established to describe the changes in current and voltage at different time points, reflecting the working characteristics of the DC converter. The voltage and current characteristics of the energy input end are analyzed to establish a model that takes into account the input impedance. The mathematical expression of the input voltage fluctuation is used to describe the characteristics of the energy input terminal. The input voltage fluctuation can be expressed as:

[0048] ;

[0049] in, is the nominal voltage at the input terminal, Indicates the fluctuation of input voltage. Input current It is expressed as:

[0050] ;

[0051] Through the above expression, we get the equivalent model of the energy input end, which reflects the influence of input voltage and input impedance on the input current of the system, especially the influence of input voltage fluctuation on the stability and energy efficiency of the whole system. At the same time, the load characteristics of the load output end are analyzed. The load usually appears as a resistive or inductive load, and its power demand changes over time. In order to describe the behavior of the load end, a load impedance is established. and power requirements The mathematical model of the load impedance and voltage is expressed as:

[0052] ;

[0053] in, is the instantaneous power demand of the load, is the output voltage at the load end, is the load impedance. This model is used to describe the changes in power demand at the load end under different operating conditions and reflect the working state of the system under different loads. The dynamic model of the DC converter, the equivalent model of the energy input end, and the equivalent model of the load output end are coupled and analyzed to establish the state space equation of the system. Through coupling, the input voltage fluctuation, the dynamic behavior of the converter, and the power demand of the load are comprehensively considered to fully reflect the operating state of the entire DC power supply system. The state space equation of the system is expressed as:

[0054] ;

[0055] in, is the system state vector, including state variables such as input voltage, output voltage, and current; is the control input, including switching frequency and duty cycle; and are the state matrix and input matrix of the system, which respectively describe the relationship between the system state and the control input. Through this state space equation, the dynamic changes of the system under different operating conditions are described. Based on the system dynamic model, the loss characteristics of multiple components are analyzed to obtain the total loss distribution data of the system. The loss analysis of the DC converter mainly includes the conduction loss and switching loss of the switching device. The conduction loss is proportional to the current size and the conduction time, while the switching loss depends on the switching frequency and the current change rate. Through the loss equation, the power loss of the DC converter under different working conditions is calculated. For the energy input end, the loss is mainly manifested as the voltage loss caused by the input impedance. When the input impedance is large, the fluctuation of the input voltage will cause greater power loss and affect the energy efficiency of the system. By analyzing the changes in voltage and current at the input end, the loss distribution of the energy input end is obtained. The loss at the load output end is related to the impedance characteristics and power demand of the load. When the load changes greatly, the system will experience instantaneous high power consumption, resulting in large losses. Therefore, through the loss analysis of the load end, the operating state of the system is optimized and unnecessary energy waste is reduced.

[0056] In one example, based on the system dynamic model, the loss characteristics of multiple components are analyzed to obtain the total system loss distribution data, including:

[0057] Based on the dynamic model of the DC converter, the conduction loss and switching loss of the switching device are analyzed, and the power loss equation including the switching frequency and duty cycle is established to obtain the DC converter loss model;

[0058] According to the energy input equivalent model, the influence of input voltage fluctuation on system efficiency is analyzed, the input loss function is established, and the energy input loss model is obtained;

[0059] Based on the equivalent model of the load output end, the influence of load change on system efficiency is analyzed, the relationship function between load power and system loss is established, and the load output end loss model is obtained;

[0060] The DC converter loss model, energy input end loss model and load output end loss model are superimposed to establish the overall system loss model;

[0061] The overall system loss model is numerically calculated under different operating conditions to generate the system total loss distribution data.

[0062] In this example, the conduction loss and switching loss of the switching device are analyzed based on the dynamic model of the DC converter. The conduction loss of the switching device refers to the power loss caused by the current passing through the device when the switching device (such as MOSFET) is in the on state. It is expressed as:

[0063] ;

[0064] in, is the load current, is the on-resistance of the switching device in the on state, is the duty cycle, which is the proportion of the time the switch device is on. The conduction loss depends mainly on the internal resistance of the device and the current passing through it, so under high current conditions, the conduction loss will increase significantly. Calculate the switching loss. The switching loss mainly occurs at the moment of turning on and off the switch device. When the switch device changes from off to on, or from on to off, the voltage and current do not immediately reach a steady state, so there is a large energy loss in a short period of time. Switching loss Calculated by the following formula:

[0065] ;

[0066] in, is the input voltage, is the switching frequency, and are the on-time and off-time of the switching device, respectively. The higher the switching frequency, the greater the switching loss. Therefore, in high-frequency applications, switching loss accounts for a large proportion of the total loss of the DC converter. Through the above analysis of conduction loss and switching loss, a total loss model of the DC converter is established. Combined with parameters such as switching frequency and duty cycle, the power loss of the converter under different operating conditions is fully described. The total loss model is expressed as:

[0067] ;

[0068] This means that the total power loss of the DC converter is the superposition of conduction loss and switching loss. According to the equivalent model of the energy input end, the impact of input voltage fluctuation on system efficiency is analyzed. The voltage at the input end usually fluctuates due to various external factors (such as power supply instability or load changes), which directly affects the energy efficiency of the system. Input voltage fluctuations will lead to unstable input current, which in turn affects the energy conversion efficiency of the system. In order to quantify this impact, an input loss function is established to describe it. Assuming the input voltage fluctuates is time-varying, the power loss at the input It is expressed as:

[0069] ;

[0070] in, is the input impedance. The greater the input voltage fluctuation, the greater the loss at the input end. Therefore, by measuring the input voltage fluctuation and determining the input impedance, the loss at the input end is calculated. This model can effectively reflect the impact of the energy input end on the overall efficiency of the system under different input conditions. For the loss at the load output end, a mathematical model is established to describe the impact of load changes on system efficiency. The loss at the load output end is mainly related to the impedance and power demand of the load. When the load changes greatly, the output voltage and current are dynamically adjusted to meet the power demand of the load. Power loss at the load end It is expressed as:

[0071] ;

[0072] in, is the output voltage, is the load impedance. When the load impedance is small or the load power demand is large, the output current of the system increases, resulting in increased losses at the load end. Therefore, changes in the load directly affect the overall power consumption of the system, especially when the load is unstable, the losses will be more significant. The loss models of the DC converter, energy input end, and load output end are superimposed to construct the overall loss model of the system. The total loss of the system It is expressed as:

[0073] ;

[0074] This formula shows that the total loss of the system is the sum of the losses of the DC converter, input end, and load output end. This model can fully reflect the power loss distribution of the entire DC power supply system under different operating conditions. Based on the overall loss model, numerical calculations are performed on the system under different operating conditions to generate the total loss distribution data of the system.

[0075] In one example, based on the total system loss distribution data, the impact of operating parameters on the loss of the DC power system is analyzed, and the optimal operating parameter combination is calculated, including:

[0076] Perform parameter sensitivity analysis on the overall system loss model, determine the key operating parameter set including switching frequency, duty cycle and output voltage, and construct an optimization objective function based on the key operating parameter set;

[0077] According to the optimization objective function, the initial parameters of the particle swarm optimization algorithm are set, including the number of particles, the number of iterations, the inertia weight and the learning factor, to obtain the initialized particle swarm;

[0078] Substitute the initialized particle swarm into the overall loss model of the system, calculate the fitness value of each particle under different load conditions, and obtain the initial optimal solution;

[0079] Based on the initial optimal solution, the speed and position of the particle swarm are updated, and iterative optimization is performed to obtain the local optimal solution and the global optimal solution for each iteration;

[0080] The convergence analysis is performed on the global optimal solution of each iteration to determine whether the termination conditions are met, and the iteration termination judgment result is obtained. Based on the iteration termination judgment result, the iterative optimization is continued or the optimal operating parameter combination is output.

[0081] In this example, a parameter sensitivity analysis is performed on the overall system loss model to determine the key operating parameter set, including switching frequency, duty cycle, and output voltage. Through sensitivity analysis, the impact of these parameters on system losses is evaluated and the most influential parameters are identified. After determining the key parameters, an optimization objective function is constructed based on these parameters, which aims to minimize the total loss of the system and provide a basis for the subsequent optimization process. For example, assuming that the total loss of the system is determined by the switching loss , conduction loss and load losses These loss functions are dependent on the switching frequency. , Duty Cycle and output voltage The total loss model can be expressed as:

[0082] ;

[0083] Through sensitivity analysis, the influence of different parameters on total loss is obtained, and the optimization objective function is determined. The optimization objective function is set as:

[0084] ;

[0085] The objective function is to find the total loss of the system The minimum parameter combination. According to the optimization objective function, the initial parameters of the particle swarm optimization algorithm are set. The particle swarm optimization algorithm finds the global optimal solution through the collaboration and update of multiple particles. The initial parameters include the number of particles, the number of iterations, the inertia weight and the learning factor. Number of particles Determines the number of candidate solutions in the algorithm, usually selected based on the complexity of the problem. Controls the influence of the particle's current velocity. A higher inertia weight allows the particle to explore a wider solution space. Learning factor and Determines the speed adjustment of the particle towards its historical best position and the global best position. For example, set the number of particles to 30, the number of iterations to 100, and the inertia weight to , learning factor and After initializing the particle swarm, the particles are substituted into the overall loss model of the system to calculate the fitness value of each particle. The fitness value is a measure of the performance of each candidate solution (particle) under the current parameter combination. The fitness value is equivalent to the total loss of the system. For each particle, its position is given by the parameter , and After substituting the loss model into the particle swarm, the fitness value of each particle is obtained. By calculating the fitness value, an initial optimal solution is selected, that is, the solution with the smallest total loss in the current particle swarm. Based on the initial optimal solution, the particle swarm optimization algorithm updates the speed and position of each particle to gradually approach the global optimal solution. The particle speed update formula is:

[0086] ;

[0087] in, Indicates The particle in The speed at the iteration, is the inertia weight, and are two random numbers used to increase randomness. It is The best position in the history of a particle, is the global best position, It is The particle in After updating the particle's position, the same method is used to calculate the fitness value of each particle, and the local optimal solution and the global optimal solution in this iteration are obtained. As the algorithm continues to iterate, the particles gradually move closer to the global optimal solution. After each iteration, a convergence analysis is performed on the current global optimal solution to determine whether the preset termination condition has been met. Common termination conditions include reaching the maximum number of iterations or the change in the global optimal solution is less than a preset threshold. For example, if the change in the global optimal solution is less than 0.0000 in several consecutive iterations, the global optimal solution is less than 0.0000. , the algorithm is judged to have converged, the iteration is terminated, and the optimal solution is output. If the termination condition is not met, the speed and position of the particle swarm are updated until the condition is met. After determining the final global optimal solution, the optimal operating parameter combination is output, that is, the switching frequency , Duty Cycle and output voltage The optimal value combination of these parameters can reduce the total loss of the system Reached minimum value.

[0088] In one example, according to the optimal operating parameter combination, a first power supply control strategy combining feedforward compensation and output voltage feedback is created, and an initial control quantity is generated, including:

[0089] Based on the optimal operating parameter combination and the system dynamic model, the system control equation including the switching frequency and duty cycle is constructed to obtain the system control model;

[0090] Linearize the system control model, design a proportional-integral controller as the output voltage feedback controller, calculate the proportional-integral parameters, and obtain the feedback control component;

[0091] Based on the overall system loss model and the optimal operating parameter combination, a feedforward compensator is constructed to calculate the impact of load disturbance on the output voltage and obtain the feedforward control component.

[0092] The feedback control component and the feedforward control component are weightedly superimposed to obtain the composite control quantity, and the stability of the composite control quantity is analyzed, the closed-loop poles are calculated, and the system stability judgment result is obtained;

[0093] According to the result of system stability judgment, the composite control quantity is adjusted and optimized to obtain the corrected composite control quantity;

[0094] Substitute the corrected composite control quantity into the system control equation to generate the initial control quantity, which is used as the output of the first power supply control strategy.

[0095] In this example, the system control equations including switching frequency and duty cycle are constructed based on the optimal operating parameter combination and the system dynamic model. The dynamic model of the DC power supply system describes the change of the system voltage and current over time, especially the effect of the switching frequency and duty cycle on the output voltage. By introducing these parameters into the system control equations, a control model that reflects the dynamic characteristics of the system is obtained. Assuming the switching frequency is , the duty cycle is , the output voltage is , establish the following system control equation:

[0096] ;

[0097] in, is the input voltage, is the load current, is the output impedance. This equation describes the output voltage As input voltage , Duty Cycle and load current The function of switching frequency and duty cycle is also clarified to determine the influence of switching frequency and duty cycle on system output. In order to simplify control design, the system control model is linearized and the nonlinear system is approximated as a local linear system so that the classical control algorithm can be used for analysis and design. Linearization is usually performed near a certain operating point. Assume that the system is and The control equation is linearized by small signal to obtain the linearized model:

[0098] ;

[0099] in, , and Represent the deviation of output voltage, duty cycle and load current respectively. The linearized model provides the dynamic response of the system under small disturbances and is used to design the controller. In order to control the output voltage of the system, a proportional integral (PI) controller is designed as the feedback controller of the output voltage. The function of the PI controller is to adjust the output voltage so that it follows the reference voltage. The PI controller uses two parameters: proportional gain and integral gain To adjust the system response. Proportional control is used to adjust the control amount immediately according to the current error, while integral control gradually corrects the system deviation according to the accumulated error. The control law of the PI controller is expressed as:

[0100] ;

[0101] in, is the reference voltage With output voltage The error between is the control output, i.e. duty cycle By adjusting and , and obtain the ideal feedback control component. On the basis of feedback control, in order to improve the system's response performance to load disturbances, a feedforward compensator is constructed. The output voltage change is predicted based on the system's prior knowledge, and compensation is performed in advance, thereby reducing the impact of load disturbances on the output voltage. According to the overall loss model of the system and the optimal operating parameter combination, the feedforward compensator adjusts the output voltage according to the load change. Directly adjust the output voltage. The feedforward control component is expressed as:

[0102] ;

[0103] This equation shows that the load current changes through the load impedance Affect the output voltage. Through feedforward compensation, the control amount is adjusted in advance when the load changes to reduce the disturbance to the output voltage. The feedback control component and the feedforward control component are weighted and superimposed to form a composite control amount. The calculation method of the composite control amount is expressed as:

[0104] ;

[0105] in, is the feedback control quantity, is the feedforward control quantity, is the weighting coefficient, which is used to adjust the weight between the two. , balance the feedback response and feedforward compensation capabilities of the system to achieve the best control effect. In order to ensure the stability of the system under the action of the composite control quantity, the composite control quantity is subjected to stability analysis and the closed-loop poles of the system are calculated. By solving the closed-loop transfer function of the system and calculating its characteristic roots (i.e., closed-loop poles), the stability of the system is determined. If all poles are located in the left half of the complex plane, the system is stable. Otherwise, the control parameters need to be further adjusted. The stability analysis is completed by solving the following closed-loop characteristic equation:

[0106] ;

[0107] in, is the state matrix of the system, is the input matrix, is the feedback gain matrix, is a complex frequency variable. By calculating the root of the characteristic equation, the stability of the system is determined. According to the results of the stability analysis, if the system is unstable or the response is not ideal, the composite control quantity needs to be adjusted and optimized. The optimization is done by re-adjusting value, or re-adjust the PI controller and Parameters are set to ensure that the system can operate stably under different load conditions. After adjustment, the corrected composite control quantity is obtained. Substitute the corrected composite control quantity into the system control equation to generate the initial control quantity. This control quantity will be used as the output of the first power supply control strategy to drive the system to adjust the duty cycle. and switching frequency , thereby achieving precise control of the output voltage.

[0108] In one example, the real-time operation data of the DC power system is input into a multi-branch graph convolutional neural network for feature extraction, domain-invariant features are constructed, and the initial control quantity is updated through an adaptive learning algorithm to obtain an intelligent power control model, including:

[0109] Perform multi-scale wavelet transform on the real-time operation data of the DC power supply system to obtain the time-frequency domain feature matrix;

[0110] Construct a multi-branch graph convolutional neural network, including spatial branch, frequency branch and temporal branch. Each branch contains a graph attention layer, a dynamic graph convolution layer and a skip connection.

[0111] The time-frequency domain feature matrix is ​​input into the spatial branch, and the node features are weighted by the spatial graph attention mechanism to obtain the spatial attention features; the spatial graph attention mechanism adopts a multi-head attention structure with 8 heads and a dimension of 64 for each head;

[0112] Perform dynamic graph convolution on the spatial attention features, dynamically update the adjacency matrix of the graph, and obtain adaptive spatial features; the dynamic graph convolution layer contains 3 layers, each of which uses a gated linear unit as the activation function;

[0113] The time-frequency domain feature matrix is ​​input into the frequency branch, and the frequency components are sorted by importance through the frequency graph attention mechanism to obtain the frequency attention feature. The frequency attention feature is then subjected to wavelet graph convolution operation, and the wavelet coefficients of different scales are combined to obtain the multi-resolution frequency feature.

[0114] The time-frequency domain feature matrix is ​​input into the timing branch, and the long-term and short-term dependencies are captured through the timing graph attention mechanism to obtain the timing attention features. The timing graph attention mechanism adopts the Transformer structure, which includes 4 encoder layers, and each layer has 8 attention heads.

[0115] Perform causal graph convolution on the temporal attention features to maintain the cause-and-effect relationship of the temporal information and obtain temporal dependency features; the causal graph convolution uses dilated convolution with a dilation rate of [1, 2, 4, 8];

[0116] The adaptive spatial features, multi-resolution frequency features and temporal dependency features are fused across branches to obtain a comprehensive feature representation, and the comprehensive feature representation is adaptively instance normalized to obtain domain-invariant features.

[0117] Based on the domain invariant features, an adaptive learning algorithm is designed, and the initial control quantity is updated to finally obtain an intelligent power supply control model.

[0118] In this example, multiscale wavelet transform is performed on the real-time operation data of the DC power supply system to extract time-frequency domain features. Multiscale wavelet transform is a signal processing technology that can decompose complex time series signals into sub-signals of different frequencies and time scales to reveal the local and global characteristics of the signal. Through multiscale wavelet transform, short-term transient and long-term trend changes in the operation of the DC power supply system are captured. Assume that the real-time signal of the DC power supply system is , through wavelet transform, we can get wavelet coefficients of different scales ,in Indicates scale, Indicates the position. These wavelet coefficients constitute the time-frequency domain feature matrix , where each row represents the feature of a time segment and each column represents the feature of the corresponding frequency. Based on the time-frequency domain feature matrix , build a multi-branch graph convolutional neural network to capture complex dependencies in space, frequency and time. The multi-branch network consists of three core branches: space branch, frequency branch and time branch. Each branch contains a graph attention layer, a dynamic graph convolution layer and a jump connection, processing the time and frequency domain features from different angles and finally fusing them. The time and frequency domain feature matrix is input to the spatial branch, which aims to capture the spatial association between features through the spatial graph attention mechanism. The core of the spatial graph attention mechanism is to weight the node features and capture global information by focusing on the importance of different nodes. A multi-head attention structure is used here, and the attention calculation of each node is completed in parallel through multiple attention heads. The number of heads is set to 8, and the dimension of each head is 64. Through these multi-head attentions, the system can more comprehensively analyze the complex relationship between features and finally generate a spatial attention feature matrix . Perform dynamic graph convolution on the spatial attention features. Dynamic graph convolution is a convolution method that dynamically adjusts the graph structure. It adaptively updates the adjacency matrix of the graph according to changes in data to improve the network's expressiveness. The dynamic graph convolution layer contains three layers. Each layer uses a gated linear unit (GLU) as an activation function. GLU can selectively control which information needs to pass through the gate mechanism. After three layers of dynamic graph convolution processing, an adaptive spatial feature matrix is ​​obtained, which can reflect the dynamic changes of the system in different spatial dimensions. At the same time, the time-frequency domain feature matrix is ​​input into the frequency branch. The frequency branch uses the frequency graph attention mechanism to sort the importance of different frequency components and extract important features in frequency. The frequency branch uses the graph attention mechanism to weight the features in the frequency dimension to generate frequency attention features. . Perform wavelet graph convolution on the frequency features. Wavelet graph convolution combines wavelet coefficients of different scales to effectively extract multi-resolution features on the frequency and generate a multi-resolution frequency feature matrix. Input the time-frequency domain feature matrix into the time series branch, and capture the long-term and short-term dependencies in the signal through the time series graph attention mechanism. The time series graph attention mechanism adopts the Transformer structure. The Transformer can efficiently capture the long-range dependencies in time series data through the self-attention mechanism. The time series graph attention contains four encoder layers, each with 8 attention heads. Through this structure, the time series characteristics of the system are understood from the time dimension, and the time series attention feature matrix is ​​generated. Perform causal graph convolution on the time series attention features. Causal graph convolution is a convolution method that maintains the cause-and-effect relationship of the time series, ensuring that the order of the time series is not destroyed. In this convolution operation, a dilated convolution is used with dilated rates of [1, 2, 4, 8] to capture time information of different scales. The resulting time series dependency feature matrix can fully reflect the time dependency characteristics of the system. The adaptive spatial features, multi-resolution frequency features and temporal dependency features are fused across branches to obtain a comprehensive feature representation. The comprehensive feature representation is adaptively normalized. Adaptive instance normalization dynamically calculates normalization parameters, including scaling factors. and offset , so that the normalized features have domain invariance. This process can eliminate the influence of different scenarios and operating conditions on feature distribution and generate a domain invariant feature matrix. Based on the domain invariant features, an adaptive learning algorithm is designed to optimize the control strategy of the system. The adaptive learning algorithm continuously adjusts the initial control amount of the system. , so that the system can maintain the best state under different operating conditions. Initial control amount is the input signal of the system, such as duty cycle or switching frequency. The adaptive learning algorithm dynamically adjusts the , so that it gradually approaches the optimal solution. Through adaptive learning, the initial control quantity of the system will be continuously updated until the best control strategy is obtained.

[0119] In one example, adaptive spatial features, multi-resolution frequency features, and temporal dependency features are cross-branch feature fusion to obtain a comprehensive feature representation, and the comprehensive feature representation is adaptively instance normalized to obtain domain invariant features, including:

[0120] Align the feature dimensions of adaptive spatial features, multi-resolution frequency features, and time-dependent features, and map each feature to the feature space of the same dimension through a fully connected layer to obtain unified dimensional features;

[0121] Based on the unified dimensional features, an attention-guided feature aggregation module is constructed to calculate the importance weight of each branch feature and obtain the weighted feature representation;

[0122] Perform nonlinear feature transformation and dimensionality reduction on the weighted feature representation to obtain preliminary fusion features, and perform residual connection and layer normalization on the preliminary fusion features to obtain comprehensive feature representation;

[0123] Based on the comprehensive feature representation, an adaptive instance normalization layer is constructed to dynamically calculate the normalization parameters, including the scaling factor and offset, to obtain the normalization parameter set;

[0124] The comprehensive feature representation and normalization parameter set are input into the adaptive instance normalization layer to perform feature normalization operation and obtain domain invariant features.

[0125] In this example, the adaptive spatial features, multi-resolution frequency features, and time-dependent features are aligned in feature dimension. The features of each branch are mapped to the same feature space through a fully connected layer (i.e., a fully connected neural network layer). Assume that the dimension of the adaptive spatial feature is The dimension of the multi-resolution frequency feature is , the dimension of the time-dependent feature is , map them to a unified dimension through the fully connected layer The process is expressed as:

[0126] ;

[0127] ;

[0128] ;

[0129] in, , , is the weight matrix of the fully connected layer of each branch, , , is the corresponding bias vector. Through the mapping operation, we get the features of uniform dimension. , and Based on the features of unified dimensions, an attention-guided feature aggregation module is constructed. Different importance weights are assigned to the features of each branch to reflect their contribution to the final output. For different feature branches, they are weighted by calculating the attention weights. is the set of all branch features, then the weight of each branch feature is defined for:

[0130] ;

[0131] in, It is The attention mechanism calculates the normalized weight of each branch through the softmax operation. Through the attention-guided aggregation method, the weighted feature representation is obtained.

[0132] ;

[0133] The weighted feature representation combines the feature information of different branches and weights them according to their importance. It is the basis for subsequent feature fusion. Nonlinear feature transformation and dimensionality reduction operations are performed on the weighted feature representation to extract deep feature patterns and reduce redundant information. Nonlinear transformation usually uses activation functions such as ReLU (rectified linear unit) or GELU (Gaussian error linear unit) to enhance the expression ability of the network. Assume that the activation function is , then the nonlinear feature transformation is expressed as:

[0134] ;

[0135] in, and are the weights and biases of the nonlinear layer. Dimensionality reduction is performed to reduce redundant information and compress feature dimensions to obtain preliminary fused features. In order to enhance the robustness of the model, residual connections and layer normalization operations are performed on the preliminary fused features. The purpose of residual connection is to add the original features to the transformed features to ensure that information is not lost and to facilitate the transfer of gradients. The residual connection is expressed as:

[0136] ;

[0137] Layer normalization is to make the features stable under different training batches and data distributions. Layer normalization makes the features smoother and more stable by standardizing the mean and variance of each feature. , construct an adaptive instance normalization layer to dynamically normalize features. Adaptive instance normalization dynamically calculates normalization parameters based on different inputs, including scaling factors and offset These parameters are learned through a small neural network and can provide personalized normalization schemes for the features of different instances. Assume that the comprehensive feature is represented as , the normalized feature is expressed as:

[0138] ;

[0139] in, and represent the mean and standard deviation of the features respectively, and is the scaling factor and offset learned by the network. Adaptive instance normalization can eliminate the differences in feature distribution and ensure the robustness of features in different scenarios and operating conditions. and the normalized parameter set Input the adaptive instance normalization layer to perform feature normalization to obtain domain-invariant features. Domain-invariant features eliminate the influence of different data domains, allowing the model to maintain stable output under different operating conditions and environments.

[0140] In one example, based on the prediction results of the first power control strategy and the intelligent power control model, the state space of the DC power system is analyzed, the state duration and the transition condition are determined, and the state transition parameters are dynamically adjusted to obtain the second power control strategy, including:

[0141] Comparing the first power supply control strategy with the prediction result of the intelligent power supply control model, calculating the control deviation, and obtaining a control error sequence;

[0142] The mean and standard deviation of the control error sequence are calculated to obtain the system steady-state error index. Based on the system steady-state error index, the state transfer threshold is set to divide the system operation state and obtain the state space division result.

[0143] Analyze the system response characteristics under each divided state, calculate the state duration and transition frequency, and obtain the state transition characteristics;

[0144] According to the state transition characteristics, an adaptive state holding time algorithm is executed to dynamically adjust the state switching conditions to obtain the optimized state transition rules;

[0145] Based on the optimized state transition rules, a state switching controller is constructed to perform smooth transitions between different operating states and obtain a multi-mode control strategy.

[0146] The multi-mode control strategy is integrated with the first power supply control strategy to obtain a comprehensive control output, and a closed-loop correction is performed on the comprehensive control output. The control parameters are adjusted according to real-time feedback to obtain a second power supply control strategy.

[0147] In this example, the first power control strategy is compared with the prediction results of the intelligent power control model to calculate the control deviation , which is expressed as:

[0148] ;

[0149] in, is the output voltage predicted by the intelligent power control model, is the actual output voltage under the first power supply control strategy. Error sequence It changes with time and reflects the difference between the model prediction and the actual control strategy. , calculate the mean and standard deviation of the series. Mean It represents the average error of the system over a long period of time, reflecting the deviation trend of the control, while the standard deviation It describes the fluctuation range of the error and reflects the stability of the system. Based on the steady-state error index of the system, the threshold of state transition is set. By setting appropriate state transition thresholds, the system operation is divided into different state spaces. In order to determine these thresholds, the maximum inter-class variance method statistical method is used to effectively distinguish different states according to the distribution of errors. Assuming that the system operation state is divided into "normal state", "transition state" and "abnormal state", the corresponding threshold is set according to the size of the error. For example, when the error is small, the system is in a normal state; when the error increases to a certain extent, the system enters a transition state; when the error is too large, the system is in an abnormal state. In this way, the state space division result of the system is obtained. After the state division is completed, the system response characteristics under each divided state are analyzed. The duration of each state and the transition frequency between states are calculated. The state duration refers to the time the system maintains in a certain state, and the transition frequency indicates how frequently the system transfers from one state to another. Through statistical analysis, the operating characteristics of the system in different states are obtained. For example, if the system often switches frequently between the normal state and the transition state, it may indicate that the system control strategy is not stable enough and needs to be adjusted. Based on the state transition characteristics, the adaptive state holding time algorithm is executed. According to the duration and frequency of system state transition, the state switching conditions are dynamically adjusted to optimize the stability of the system. The adaptive state retention time algorithm sets a minimum state retention time, that is, the shortest duration of the system in any state cannot be less than T, so as to avoid instability caused by frequent system switching. Through the optimized state transition rules, the frequency of system jumps between different states is effectively reduced, and the operation stability of the system is improved. Based on the optimized state transition rules, a state switching controller is constructed. The function of the controller is to dynamically adjust the control strategy according to the system behavior in different states, so that the system can smoothly transition between different operating states. The state switching controller adjusts the control output according to the threshold of state transition and the real-time state of the system to ensure the stability during the state switching process. For example, when the system transfers from the normal state to the transition state, the controller appropriately adjusts the switching frequency or duty cycle according to the characteristics of the transition state to ensure the smooth change of the output voltage and reduce the voltage fluctuation during the transition process. The state switching controller is integrated with the first power supply control strategy to obtain a multi-mode control strategy. The multi-mode control strategy can more flexibly cope with complex system environments by combining the optimal control methods under different states. For example, in normal state, the system uses basic PI control strategy, while in transition state or abnormal state, more sophisticated control methods such as model predictive control or adaptive control are used. In this way, the multi-mode control strategy can better respond to the needs of the system under different operating conditions. The multi-mode control strategy is integrated with the first power supply control strategy to obtain a comprehensive control output.The comprehensive control output combines the control requirements under different states and can maintain the stable operation of the system under complex operating conditions. At the same time, in order to ensure the effectiveness of the comprehensive control strategy, a closed-loop correction mechanism is adopted to continuously adjust the control parameters according to real-time feedback. The closed-loop correction dynamically adjusts the controller parameters, such as proportional gain or integral gain, by monitoring the output voltage and control error of the system in real time to ensure that the control system can maintain stable and efficient operation under different operating conditions. Through the closed-loop correction process, the system can adaptively adjust the control parameters to obtain the final second power supply control strategy. This strategy can optimize the control output of the system under different operating conditions, and can also adjust the control strategy through the real-time feedback mechanism to achieve the optimization of dynamic response.

[0150] In one example, the mean and standard deviation of the control error sequence are calculated to obtain a system steady-state error index, and based on the system steady-state error index, a state transition threshold is set to divide the system operation state, and a state space division result is obtained, including:

[0151] The control error sequence is processed by sliding window to obtain multiple subsequences, and the mean and standard deviation of each subsequence are calculated to obtain the system steady-state error index;

[0152] The system steady-state error index is input into the kernel density estimation algorithm, the probability density function of the error distribution is calculated, and the maximum inter-class variance method is used to determine the initial threshold to obtain a set of candidate thresholds;

[0153] Each threshold in the candidate threshold set is evaluated, the ratio of the intra-class variance to the inter-class variance is calculated, and the threshold evaluation index is obtained. According to the threshold evaluation index, the optimal threshold is selected as the state transition threshold to obtain the final threshold;

[0154] Based on the final threshold, the control error sequence is segmented into different operating states to obtain a preliminary state division result. The adjacent states of the preliminary state division result are merged, and the minimum state duration is set to obtain a state space division result.

[0155] In this example, the control error sequence is processed by sliding window, the original error sequence is divided into multiple subsequences, and statistical analysis is performed on each subsequence to more accurately capture the error changes in different time periods during system operation. Assume that the control error sequence is , the length is , the size of the sliding window is , the sliding step length is , define the subsequence of each window as:

[0156] ;

[0157] Through sliding window processing, we get subsequences, each of which contains For each subsequence , calculate the mean and standard deviation

[0158] ;

[0159] The mean and standard deviation reflect the steady-state error of the system in each sub-time period, and the steady-state error index of the system is obtained. The steady-state error index of the system is input into the kernel density estimation algorithm to calculate the probability density function of the error distribution. The kernel density estimation algorithm is a non-parametric estimation method that intuitively represents the probability distribution of data by smoothing the distribution around each data point. For the sequence of mean and standard deviation , through the kernel function Calculate the probability density function

[0160] ;

[0161] in, is the sample size, is the smoothing parameter (bandwidth), Gaussian kernel is usually used. By calculating the kernel density function, the probability distribution of the error is obtained. In order to determine the initial threshold of state transition, the maximum inter-class variance method is used to maximize the inter-class variance of different error intervals to determine an optimal threshold so that the errors are divided into different categories. Represents the difference between different categories, intra-class variance Indicates the degree of dispersion within the same category. Assume that the errors are divided into two categories and , and the corresponding between-class variance is:

[0162] ;

[0163] in, and are the weights of the two categories, and are the means of the two categories, is the overall mean. By maximizing the inter-class variance, the initial candidate threshold set is obtained. Each threshold in the candidate threshold set is evaluated, and the ratio of the intra-class variance to the inter-class variance is calculated to obtain the threshold evaluation index. Intra-class variance The formula is:

[0164] ;

[0165] in, and are the variances of the two classes respectively. Calculate the ratio of the between-class variance to the within-class variance:

[0166] ;

[0167] The larger the ratio is, the better the threshold can distinguish the error characteristics of different categories. The threshold with the largest ratio is selected as the optimal threshold to obtain the final state transition threshold. Based on the final determined state transition threshold, the original control error sequence is segmented to distinguish the system error into different operating states. For example, when the error value is less than a certain threshold, the system is in a "normal state"; when the error value is in a medium range, the system enters a "transition state"; when the error value exceeds a larger threshold, the system is considered to be in an "abnormal state". In this way, the preliminary state division results of the system are obtained. In order to ensure the stability and continuity of the system's operating state, the adjacent states of the preliminary state division results are merged. Set a minimum state duration , ensuring that the system's running time in any state does not fall below the threshold. If the duration of a state is less than , then merge it into the previous or next adjacent state to avoid frequent state switching. By merging adjacent states and setting the minimum state duration, the state space division result of the system is obtained. This result can stably reflect the state changes of the system under different operating conditions and provide a basis for subsequent control strategy adjustments.

[0168] Reference Figure 2 This embodiment provides a DC power supply operation state intelligent control device, including:

[0169] Modeling module 1 is used to model multiple components of the DC power supply system to obtain a system dynamic model, and based on the system dynamic model, analyze the loss characteristics of the multiple components to obtain system total loss distribution data;

[0170] Calculation module 2, used for analyzing the influence of operating parameters on the loss of the DC power supply system based on the total loss distribution data of the system, and calculating the optimal combination of operating parameters;

[0171] Creating module 3, used for creating a first power supply control strategy combining feedforward compensation and output voltage feedback according to the optimal operating parameter combination, and generating an initial control amount;

[0172] Extraction module 4, used for inputting the real-time operation data of the DC power supply system into the multi-branch graph convolutional neural network for feature extraction, constructing domain invariant features, updating the initial control quantity through an adaptive learning algorithm, and obtaining an intelligent power supply control model;

[0173] The adjustment module 5 is used to analyze the state space of the DC power supply system based on the prediction results of the first power supply control strategy and the intelligent power supply control model, determine the state duration and transfer conditions, dynamically adjust the state transfer parameters, and obtain the second power supply control strategy.

[0174] 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.

[0175] 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.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] 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.

[0180] 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 method for intelligently controlling the operation status of a DC power supply, characterized in that: The following steps are involved: Modeling multiple components of a DC power supply system to obtain a system dynamic model, and based on the system dynamic model, performing loss characteristic analysis on the multiple components to obtain system total loss distribution data; Based on the total loss distribution data of the system, analyzing the impact of operating parameters on the loss of the DC power supply system, and calculating the optimal operating parameter combination; According to the optimal operating parameter combination, a first power supply control strategy combining feedforward compensation and output voltage feedback is created, and an initial control quantity is generated; specifically, the following steps are performed: based on the optimal operating parameter combination and the system dynamic model, a system control equation including a switching frequency and a duty cycle is constructed to obtain a system control model; the system control model is linearized, and a proportional-integral controller is designed as an output voltage feedback controller, and a proportional-integral parameter is calculated to obtain a feedback control component; based on the overall system loss model and the optimal operating parameter combination, a feedforward compensator is constructed to calculate the influence of load disturbance on the output voltage to obtain a feedforward control component; the feedback control component is weightedly superimposed with the feedforward control component to obtain a composite control quantity, and a stability analysis is performed on the composite control quantity, and the closed-loop poles are calculated to obtain a system stability judgment result; according to the system stability judgment result, the composite control quantity is parameter adjusted and optimized to obtain a modified composite control quantity; the modified composite control quantity is substituted into the system control equation to generate an initial control quantity, and used as the output of the first power supply control strategy; Inputting the real-time operation data of the DC power supply system into a multi-branch graph convolutional neural network for feature extraction, constructing domain-invariant features, updating the initial control quantity through an adaptive learning algorithm, and obtaining an intelligent power supply control model; Based on the prediction results of the first power control strategy and the intelligent power control model, the state space of the DC power supply system is analyzed, the state duration and transfer conditions are determined, and the state transfer parameters are dynamically adjusted to obtain the second power control strategy.

2. The method for intelligently controlling the operation status of a DC power supply according to claim 1, characterized in that: The multiple components of the DC power supply system are modeled to obtain a system dynamic model, and based on the system dynamic model, the multiple components are analyzed for loss characteristics to obtain system total loss distribution data, including: The multiple components of the DC power supply system include a DC converter, an energy input terminal and a load output terminal; Performing circuit topology analysis on the DC converter, establishing a state equation including the on-state and off-state of the switch device, and obtaining a dynamic model of the DC converter; Analyzing the voltage and current characteristics of the energy input terminal, establishing a mathematical expression that takes into account input impedance and input voltage fluctuations, and obtaining an equivalent model of the energy input terminal; Performing load characteristic analysis on the load output terminal, establishing a mathematical description including load impedance and power demand, and obtaining an equivalent model of the load output terminal; Conduct coupling analysis on the dynamic model of the DC converter, the energy input end equivalent model and the load output end equivalent model, establish a system state space equation, and obtain a system dynamic model; Based on the system dynamic model, loss characteristics of the multiple components are analyzed to obtain system total loss distribution data.

3. The intelligent control method for the operation status of a DC power supply according to claim 2, characterized in that: The loss characteristics of the plurality of components are analyzed based on the system dynamic model to obtain total system loss distribution data, including: Based on the DC converter dynamic model, the conduction loss and switching loss of the switch device are analyzed, a power loss equation including the switching frequency and the duty cycle is established, and a DC converter loss model is obtained; According to the energy input end equivalent model, the influence of input voltage fluctuation on system efficiency is analyzed, an input end loss function is established, and an energy input end loss model is obtained; Based on the load output end equivalent model, the influence of load change on system efficiency is analyzed, a relationship function between load power and system loss is established, and a load output end loss model is obtained; The DC converter loss model, the energy input end loss model and the load output end loss model are superimposed to establish a system overall loss model; The overall loss model of the system is numerically calculated under different operating conditions to generate system total loss distribution data.

4. The method for intelligently controlling the operation status of a DC power supply according to claim 3, characterized in that: The analyzing the influence of operating parameters on the loss of the DC power supply system based on the total loss distribution data of the system and calculating the optimal operating parameter combination includes: Performing parameter sensitivity analysis on the overall loss model of the system, determining a set of key operating parameters including switching frequency, duty cycle and output voltage, and constructing an optimization objective function based on the set of key operating parameters; According to the optimization objective function, initial parameters of the particle swarm optimization algorithm are set, including the number of particles, the number of iterations, the inertia weight and the learning factor, to obtain an initialized particle swarm; Substituting the initialized particle swarm into the overall loss model of the system, calculating the fitness value of each particle under different load conditions, and obtaining an initial optimal solution; Based on the initial optimal solution, the speed and position of the particle swarm are updated, and iterative optimization is performed to obtain a local optimal solution and a global optimal solution for each iteration; The global optimal solution of each iteration is subjected to convergence analysis to determine whether the termination condition is met, and an iteration termination judgment result is obtained. Based on the iteration termination judgment result, the iterative optimization is continued or the optimal operating parameter combination is output.

5. The method for intelligently controlling the operation status of a DC power supply according to claim 1, characterized in that: The real-time operation data of the DC power supply system is input into a multi-branch graph convolutional neural network for feature extraction, domain-invariant features are constructed, and the initial control quantity is updated through an adaptive learning algorithm to obtain an intelligent power supply control model, including: Performing multi-scale wavelet transform on the real-time operation data of the DC power supply system to obtain a time-frequency domain feature matrix; Construct a multi-branch graph convolutional neural network, including spatial branch, frequency branch and temporal branch. Each branch contains a graph attention layer, a dynamic graph convolution layer and a skip connection. The time-frequency domain feature matrix is ​​input into the spatial branch, and the node features are weighted by a spatial graph attention mechanism to obtain spatial attention features; the spatial graph attention mechanism adopts a multi-head attention structure with 8 heads and a dimension of 64 for each head; Performing a dynamic graph convolution operation on the spatial attention feature, dynamically updating the adjacency matrix of the graph, and obtaining an adaptive spatial feature; the dynamic graph convolution layer includes 3 layers, and each layer uses a gated linear unit as an activation function; Inputting the time-frequency domain feature matrix into the frequency branch, sorting the frequency components by importance through the frequency graph attention mechanism to obtain frequency attention features, performing wavelet graph convolution operation on the frequency attention features, and combining wavelet coefficients of different scales to obtain multi-resolution frequency features; The time-frequency domain feature matrix is ​​input into the timing branch, and the long-term and short-term dependencies are captured by the timing graph attention mechanism to obtain the timing attention feature; the timing graph attention mechanism adopts a Transformer structure, including 4 encoder layers, each layer has 8 attention heads; Performing a causal graph convolution operation on the temporal attention feature to maintain the cause-and-effect relationship of the temporal information and obtain a temporal dependency feature; the causal graph convolution uses a dilated convolution with a dilated rate of [1, 2, 4, 8]; Performing cross-branch feature fusion on the adaptive spatial features, the multi-resolution frequency features and the temporal dependency features to obtain a comprehensive feature representation, and performing adaptive instance normalization on the comprehensive feature representation to obtain a domain invariant feature; Based on the domain invariant features, an adaptive learning algorithm is designed, and the initial control quantity is updated to finally obtain an intelligent power supply control model.

6. The method for intelligently controlling the operation status of a DC power supply according to claim 5, characterized in that: The step of performing cross-branch feature fusion on the adaptive spatial features, the multi-resolution frequency features and the temporal dependency features to obtain a comprehensive feature representation, and performing adaptive instance normalization on the comprehensive feature representation to obtain a domain invariant feature includes: Aligning the feature dimensions of the adaptive spatial features, the multi-resolution frequency features, and the temporal dependency features, and mapping each feature to a feature space of the same dimension through a fully connected layer to obtain a unified dimensional feature; Based on the unified dimensional features, an attention-guided feature aggregation module is constructed to calculate the importance weight of each branch feature to obtain a weighted feature representation; Performing nonlinear feature conversion and dimensionality reduction on the weighted feature representation to obtain preliminary fusion features, and performing residual connection and layer normalization on the preliminary fusion features to obtain comprehensive feature representation; Based on the comprehensive feature representation, an adaptive instance normalization layer is constructed, and normalization parameters, including a scaling factor and an offset, are dynamically calculated to obtain a normalization parameter set; The comprehensive feature representation and the normalization parameter set are input into an adaptive instance normalization layer to perform a feature normalization operation to obtain domain invariant features.

7. The method for intelligently controlling the operation status of a DC power supply according to claim 1, characterized in that: The method of analyzing the state space of the DC power supply system based on the prediction result of the first power supply control strategy and the intelligent power supply control model, determining the state duration and the transition condition, and dynamically adjusting the state transition parameter to obtain the second power supply control strategy includes: Comparing the first power supply control strategy with the prediction result of the intelligent power supply control model, calculating the control deviation, and obtaining a control error sequence; Calculating the mean and standard deviation of the control error sequence to obtain a system steady-state error index, and setting a state transfer threshold based on the system steady-state error index to divide the system operating state and obtain a state space division result; Analyze the system response characteristics under each divided state, calculate the state duration and transition frequency, and obtain the state transition characteristics; According to the state transition characteristics, an adaptive state holding time algorithm is executed to dynamically adjust the state switching conditions to obtain an optimized state transition rule; Based on the optimized state transition rules, a state switching controller is constructed to perform smooth transitions between different operating states to obtain a multi-mode control strategy; The multi-mode control strategy is integrated with the first power supply control strategy to obtain a comprehensive control output, and a closed-loop correction is performed on the comprehensive control output, and control parameters are adjusted according to real-time feedback to obtain a second power supply control strategy.

8. The method for intelligently controlling the operation status of a DC power supply according to claim 7, characterized in that: The method of calculating the mean and standard deviation of the control error sequence to obtain a system steady-state error index, and setting a state transfer threshold based on the system steady-state error index to divide the system operation state to obtain a state space division result includes: Performing sliding window processing on the control error sequence to obtain multiple subsequences, and calculating the mean and standard deviation of each subsequence to obtain a system steady-state error index; Inputting the system steady-state error index into a kernel density estimation algorithm, calculating the probability density function of the error distribution, and using the maximum inter-class variance method to determine the initial threshold to obtain a candidate threshold set; Evaluate each threshold in the candidate threshold set, calculate the ratio of the intra-class variance to the inter-class variance, obtain a threshold evaluation index, and select the optimal threshold as the state transition threshold according to the threshold evaluation index to obtain a final threshold; The control error sequence is segmented based on the final threshold value into different operating states to obtain a preliminary state division result, and adjacent states of the preliminary state division result are merged, and a minimum state duration is set to obtain a state space division result.

9. An intelligent control device for the operation status of a DC power supply, characterized in that: For implementing the steps of the method according to any one of claims 1 to 8, the device comprises: A modeling module, used to model multiple components of a DC power supply system to obtain a system dynamic model, and based on the system dynamic model, perform loss characteristic analysis on the multiple components to obtain system total loss distribution data; A calculation module, used for analyzing the influence of operating parameters on the loss of the DC power supply system based on the total loss distribution data of the system, and calculating the optimal combination of operating parameters; A creation module, used for creating a first power supply control strategy combining feedforward compensation and output voltage feedback according to the optimal operating parameter combination, and generating an initial control amount; An extraction module, used for inputting the real-time operation data of the DC power supply system into a multi-branch graph convolutional neural network for feature extraction, constructing domain-invariant features, updating the initial control amount through an adaptive learning algorithm, and obtaining an intelligent power supply control model; An adjustment module is used to analyze the state space of the DC power supply system based on the prediction results of the first power supply control strategy and the intelligent power supply control model, determine the state duration and transfer conditions, dynamically adjust the state transfer parameters, and obtain the second power supply control strategy.

Images