Control method for emergency mobile power supply and emergency mobile power supply

Abstract

The present invention relates to the technical field of emergency mobile power supplies, and discloses a control method for an emergency mobile power supply and an emergency mobile power supply. Among them, the method includes: collecting operation parameter data of the emergency mobile power supply and performing adaptive wavelet transform denoising to obtain a standardized feature matrix; performing hierarchical feature extraction through an extended convolutional capsule network to obtain global power supply features and local power supply features; establishing a deep deterministic policy network for power supply state scoring to obtain a state scoring result; inputting historical data in the experience replay pool into an action generation module to optimize and iterate control actions to obtain a first control strategy, and recording state transition data during the control process; performing power loss compensation control analysis to obtain a compensation control amount, and performing superposition to obtain a second control strategy, and performing PWM modulation to generate charge and discharge control signals. The present invention realizes precise power regulation and energy storage unit control, and ensures the stable operation of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency mobile power supplies, and particularly to a control method for an emergency mobile power supply and an emergency mobile power supply. Background Art

[0002] As a key portable power supply device, the emergency mobile power supply plays an important role in scenarios such as emergency rescue, disaster handling, and temporary power use. However, the current emergency mobile power supply control systems generally suffer from problems such as insufficient control accuracy, slow dynamic response, and poor power supply stability. Traditional control methods mainly rely on optimizing local power supply characteristics and are difficult to effectively handle dynamic load changes in complex power supply scenarios, resulting in slow system response speed and large power supply fluctuations.

[0003] In the existing emergency mobile power supply control methods, the problem of power loss compensation has not been fully emphasized. In practical applications, line losses, switch losses, and energy storage unit losses will significantly affect the power supply efficiency and stability of the system. Especially during high-power output and frequent switching of working states, the superimposed effect of various losses will cause a sharp decline in system performance and affect the reliability of emergency power supply. In addition, the existing control strategies have insufficient perception ability of the global state of the system and are difficult to accurately capture key characteristic information during the power supply process. Summary of the Invention

[0004] The present invention provides a control method for an emergency mobile power supply and an emergency mobile power supply, which realizes precise power regulation and energy storage unit control and ensures the stable operation of the system.

[0005] In a first aspect, the present invention provides a control method for an emergency mobile power supply, and the control method for the emergency mobile power supply includes:

[0006] Collect operation parameter data of the emergency mobile power supply and perform adaptive wavelet transform denoising to obtain a standardized feature matrix;

[0007] Construct a multi-dimensional feature vector using the standardized feature matrix, and input the multi-dimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain global power supply characteristics and local power supply characteristics;

[0008] Establish a deep deterministic policy network based on the global power supply characteristics and the local power supply characteristics, and the deep deterministic policy network includes a value evaluation module and an action generation module; use the value evaluation module to perform a power supply state score to obtain a state score result, and store the state score result and the corresponding control action in an experience replay pool;

[0009] Input the historical data in the experience replay pool into the action generation module to optimize and iterate the control actions, and obtain the first control strategy; use the first control strategy for online control, and record the state transition data during the control process;

[0010] Input the state transition data into the power loss compensation model for power loss compensation control analysis to obtain the compensation control quantity, and superimpose the compensation control quantity on the first control strategy to obtain the second control strategy, and perform PWM modulation based on the second control strategy to generate the corresponding charge and discharge control signals.

[0011] In a second aspect, the present invention provides an emergency mobile power supply, and the emergency mobile power supply includes:

[0012] An acquisition module, configured to collect operation parameter data of the emergency mobile power supply and perform adaptive wavelet transform denoising to obtain a standardized feature matrix;

[0013] A feature extraction module, configured to construct a multi-dimensional feature vector by using the standardized feature matrix, and input the multi-dimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain a global power supply feature and a local power supply feature;

[0014] A building module, configured to establish a deep deterministic policy network based on the global power supply feature and the local power supply feature, where the deep deterministic policy network includes a value evaluation module and an action generation module; use the value evaluation module to perform a power supply state score to obtain a state score result, and store the state score result and the corresponding control action in the experience replay pool;

[0015] An iteration module, configured to input the historical data in the experience replay pool into the action generation module to optimize and iterate the control actions, and obtain the first control strategy; use the first control strategy for online control, and record the state transition data during the control process;

[0016] A generation module, configured to input the state transition data into the power loss compensation model for power loss compensation control analysis to obtain the compensation control quantity, and superimpose the compensation control quantity on the first control strategy to obtain the second control strategy, and perform PWM modulation based on the second control strategy to generate the corresponding charge and discharge control signals.

[0017] In the technical solution provided by the present invention, through the introduction of adaptive wavelet transform for data denoising and feature extraction, combined with maximum-minimum normalization calculation and time series analysis, a complete data preprocessing process is established, effectively improving the quality and usability of the original data. The extended convolutional capsule network is used for hierarchical feature extraction. Through multi-level feature mapping and dynamic routing mechanisms, the effective separation of global power supply features and local power supply features is achieved, enhancing the system's perception ability of the power supply state; a control architecture based on the deep deterministic policy network is designed. Through the collaborative action of the value evaluation module and the action generation module, a complete state evaluation and action generation mechanism is established, improving the accuracy and adaptability of the control strategy; an experience replay pool is constructed to store historical data, and the control strategy is optimized by using random sampling and temporal difference calculation methods. Through multiple rounds of iterative training and soft update mechanisms, the continuous optimization and stable convergence of the control strategy are ensured; a multi-dimensional power loss compensation scheme is proposed. Through the accurate modeling and compensation of line loss, switch loss, and energy storage loss, the power loss of the system is significantly reduced, and the power supply efficiency is improved; a charge and discharge control mechanism based on PWM modulation is designed. Through dead-time compensation and triangular wave carrier modulation, accurate power regulation and energy storage unit control are achieved, ensuring the stable operation of the system.

[0018] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are realized and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0019] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, detailed descriptions are as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of an embodiment of the control method of the emergency mobile power supply in the embodiment of the present invention;

[0021] Figure 2 It is a schematic diagram of an embodiment of the emergency mobile power supply in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] As used in the embodiments of the present invention, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes other steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0024] For the convenience of understanding this embodiment, first, a control method for an emergency mobile power supply disclosed in the embodiments of the present invention will be introduced in detail. As Figure 1 shown, the method includes the following steps:

[0025] 101. Collect operation parameter data of the emergency mobile power supply and perform adaptive wavelet transform noise reduction to obtain a standardized feature matrix;

[0026] It can be understood that the execution subject of the present invention can be an emergency mobile power supply, or a terminal or a server, and specific limitations are not made here. In the embodiments of the present invention, the server is taken as an example of the execution subject for illustration.

[0027] Specifically, data collection is carried out to collect key operating parameters such as the input voltage, output current, load power, and battery status of the emergency mobile power supply, which reflect the overall load situation of the power supply and the health status of the battery. The original operating parameter data is processed in segments. The continuous original data is cut and divided into several small time windows or data segments to analyze and process each feature within a more precise time range. Wavelet transform is performed on each segment of data to extract the frequency-domain features of the signal. Wavelet transform is a signal analysis tool that can perform localized analysis in both the time and frequency dimensions. By calculating the wavelet coefficients of each segment of data, a preliminary wavelet decomposition result is obtained, which reflects the contribution of different frequency components to the original data and helps identify the noise and detailed parts in the signal. The wavelet decomposition result is optimized by setting an adaptive threshold function. The adaptive threshold function dynamically adjusts the threshold according to the local characteristics of the data, thereby distinguishing between the signal and the noise. By calculating the magnitudes of the wavelet coefficients of each segment of data, a suitable threshold is determined. When the absolute value of the coefficient is lower than this threshold, it is considered to belong to the noise and is filtered out; otherwise, the important signal components are retained. In this way, the noise is effectively removed and the signal-to-noise ratio of the data is improved. The optimized threshold parameters are substituted into the wavelet inverse transform equation to reconstruct the initial wavelet decomposition result, obtaining the denoised operating parameter data. The maximum-minimum normalization calculation is performed on the denoised operating parameter data to unify the value range of each item of data into a fixed interval, usually [0,1], eliminating the differences in dimensions between different parameters, enabling the data to be analyzed on the same scale, and avoiding unfair influences on subsequent analysis caused by certain features having too large or too small numerical ranges. Through the normalization process, a set of standardized data is obtained. Based on the standardized data, the static characteristic parameters of the power supply system are extracted. The static characteristics reflect the behavior of the power supply at a certain moment or in a certain fixed state, such as the amplitude of the input voltage, the average value of the output current, and the remaining battery power. The static characteristics can provide the basic operating conditions of the power supply, such as whether there are abnormal voltage fluctuations or battery status problems. After extracting the static characteristics, the dynamic characteristics of the power supply are considered. These characteristics mainly reflect the changes of the power supply in the time series, helping to capture the volatility and operating trends of the system. For example, by calculating indicators such as voltage volatility, power change rate, and load change trend, the response of the power supply when the load changes is identified, and further understanding of whether the power supply can adapt to the load change in a timely manner. The static characteristic parameter set and the dynamic characteristic parameter set are subjected to matrix splicing operation to obtain a standardized feature matrix.

[0028] 102. Construct a multi-dimensional feature vector using the standardized feature matrix and input the multi-dimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain global power supply characteristics and local power supply characteristics;

[0029] Specifically, dimension reconstruction and feature combination operations are performed on the standardized feature matrix to fuse the static feature data and dynamic feature data in the matrix into a unified multi-dimensional feature vector. The multi-dimensional feature vector is input into the extended convolutional capsule network for hierarchical feature extraction. The feature vector enters the first convolutional layer of the network for feature mapping. In this layer, 32 3×3 convolutional kernels are used to process the input features through convolutional operations. At the same time, the ReLU activation function is used to introduce non-linearity, and batch normalization is used to standardize the convolutional results, accelerating the training process and reducing the risk of overfitting, resulting in a primary feature map. Max pooling operation is performed on the primary feature map. The role of the pooling operation is to reduce the dimension of the feature map. Through a 2×2 pooling kernel with a stride of 2, the pooling operation reduces the size of the feature map by half and retains important feature information. The downsampled feature map obtained after pooling is input into the primary capsule layer of the extended convolutional capsule network for processing. Through the tensor transformation and direction voting calculation of the primary capsule layer, more complex feature information is captured. In this process, the primary capsule layer contains 8 capsule units, and each capsule unit outputs a 16-dimensional vector, representing different aspects of the features captured by the capsule unit, effectively describing the detailed information of the power system, and obtaining the primary capsule features. The primary capsule features are input into the digital capsule layer of the extended convolutional capsule network. The digital capsule layer performs feature routing through the dynamic routing algorithm to extract more abstract features. Dynamic routing is a process of iterative optimization that weights the outputs of different capsule units according to similarity to obtain more representative features. To ensure the stability of the routing process, the number of routing iterations is set to 3 times. Through multiple iterations, the routing algorithm can form reasonable connections between capsule units, ensuring more accurate and efficient feature transmission. Global and local feature extraction are performed. In global feature extraction, global average pooling operation is performed on the feature vector obtained by routing to obtain global power supply features. The process of global average pooling is to perform weighted averaging on all features to obtain a feature representing the working state of the entire power system, reflecting the overall operation performance of the power supply, such as the overall change trends of voltage, current, and load. To further refine the state of the power supply, information is extracted from the local level. In the process of local feature extraction, local attention weighting calculation is performed on the routed feature vector. By introducing the attention mechanism, it focuses on the features that have a greater impact on the power supply performance, such as sudden changes in current at a certain moment or abnormal fluctuations in a certain power supply unit. Local features can describe the specific working states of each subsystem or power supply unit in the power supply, helping to identify potential faults or unstable factors. Through the above steps, global power supply features and local power supply features are obtained.

[0030] 103. Establish a deep deterministic policy network based on global power supply characteristics and local power supply characteristics. The deep deterministic policy network includes a value evaluation module and an action generation module; use the value evaluation module to score the power supply state, obtain the state scoring result, and store the state scoring result and the corresponding control action in the experience replay pool;

[0031] Specifically, the global power supply characteristics and local power supply characteristics are concatenated to obtain a combined feature vector, which includes the overall operating condition of the power supply system and the specific performance of each power supply unit, reflecting the operating conditions of the power supply system under different conditions. Design the structure of the deep deterministic policy network. In the value evaluation module, perform dimensionality mapping on the combined feature vector. To ensure the effective expression of the input features in the network, process them through a three-layer fully connected neural network. The fully connected layer can extract and integrate the input feature information layer by layer, mapping the high-dimensional input data to a feature space suitable for evaluating the current power supply state. Through dimensionality mapping, the network can more accurately capture the key state information of the power supply. After processing, the value evaluation module evaluates the operating state of the current power supply system according to the input feature vector and calculates the state score result, including the power supply stability score, the power balance score, and the response speed score. At the same time, input the combined feature vector into the action generation module. This module generates corresponding control actions according to the current operating state of the power supply. To achieve this goal, perform dimensionality reduction processing on the combined feature vector through a four-layer fully connected neural network to ensure that the expression of the features is more concise without losing important information. Input the dimensionality-reduced feature vector into the action generation module to generate corresponding control action parameters. The control action parameters include power adjustment instructions and charge and discharge instructions for energy storage units, ensuring that the power supply can adjust its power output and battery charge and discharge strategies according to the current demand to meet the stable operation requirements of the system. Set a reward function for the state score result to measure the effect of the control strategy. The reward function comprehensively considers multiple factors such as power supply stability, power balance, and response speed, and calculates a reward value based on these indicators. The reward value reflects the effect of the current policy execution, which can guide the deep reinforcement learning algorithm to further optimize the policy, enabling the power supply system to achieve the optimal power supply performance under different working conditions. To achieve continuous optimization of the policy, construct an experience replay pool. The role of the experience replay pool is to store historical state, action, and reward information for sampling during the training process to update the model parameters. The storage capacity of the experience replay pool is set to 10,000, and the system can store a large amount of historical experience, providing sufficient data support for the training process. Whenever the network undergoes a training session, the data in the experience replay pool is stored in chronological order, including the current state score result, control action parameters, and the state value evaluation result calculated according to the reward function. Through the accumulation of these historical data, continuously learn and improve the control strategy, gradually enhancing the power supply stability and response ability of the power supply system. Through the above steps, the deep deterministic policy network is continuously optimized according to historical experience and the current system state, enabling the power supply to more intelligently adjust its working strategy to meet the power supply requirements in emergency situations.

[0032] 104. Input the historical data in the experience replay pool into the action generation module, optimize and iterate the control actions to obtain the first control strategy; use the first control strategy for online control and record the state transition data during the control process;

[0033] Specifically, randomly sample the historical data in the experience replay pool to obtain a batch of training data. The batch of training data contains historical state, action, and reward information, representing the performance and decision-making process of the power system in different operating environments. Perform temporal difference calculation on the batch of training data, and calculate the target Q-value through the current state transition sequence. The Q-value is a criterion for evaluating the value of a state-action pair, indicating the long-term reward that can be obtained by taking a certain action in a certain state. The target Q-value is used as the optimization target of the value evaluation module to guide the subsequent learning process. By calculating the difference between the current state and the target state, the model is continuously updated, so that the policy is gradually optimized. Input the batch of training data into the action generation module, and update the network parameters of the action generation module by gradient. Use the Adam optimizer to optimize the network parameters. The Adam optimizer is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate according to the gradient of each parameter, thereby improving the efficiency and stability of training. To ensure the smooth progress of the training process, the learning rate is set to 0.001, ensuring that the amplitude of each parameter update is moderate, avoiding local optimality caused by too fast convergence and preventing the problem of too slow convergence. Through this step, the network parameters of the action generation module are gradually updated, making the generated control actions more in line with the requirements of the power system. At the same time, perform a soft update on the value evaluation module. The soft update strategy updates the parameters of the value evaluation module through a coefficient with a certain step size, making the parameters of the target network approach the parameters of the current network, but not making a large-scale update at one time, avoiding parameter instability. The update step size coefficient is set to 0.01. By gradually approaching the target network parameters, the parameters of the value evaluation module are gradually optimized, improving the accuracy of its evaluation of the power state. After updating the action generation module and the value evaluation module, the entire deep deterministic policy network undergoes multiple rounds of iterative training. Each iteration is based on the latest network parameters for training, enabling the system to continuously optimize the control strategy and gradually obtain the first control strategy. After multiple rounds of training, the deep deterministic policy network can make accurate decisions under changing power demands and environmental conditions, providing a more efficient and stable control strategy for the power system. Deploy the first control strategy to the control system of the emergency mobile power supply. In the control system, the input voltage, output current, and load power are adjusted in real time according to the first control strategy to ensure that the power supply responds quickly according to factors such as load changes and environmental changes during actual use, ensuring the stability and reliability of power supply. During the control process, record the system response data during the adjustment process. Perform segmented sampling on the system response data. Sample the system state according to the time step, record the power supply state parameters and control parameters at each sampling moment, and form complete state transition data. The state transition data within each period contains the current power supply state of the system, the control actions taken, and the corresponding reward information. By analyzing these data, evaluate the effect of the current control strategy and provide the necessary data support for the next round of policy optimization.

[0034] 105. Input the state transition data into the power loss compensation model for power loss compensation control analysis to obtain the compensation control quantity, and superimpose the compensation control quantity with the first control strategy to obtain the second control strategy. Based on the second control strategy, perform PWM modulation to generate the corresponding charge and discharge control signals.

[0035] Specifically, parse the state transition data. By extracting the data of line loss, switch loss, and energy storage unit loss, form multi-dimensional loss data, which represents the energy loss situation of each component in the power supply system during operation. Build a power loss compensation model based on the multi-dimensional loss data. This model models the power loss of the power supply system based on the loss data and calculates the distribution matrix of power loss. This matrix reflects the spatial distribution and temporal variation trend of various losses, and shows the performance of each loss type under different loads and operating states. Through this step, identify which losses have the greatest impact on the overall power supply efficiency, so as to perform targeted compensation. Perform a weighted summation operation on the power loss distribution matrix to calculate the compensation weight coefficient. Each loss term is assigned an appropriate compensation coefficient according to its impact on system stability. After weighted summation, obtain the compensation parameter matrix, and each element in it represents the compensation amount of a specific loss type. Based on this compensation parameter matrix, calculate various compensation control quantities, including line compensation quantity, switch compensation quantity, and energy storage compensation quantity. These compensation control quantities specifically reduce the negative impact of various losses on the overall performance of the system, ensuring that the power supply can operate with higher efficiency. Superimpose the compensation control quantity linearly with the first control strategy to obtain the second control strategy. Based on the second control strategy, perform PWM modulation to generate the corresponding charge and discharge control signals. To implement PWM modulation, use a triangular wave carrier signal to modulate and generate a PWM duty cycle sequence. In each cycle of the duty cycle sequence, the rising edge and the falling edge respectively correspond to the intersection points of the triangular wave carrier. This modulation method can accurately control the power output of the power supply. Perform dead-time compensation on the power regulation pulse signal and the charge and discharge control signal of the energy storage unit. The dead time refers to the response lag area of the signal when the switching element changes from the off state to the on state during the change process of the switching signal. By compensating for the dead time, avoid the unstable phenomenon caused by signal lag, ensure that each switching operation can be accurately executed, and finally obtain a stable charge and discharge control signal.

[0036] Normalize the line loss data in the power loss distribution matrix to ensure that all data types have the same dimension and comparability. Based on the input voltage and output current, calculate the impedance of the line, thereby determining the weight coefficient of the line loss, which reflects the relationship between the energy loss in the line and the input current and voltage. Conduct a frequency-domain analysis on the switching loss data in the power loss distribution matrix. The switching loss is mainly affected by the switching frequency and current stress. Using the frequency-domain analysis method, identify the switching losses at different frequencies through frequency-domain analysis. According to the switching frequency and current stress, calculate the conduction loss and turn-off loss during the switching process to obtain the weight coefficient of the switching loss, which measures the energy loss during the switching process. Similarly, conduct a time-series analysis on the energy storage unit loss data. The loss of the energy storage unit is affected by the charge-discharge state and the change in internal resistance. Monitor the change in the internal resistance of the battery during different charge-discharge processes and calculate the energy storage loss based on these changes. The change in internal resistance directly affects the energy loss of the battery. Through comprehensive analysis of the charge-discharge state and the change in internal resistance, calculate the weight coefficient of the energy storage loss, which reflects the loss degree of the energy storage unit under different working conditions. Combine the line loss weight coefficient, switching loss weight coefficient, and energy storage loss weight coefficient to construct a weight matrix. Normalize the weight matrix to obtain a compensation parameter matrix. Based on the compensation parameter matrix, calculate various compensation control quantities. Based on the compensation parameter matrix, calculate the line power compensation value. The calculation formula of the line compensation amount shows that the line compensation amount is proportional to the product of the input voltage and output current. According to the real-time data of the input voltage and output current, calculate the corresponding line power compensation value to effectively compensate for the energy loss caused by the line impedance, thereby optimizing the efficiency of the power supply system. Similarly, based on the compensation parameter matrix, calculate the switching power compensation value. The calculation basis of the switching compensation amount is the product of the switching frequency and the switching current. According to the real-time frequency and current data during the switching process, calculate the corresponding switching power compensation value. This compensation amount is used to compensate for the loss generated by frequent switching operations, thereby avoiding the negative impact of high-frequency switching on the stability of the power supply system. Calculate the energy storage power compensation value based on the compensation parameter matrix. The energy storage compensation amount is proportional to the product of the charge-discharge current and the equivalent internal resistance. Based on the current data during the charge-discharge process and the change in the battery internal resistance, calculate the energy storage power compensation value to effectively reduce the energy loss during the battery charge-discharge process and improve the battery usage efficiency. Linearly superimpose the line power compensation value, switching power compensation value, and energy storage power compensation value to obtain the compensation control quantity.

[0037] In the embodiments of the present invention, by introducing adaptive wavelet transform for data denoising and feature extraction, combining maximum-minimum normalization calculation and time series analysis, a complete data preprocessing process is established, effectively improving the quality and usability of the original data. An extended convolutional capsule network is used for hierarchical feature extraction. Through multi-level feature mapping and dynamic routing mechanisms, the effective separation of global power supply features and local power supply features is achieved, enhancing the system's perception ability of the power supply state. A control architecture based on a deep deterministic policy network is designed. Through the collaborative action of the value evaluation module and the action generation module, a complete state evaluation and action generation mechanism is established, improving the accuracy and adaptability of the control strategy. An experience replay pool is constructed to store historical data. The control strategy is optimized by using random sampling and temporal difference calculation methods. Through multiple rounds of iterative training and soft update mechanisms, the continuous optimization and stable convergence of the control strategy are ensured. A multi-dimensional power loss compensation scheme is proposed. By accurately modeling and compensating line losses, switch losses, and energy storage losses, the power loss of the system is significantly reduced, and the power supply efficiency is improved. A charge and discharge control mechanism based on PWM modulation is designed. Through dead time compensation and triangular wave carrier modulation, accurate power regulation and energy storage unit control are achieved, ensuring the stable operation of the system.

[0038] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0039] Collect data on the input voltage, output current, load power, and battery state of the emergency mobile power supply to obtain the original operating parameter data;

[0040] Perform data segmentation processing on the original operating parameter data and calculate the wavelet coefficients of each segment of data to obtain the initial wavelet decomposition result;

[0041] Set an adaptive threshold function based on the initial wavelet decomposition result to obtain the optimized threshold parameters, and substitute the optimized threshold parameters into the wavelet inverse transform equation to perform reconstruction operations on the initial wavelet decomposition result to obtain the denoised operating parameter data;

[0042] Perform maximum-minimum normalization calculation on the denoised operating parameter data to obtain the standardized data, and extract static feature parameters according to the inherent parameters of the power supply system from the standardized data to obtain a set of static feature parameters;

[0043] Perform time series analysis on the standardized data, calculate the voltage volatility, power change rate, and load change trend to obtain a set of dynamic feature parameters, and perform matrix splicing operations on the set of static feature parameters and the set of dynamic feature parameters to obtain the standardized feature matrix.

[0044] Specifically, the input voltage of the power supply is collected in real time from the sensor 、output current , load power and battery status , which respectively represent the input electrical energy of the system, the output current flow rate, the power consumed by the system load, and the current working state of the battery (such as remaining power or temperature). Assume that the collected time series data is , where represents time. Segment these data, and divide the entire time series into data segments of length . Each data segment is represented as , where is the segment index, . Then perform wavelet transform on each data segment to calculate the wavelet coefficients. Wavelet transform is a time-frequency analysis method. By selecting a wavelet basis function , decompose the input data . The continuous wavelet transform formula is:

[0045] ;

[0046] where, is the wavelet coefficient, is the scale parameter (controlling the frequency resolution), is the translation parameter (controlling the time resolution), is the complex conjugate of the wavelet basis function. By calculating , obtain the initial wavelet decomposition result. Set an adaptive threshold function based on the initial wavelet decomposition result. The set threshold function adaptively adjusts according to the amplitude of the wavelet coefficients. For example, use the soft threshold method:

[0047] ;

[0048] where, is the standard deviation of the noise, is the data length. For each wavelet coefficient , if , then consider it as noise and set it to zero; if , then reduce it. Substitute the processed wavelet coefficients into the wavelet inverse transform equation:

[0049] ;

[0050] where, is the data after noise reduction, is the regularization constant of the wavelet basis function. After completing noise reduction, perform normalization processing on the noise-reduced data. Use the maximum-minimum normalization method to map all data to the [0,1] interval. The normalization formula is:

[0051] ;

[0052] Among them, and are the minimum and maximum values of the data respectively, and is the normalized data. The normalized data is subjected to static feature extraction according to the inherent parameters of the power system. Static features are parameters reflecting the overall state of the data, such as the average value of voltage , the variance of the output current , the maximum value of the load power , etc. The static feature set is expressed as:

[0053] ;

[0054] Perform time series analysis on the normalized data to extract dynamic features. Dynamic features describe the changing trend of the data over time, such as voltage volatility , power change rate , load change trend , etc. The dynamic feature set is expressed as:

[0055] ;

[0056] Concatenate the static feature set with the dynamic feature set to form a standardized feature matrix :

[0057] ;

[0058] The standardized feature matrix describes the operating state of the emergency mobile power supply, including static overall features and dynamic changing trends.

[0059] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0060] Perform dimension reconstruction and feature combination operations on the standardized feature matrix to obtain a multi-dimensional feature vector, which includes static feature data and dynamic feature data;

[0061] Input the multi-dimensional feature vector into the first convolutional layer of the extended convolutional capsule network for feature mapping. The first convolutional layer contains 32 3×3 convolutional kernels, uses the ReLU activation function and batch normalization processing to obtain a primary feature map;

[0062] Perform a max pooling operation on the primary feature map. The pooling kernel size is 2×2 and the stride is 2 to obtain a downsampled feature map, and input the downsampled feature map into the main capsule layer of the extended convolutional capsule network;

[0063] Perform tensor transformation and direction voting calculation on the dimensionality-reduced feature map in the main capsule layer. The main capsule layer contains 8 capsule units, and each capsule unit outputs a 16-dimensional vector to obtain the main capsule features;

[0064] Input the main capsule features into the digital capsule layer of the extended convolutional capsule network. The digital capsule layer contains 10 capsule units, and use the dynamic routing algorithm for feature routing. The number of routing iterations is 3 to obtain the routed feature vector;

[0065] Perform global average pooling operation on the routed feature vector to obtain the global power supply feature, and perform local attention weighting calculation on the routed feature vector to obtain the local power supply feature. The global power supply feature reflects the overall power supply state of the system, and the local power supply feature characterizes the working state of each power supply unit.

[0066] Specifically, reconstruct the input standardized feature matrix Assume is a matrix, where represents the number of features, represents the time step. Through the feature combination operation, the static feature data and the dynamic feature data of each column are concatenated into a group of multi-dimensional feature vectors . The multi-dimensional feature vector is expressed as:

[0067] ;

[0068] where, represents the static feature of the th column, such as the average voltage or load power; represents the dynamic feature of the th column, such as the voltage change rate or power fluctuation trend. In this way, the static and dynamic information of the feature matrix is fused to obtain a multi-dimensional feature vector set containing two parts of features. Input the multi-dimensional feature vector into the first convolutional layer of the extended convolutional capsule network for feature mapping. The first convolutional layer contains 32 convolution kernels, and the convolution operation is represented by the following formula:

[0069] ;

[0070] where, is the output feature of the th convolution kernel at the position , is the weight of the convolution kernel, is the bias, is the ReLU activation function, which is used to introduce non-linearity. Through this convolution operation, each convolution kernel extracts local patterns in the input data, such as the local law of voltage fluctuations or the instantaneous change trend of load power. At the same time, the convolution layer adopts the batch normalization technique to normalize each convolution feature, and the formula is:

[0071] ;

[0072] where, is the mean of the mini-batch data, is the variance, is a very small value to prevent the denominator from being zero. After ReLU and batch normalization processing, the primary feature map is obtained. The size of each feature map is obtained by subtracting the convolution kernel size from the input matrix, and the formula is:

[0073] ;

[0074] where, and are the height and width of the input matrix respectively, is the size of the convolution kernel. The primary feature map is input into the max-pooling layer after passing through the convolution layer for dimensionality reduction. The max-pooling operation is implemented by a pooling kernel and a sliding window with a stride of 2. The formula is:

[0075] ;

[0076] where, is the value of the pooled feature map. By selecting the maximum value within the sliding window, the size of the feature map is effectively reduced while retaining the most significant features. The dimensionality-reduced feature map is input into the main capsule layer of the extended convolutional capsule network. In the main capsule layer, the dimensionality-reduced feature map is processed through a tensor transformation and a direction voting mechanism. The main capsule layer contains 8 capsule units, and each capsule unit outputs a 16-dimensional vector , and its formula is:

[0077] ;

[0078] where, is the input feature vector, is the weight matrix, is the direction voting score, is the scaling function used to normalize the amplitude of the output. In this way, each capsule unit captures a specific power supply mode, such as global voltage fluctuations or power anomalies in a certain local unit. The main capsule features are input into the digital capsule layer for processing. The digital capsule layer contains 10 capsule units and uses a dynamic routing algorithm for feature routing. The formula for dynamic routing is:

[0079] ;

[0080] Among them, is the routing score, and are the feature vectors of the main capsule layer and the digital capsule layer respectively. After 3 routing iterations, the output routing feature vector contains comprehensive information of multi-layer features. Perform global average pooling operation on the routing feature vector, and the formula is:

[0081] ;

[0082] Global power supply feature reflects the overall power supply status of the power system, such as total power stability and average load volatility. At the same time, perform local attention weighting calculation on the routing feature vector, and the formula is:

[0083] ;

[0084] Among them, the attention weight is calculated through a weighting function and is used to focus on the description of certain key features. The local power supply feature reflects the working status of each power supply unit, such as the charge and discharge characteristics of a specific battery unit or the abnormal change of a certain path current.

[0085] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0086] Concatenate the global power supply feature and the local power supply feature to obtain a combined feature vector;

[0087] Perform dimension mapping on the combined feature vector, and construct a value evaluation module of the deep deterministic policy network through a three-layer fully connected neural network;

[0088] Perform dimensionality reduction processing on the combined feature vector, and construct an action generation module of the deep deterministic policy network through a four-layer fully connected neural network;

[0089] Input the combined feature vector into the value evaluation module, calculate the score for the current power supply status, and obtain a state score result. The state score result includes a power supply stability score, a power balance score, and a response speed score;

[0090] Input the combined feature vector into the action generation module to generate corresponding control action parameters. The control action parameters include a power adjustment instruction and a charge and discharge instruction for the energy storage unit;

[0091] Set a reward function for the state scoring result, calculate the reward value based on the power supply stability, power balance degree, and system response speed, and obtain the state value evaluation result;

[0092] Construct an experience replay pool with a storage capacity of 10,000. Store the state scoring result, control action parameters, and state value evaluation result in the experience replay pool in chronological order.

[0093] Specifically, splice the global power supply characteristics and local power supply characteristics. The global power supply characteristics represent the overall power supply state of the power system, such as the average stability of the system voltage and the volatility of the power output, while the local power supply characteristics refine to the state of each power supply unit, including the charge and discharge conditions of the energy storage unit and the power supply efficiency of the local load. The feature splicing generates a combined feature vector by merging these two parts of feature vectors by dimension, and its expression is:

[0094] ;

[0095] where the combined feature vector contains the global and local feature information of the system and can comprehensively describe the operating state of the emergency mobile power supply. Input the combined feature vector into the value evaluation module and the action generation module. Perform dimension mapping on to reduce the complexity of the features and extract meaningful features. In the value evaluation module, construct a three-layer fully connected neural network, and the output of each layer represents the gradual mapping and feature extraction of the input features. For the output of the th layer, it is expressed as:

[0096] ;

[0097] where is the weight matrix of the th layer, is the bias term, is the activation function (such as ReLU). The first layer maps the input features to the hidden space, the second layer extracts more abstract features, and the third layer outputs the score of the current power supply state. The scoring results include the power supply stability score , the power balance score , and the response speed score . The state scoring result is expressed as:

[0098] ;

[0099] Among them, Measure the voltage and power stability of the system, Reflect the power supply balance degree of the load, Indicate the response speed of the power supply to the load change. At the same time, the combined feature vector Is input into the action generation module for dimensionality reduction processing to generate corresponding control action parameters. In the action generation module, through a four-layer fully connected neural network for Perform step-by-step dimensionality reduction, and the output is the control action parameter . The output calculation formula of the action generation module is similar to that of the value evaluation module, expressed as:

[0100] ;

[0101] Among them, Contains two main control instructions: the power regulation instruction And the charge and discharge instruction of the energy storage unit . The power regulation instruction Controls the overall power output of the power supply system, while the charge and discharge instruction Determines the charging or discharging strategy of the energy storage unit at the current moment. Define a reward function For the state scoring result , which is used to measure the quality of the current control strategy. The reward function comprehensively considers power supply stability, power balance degree and system response speed, and the formula is expressed as:

[0102] ;

[0103] Among them, , And Are the weight parameters of the reward function, used to adjust the influence of different scoring items on the total reward. By calculating the reward value , evaluate the control strategy according to the state scoring result, and obtain the state value evaluation result , and its value is:

[0104] ;

[0105] Among them, Is the discount factor, is the value estimate of the next state, which is used to balance short-term benefits and long-term benefits. In order to optimize the control strategy, an experience replay pool is constructed to store historical states, actions, and reward information. The storage capacity of the experience replay pool is set to 10,000, which can save enough historical samples to support subsequent training optimization. Each execution of the control strategy generates a new sample, including the state scoring result S, the control action parameter A, and the state value evaluation result Q. These samples are stored in the experience replay pool in time sequence to provide training data for the deep reinforcement learning algorithm.

[0106] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0107] Randomly sample historical data in the experience replay pool to obtain training data batches;

[0108] Perform time series difference calculation on the training data batches, calculate the target Q value based on the state transition sequence, and use the target Q value as the optimization target of the value assessment module;

[0109] The training data batches are input into the action generation module, and the network parameters of the action generation module are gradient updated. The Adam optimizer is used for parameter optimization, and the learning rate is set to 0.001;

[0110] The value assessment module is soft-updated, the update step coefficient is set to 0.01, and the parameters of the value assessment module are brought closer to the target network parameters to obtain the updated value assessment module;

[0111] Based on the updated value assessment module and action generation module, the deep deterministic policy network is trained for multiple rounds to obtain the first control strategy;

[0112] Deploy the first control strategy to the control system of the emergency mobile power supply, adjust the input voltage, output current and load power in real time, and record the system response data during the adjustment process;

[0113] The system response data is sampled in sections, and the power supply state parameters and control parameters at each sampling moment are recorded to obtain the state transfer data.

[0114] Specifically, historical data is randomly sampled from the experience replay pool to generate a training data batch. This data contains the states recorded during the previous system operation. ,action ,award and the next state . Assume that the size of the experience replay pool is , the size of the training data batch is , the random sampling process is represented by drawing from the playback pool without replacement A sample, denoted as:

[0115] ;

[0116] Among them, is the training data set of the current batch, represents the state of the th sample, is the corresponding control action, is the reward obtained after executing the action, is the next state after the state transition. Temporal difference calculation is performed on each sample in the training data batch to estimate the target Q value. The temporal difference method defines the target Q value according to the Bellman equation , and the formula is:

[0117] ;

[0118] Among them, is the discount factor, used to balance short-term rewards and long-term rewards, is the next state under the maximum Q value. Through temporal difference calculation, combined with the current reward and future potential benefits, the target Q value of each sample is estimated. This target value is used as the optimization target of the value evaluation module to guide the update of network parameters. The training data batch is input into the action generation module to update the gradients of its network parameters. The action generation module predicts the optimal action by fitting the function, where is the policy function, representing the mapping relationship of selecting an action under the state . In order to optimize the parameters of the action generation module, the Adam optimizer is used. Adam is an adaptive optimization algorithm based on the first-order gradient, and its update formula is:

[0119] ;

[0120] Among them, is the parameter at the th iteration, is the learning rate (set to 0.001), and are the first-order momentum and second-order momentum of the gradient respectively, It is a small value to avoid a zero denominator. Through the Adam optimizer, the network parameters of the action generation module can be updated in an efficient and stable manner. At the same time, the value evaluation module is softly updated to improve its prediction ability. The principle of soft update is to gradually move the parameters of the current network closer to the parameters of the target network, thereby smoothly updating the model and avoiding instability caused by large parameter adjustments. The formula for soft update is:

[0121] ;

[0122] where, is the update step coefficient (set to 0.01), are the parameters of the current network, are the parameters of the target network. By approaching gradually, the updated value evaluation module can more accurately predict the Q value, providing reliable support for policy optimization. After optimizing the action generation module and the value evaluation module, the deep deterministic policy network is trained iteratively for multiple rounds based on the updated modules. Each round of iteration goes through a cycle of random sampling, temporal difference calculation, and parameter update, enabling the network to gradually learn the optimal control strategy for the emergency mobile power supply. After multiple rounds of training, the obtained first control strategy can output the optimal control action in real time according to the input state of the system. After deploying the first control strategy to the control system of the emergency mobile power supply, the input voltage , output current , and load power are adjusted in real time. Each adjustment performs power distribution and charge-discharge control of the energy storage unit according to the current state and the action generated by the policy . During the control process, the system response data at each adjustment moment is recorded in real time, including state parameters and control parameters. The system response data is sampled in segments, and each segment of data is used as a state transition sample, recording the current state, the executed action, and the state information after the transition. By continuously accumulating these data, the control strategy is optimized.

[0123] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0124] Parse and process the state transition data, extract the line loss data, switch loss data, and energy storage unit loss data to obtain multi-dimensional loss data;

[0125] Construct a power loss compensation model based on the multi-dimensional loss data, and input the multi-dimensional loss data into the power loss compensation model to obtain a power loss distribution matrix;

[0126] Perform weighted summation calculation on the power loss distribution matrix, determine the compensation weight coefficients for various types of losses, obtain the compensation parameter matrix, and calculate the compensation control quantity according to the compensation parameter matrix. The compensation control quantity includes line compensation quantity, switch compensation quantity, and energy storage compensation quantity;

[0127] Perform linear superposition operation on the compensation control quantity and the first control strategy to obtain the second control strategy, and perform triangular wave carrier modulation based on the second control strategy to obtain the PWM duty cycle sequence;

[0128] Generate a power regulation pulse signal and an energy storage unit charge and discharge control signal based on the PWM duty cycle sequence. The rising edge and falling edge of the power regulation pulse signal correspond to the intersection points of the triangular wave carrier;

[0129] Perform dead-time compensation on the power regulation pulse signal and the energy storage unit charge and discharge control signal to obtain the final charge and discharge control signal.

[0130] Specifically, identify the key parameters related to various types of losses from the state transition data. Assume that the state transition data includes input voltage , output current , switching frequency , charge and discharge current , as well as equivalent line impedance and the internal resistance of the energy storage unit . Among them and represent the input power state of the power supply system, reflects the operating frequency of the switch, reflects the working intensity of the energy storage unit, and are respectively the main factors of line loss and energy storage unit loss. For line loss, the line loss power is calculated by the following formula:

[0131] ;

[0132] Among them, is the output current, is the line impedance. This formula indicates that the line loss is proportional to the square of the output current and the line impedance. For switch loss, the switch loss is divided into two parts: conduction loss and turn-off loss. The conduction loss is calculated by the formula:

[0133] ;

[0134] The turn-off loss is calculated by the formula:

[0135] ;

[0136] Among them, is the switching voltage, and are the on-time and off-time of the switch respectively, is the switching frequency. The total switching loss is the sum of the conduction loss and the turn-off loss:

[0137] ;

[0138] For the loss of the energy storage unit, it is mainly determined by the charge and discharge current and the internal resistance of the energy storage unit. The energy storage loss power is expressed as:

[0139] ;

[0140] The above three formulas respectively obtain the line loss, switching loss and energy storage loss data, and these data constitute a multi-dimensional loss data set Based on the multi-dimensional loss data , a power loss compensation model is constructed. After the loss data is input into the model, the power loss distribution matrix is calculated. Its rows represent the time series, and its columns represent different loss types. Each element represents the th type of loss value at the th time point. The power loss distribution matrix is weighted and summed to determine the compensation weight coefficients

[0141] ;

[0142] Among them, is the weight of the th type of loss, is the length of the time series. The larger the weight, the more significant the impact of this type of loss on the system. Based on the compensation weight coefficients, a compensation parameter matrix is constructed, and the compensation control quantity is calculated. The line compensation quantity is proportional to the product of the input voltage and the output current, and the formula is:

[0143] ;

[0144] The switch compensation quantity is proportional to the product of the switching frequency and the switching current, and the formula is:

[0145] ;

[0146] Energy storage compensation amount Is proportional to the product of the charge and discharge current and the internal resistance, and the formula is:

[0147] ;

[0148] Linearly superimpose the compensation control amount and the first control strategy To obtain the second control strategy :

[0149] ;

[0150] Based on the second control strategy, generate a PWM duty cycle sequence through triangular wave carrier modulation. The duty cycle sequence is determined by the intersection of the control signal and the triangular wave:

[0151] ;

[0152] Among them, Is the triangular wave signal, Is the output signal of PWM modulation. Perform dead-time compensation on the power regulation pulse signal and the charge and discharge signal of the energy storage unit, adjust the switch timing, and ensure the non-response time between the rising edge and the falling edge. The compensated signal is used as the final charge and discharge control signal and applied to the control of the emergency mobile power supply to optimize the system operation efficiency.

[0153] In a specific embodiment, the process of performing weighted summation calculation on the power loss distribution matrix, determining the compensation weight coefficients for various types of losses, obtaining the compensation parameter matrix, and calculating the compensation control amount according to the compensation parameter matrix can specifically include the following steps:

[0154] Normalize the line loss data in the power loss distribution matrix, calculate the line impedance based on the input voltage and output current, and obtain the line loss weight coefficient;

[0155] Perform frequency-domain analysis on the switch loss data in the power loss distribution matrix, calculate the conduction loss and turn-off loss based on the switching frequency and current stress, and obtain the switch loss weight coefficient;

[0156] Perform time-series analysis on the energy storage unit loss data in the power loss distribution matrix, calculate the energy storage loss based on the charge and discharge state and the change of internal resistance, and obtain the energy storage loss weight coefficient;

[0157] Combine the line loss weight coefficient, the switch loss weight coefficient, and the energy storage loss weight coefficient to construct a weight matrix, and normalize the weight matrix to obtain the compensation parameter matrix;

[0158] Calculate the line power compensation value based on the compensation parameter matrix. The line compensation amount is proportional to the product of the input voltage and the output current, and calculate the switch power compensation value based on the compensation parameter matrix. The switch compensation amount is proportional to the product of the switching frequency and the switching current;

[0159] Calculate the energy storage power compensation value based on the compensation parameter matrix. The energy storage compensation amount is proportional to the product of the charge and discharge current and the equivalent internal resistance, and linearly superimpose the line power compensation value, the switch power compensation value and the energy storage power compensation value to obtain the compensation control amount.

[0160] Specifically, normalize the line loss data in the power loss distribution matrix. Assume the power loss distribution matrix The element represents the data of the th type of loss at the th time point, where represents the line loss, represents the switch loss, represents the energy storage unit loss. When normalizing the line loss data, use the maximum-minimum normalization method, and the formula is as follows:

[0161] ;

[0162] Among them, is the normalized line loss data, and represent the minimum and maximum values of the line loss data respectively. The normalization process eliminates the influence of different data dimensions, making subsequent analysis more unified. After obtaining the normalized line loss data, calculate the line impedance based on the input voltage and the output current , and its formula is:

[0163] ;

[0164] The line loss weight coefficient is calculated by the proportion of the line loss in the total loss, and the formula is:

[0165] ;

[0166] Among them, is the time step, is the line loss weight coefficient, indicating the importance of the line loss in the multi-dimensional loss. Perform frequency domain analysis on the switch loss data in the power loss distribution matrix to calculate the conduction loss and the turn-off loss. The two main components of the switch loss are the conduction loss and the turn-off loss 。The calculation formula for conduction loss is:

[0167] ;

[0168] The calculation formula for turn-off loss is:

[0169] ;

[0170] Among them, is the switch voltage, and are the conduction time and turn-off time respectively, is the switching frequency. represents the switching loss weight coefficient, and the switching loss weight coefficient is calculated through normalization, and the formula is:

[0171] ;

[0172] The switching loss weight coefficient reflects the influence degree of the switching frequency on the total loss. At the same time, perform time series analysis on the loss data of the energy storage unit. The energy storage loss is mainly determined by the charge and discharge current and the equivalent internal resistance of the energy storage unit. The energy storage loss power The calculation formula is:

[0173] ;

[0174] In the time series analysis, the charge and discharge state of the energy storage unit is represented by the positive and negative values of , positive value represents charging, and negative value represents discharging. represents the energy storage loss weight coefficient, and the calculation formula for the energy storage loss weight coefficient is:

[0175] ;

[0176] The weight coefficient reflects the proportion of the energy storage unit loss in the total loss. Combine the line loss weight coefficient , the switching loss weight coefficient and the energy storage loss weight coefficient into a weight matrix . After normalizing the weight matrix, the compensation parameter matrix is obtained:

[0177] ;

[0178] Among them, is the normalized compensation parameter. Based on the compensation parameter matrix, calculate the line power compensation value Its formula is:

[0179] ;

[0180] Calculate the switching power compensation value :

[0181] ;

[0182] Calculate the energy storage power compensation value :

[0183] ;

[0184] Among them, represents the line compensation amount, represents the switching compensation amount, represents the energy storage compensation amount. The line power compensation value, the switching power compensation value, and the energy storage power compensation value are linearly superimposed to obtain the compensation control amount :

[0185] .

[0186] The control method of the emergency mobile power supply in the embodiment of the present invention is described above. Next, the emergency mobile power supply in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the emergency mobile power supply in the embodiment of the present invention includes:

[0187] The acquisition module 201 is used to collect the operation parameter data of the emergency mobile power supply and perform adaptive wavelet transform noise reduction to obtain a standardized feature matrix;

[0188] The feature extraction module 202 is used to construct a multi-dimensional feature vector by using the standardized feature matrix, and input the multi-dimensional feature vector into the extended convolutional capsule network for hierarchical feature extraction to obtain the global power supply feature and the local power supply feature;

[0189] The establishment module 203 is used to establish a deep deterministic policy network based on the global power supply feature and the local power supply feature. The deep deterministic policy network includes a value evaluation module and an action generation module; use the value evaluation module to perform a power supply state score to obtain a state score result, and store the state score result and the corresponding control action in the experience replay pool;

[0190] The iteration module 204 is used to input the historical data in the experience replay pool into the action generation module, optimize and iterate the control action to obtain the first control strategy; use the first control strategy for online control and record the state transition data during the control process;

[0191] A generation module 205 is configured to input the state transition data into a power loss compensation model for power loss compensation control analysis, obtain a compensation control quantity, superimpose the compensation control quantity with a first control strategy to obtain a second control strategy, and perform PWM modulation based on the second control strategy to generate corresponding charge and discharge control signals.

[0192] Through the collaborative cooperation of the above-mentioned various components, by introducing adaptive wavelet transform for data denoising and feature extraction, combining maximum-minimum normalization calculation and time series analysis, a complete data preprocessing process is established, effectively improving the quality and usability of the original data. An extended convolutional capsule network is used for hierarchical feature extraction. Through multi-level feature mapping and dynamic routing mechanisms, the effective separation of global power supply features and local power supply features is achieved, enhancing the system's perception ability of the power supply state; a control architecture based on a deep deterministic policy network is designed. Through the collaborative action of a value evaluation module and an action generation module, a complete state evaluation and action generation mechanism is established, improving the accuracy and adaptability of the control strategy; an experience replay pool is constructed to store historical data, and random sampling and temporal difference calculation methods are used to optimize the control strategy. Through multiple rounds of iterative training and soft update mechanisms, the continuous optimization and stable convergence of the control strategy are ensured; a multi-dimensional power loss compensation scheme is proposed. By accurately modeling and compensating line losses, switch losses, and energy storage losses, the power loss of the system is significantly reduced, improving the power supply efficiency; a charge and discharge control mechanism based on PWM modulation is designed. Through dead time compensation and triangular wave carrier modulation, accurate power regulation and energy storage unit control are achieved, ensuring the stable operation of the system.

[0193] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0194] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0195] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for an emergency mobile power supply, characterized in that: The method comprises: The operating parameter data of the emergency mobile power supply is collected and denoised by adaptive wavelet transform to obtain a standardized feature matrix; Using the standardized feature matrix to construct a multidimensional feature vector, and inputting the multidimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain global power supply features and local power supply features; A deep deterministic strategy network is established based on the global power supply characteristics and the local power supply characteristics, wherein the deep deterministic strategy network includes a value assessment module and an action generation module; the value assessment module is used to score the power supply status to obtain a status scoring result, and the status scoring result and the corresponding control action are stored in an experience replay pool; Inputting the historical data in the experience replay pool into the action generation module, optimizing and iterating the control action to obtain a first control strategy; performing online control using the first control strategy, and recording state transition data during the control process; The state transfer data is input into a power loss compensation model for power loss compensation control analysis to obtain a compensation control amount, and the compensation control amount is superimposed on the first control strategy to obtain a second control strategy. PWM modulation is performed based on the second control strategy to generate corresponding charge and discharge control signals.

2. The control method of the emergency mobile power supply according to claim 1, characterized in that: The operation parameter data of the emergency mobile power supply is collected and the adaptive wavelet transform is used for noise reduction to obtain a standardized feature matrix, including: Collect data on the input voltage, output current, load power and battery status of the emergency mobile power supply to obtain original operating parameter data; Performing data segmentation processing on the original operating parameter data, and calculating the wavelet coefficient of each segment of data to obtain an initial wavelet decomposition result; An adaptive threshold function is set based on the initial wavelet decomposition result to obtain an optimized threshold parameter, and the optimized threshold parameter is substituted into an inverse wavelet transform equation to perform a reconstruction operation on the initial wavelet decomposition result to obtain noise-reduced operating parameter data; Performing maximum and minimum value normalization calculation on the noise-reduced operating parameter data to obtain standardized data, and performing static feature extraction on the standardized data according to inherent parameters of the power supply system to obtain a static feature parameter set; The standardized data is subjected to time series analysis, the voltage fluctuation rate, the power change rate and the load change trend are calculated to obtain a dynamic characteristic parameter set, and the static characteristic parameter set is subjected to matrix concatenation operation with the dynamic characteristic parameter set to obtain the standardized characteristic matrix.

3. The control method of the emergency mobile power supply according to claim 1, characterized in that: The method of constructing a multidimensional feature vector by using the standardized feature matrix and inputting the multidimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain global power supply features and local power supply features includes: Performing dimension reconstruction and feature combination operations on the standardized feature matrix to obtain a multidimensional feature vector, wherein the multidimensional feature vector includes static feature data and dynamic feature data; Inputting the multidimensional feature vector into the first convolutional layer of the extended convolutional capsule network for feature mapping, wherein the first convolutional layer includes 32 3×3 convolution kernels, uses a ReLU activation function and batch normalization processing, and obtains a primary feature map; Performing a maximum pooling operation on the primary feature map, with a pooling kernel size of 2×2 and a step size of 2, to obtain a reduced-dimensional feature map, and inputting the reduced-dimensional feature map into the main capsule layer of the extended convolutional capsule network; Performing tensor transformation and direction voting calculation on the dimension reduction feature map in the main capsule layer, the main capsule layer includes 8 capsule units, each capsule unit outputs a 16-dimensional vector, and obtaining a main capsule feature; Input the main capsule feature into the digital capsule layer of the extended convolutional capsule network, the digital capsule layer includes 10 capsule units, use a dynamic routing algorithm to perform feature routing, the number of routing iterations is 3, and obtain a routing feature vector; A global average pooling operation is performed on the routing feature vector to obtain a global power supply feature, and a local attention weighted calculation is performed on the routing feature vector to obtain a local power supply feature, wherein the global power supply feature reflects the overall power supply status of the system, and the local power supply feature characterizes the working status of each power supply unit.

4. The control method of the emergency mobile power supply according to claim 1, characterized in that: The method includes establishing a deep deterministic strategy network based on the global power supply characteristics and the local power supply characteristics, wherein the deep deterministic strategy network includes a value assessment module and an action generation module; using the value assessment module to score the power supply status, obtaining a status scoring result, and storing the status scoring result and the corresponding control action in an experience replay pool, including: Perform feature concatenation on the global power supply feature and the local power supply feature to obtain a combined feature vector; Performing dimension mapping on the combined feature vector, and constructing a value assessment module of a deep deterministic strategy network through a three-layer fully connected neural network; Performing dimensionality reduction processing on the combined feature vector, and constructing an action generation module of the deep deterministic strategy network through a four-layer fully connected neural network; The combined feature vector is input into the value assessment module to score the current power supply status to obtain a status scoring result, wherein the status scoring result includes a power supply stability score, a power balance score and a response speed score; Inputting the combined feature vector into the action generation module to generate corresponding control action parameters, wherein the control action parameters include power adjustment instructions and energy storage unit charge and discharge instructions; A reward function is set for the state scoring result, and a reward value is calculated according to power supply stability, power balance and system response speed to obtain a state value evaluation result; An experience replay pool is constructed, the storage capacity of the experience replay pool is 10000, and the state scoring result, the control action parameter and the state value evaluation result are stored in the experience replay pool in chronological order.

5. The control method of the emergency mobile power supply according to claim 4, characterized in that: The historical data in the experience replay pool is input into the action generation module, and the control action is optimized and iterated to obtain a first control strategy; The first control strategy is used to perform online control and record state transition data during the control process, including: Randomly sampling historical data in the experience replay pool to obtain training data batches; Performing time series difference calculation on the training data batch, calculating the target Q value based on the state transition sequence, and the target Q value is used as the optimization target of the value evaluation module; Input the training data batches into the action generation module, perform gradient update on the network parameters of the action generation module, use the Adam optimizer to perform parameter optimization, and set the learning rate to 0.001; The value assessment module is soft-updated, the update step coefficient is set to 0.01, and the parameters of the value assessment module are brought closer to the target network parameters to obtain an updated value assessment module; Based on the updated value assessment module and the action generation module, performing multiple rounds of iterative training on the deep deterministic strategy network to obtain a first control strategy; Deploy the first control strategy to the control system of the emergency mobile power supply, adjust the input voltage, output current and load power in real time, and record the system response data during the adjustment process; The system response data is sampled in sections, and the power supply state parameters and control parameters at each sampling moment are recorded to obtain state transfer data.

6. The control method of the emergency mobile power supply according to claim 1, characterized in that: The step of inputting the state transition data into a power loss compensation model to perform power loss compensation control analysis to obtain a compensation control amount, superimposing the compensation control amount with the first control strategy to obtain a second control strategy, and performing PWM modulation based on the second control strategy to generate a corresponding charge and discharge control signal includes: Analyze and process the state transition data, extract line loss data, switch loss data and energy storage unit loss data, and obtain multi-dimensional loss data; Constructing a power loss compensation model based on the multi-dimensional loss data, and inputting the multi-dimensional loss data into the power loss compensation model to obtain a power loss distribution matrix; Performing weighted sum calculation on the power loss distribution matrix, determining compensation weight coefficients of various types of losses, obtaining a compensation parameter matrix, and calculating a compensation control amount according to the compensation parameter matrix, wherein the compensation control amount includes a line compensation amount, a switch compensation amount, and an energy storage compensation amount; Performing a linear superposition operation on the compensation control amount and the first control strategy to obtain a second control strategy, and performing triangular wave carrier modulation based on the second control strategy to obtain a PWM duty cycle sequence; Generate a power regulation pulse signal and an energy storage unit charge and discharge control signal based on the PWM duty cycle sequence, wherein the rising edge and the falling edge of the power regulation pulse signal correspond to the intersection of the triangular wave carrier; Dead time compensation is performed on the power regulation pulse signal and the energy storage unit charge and discharge control signal to obtain a final charge and discharge control signal.

7. The control method of the emergency mobile power supply according to claim 6, characterized in that: The power loss distribution matrix is ​​weighted and summed to determine the compensation weight coefficients of various types of losses, to obtain a compensation parameter matrix, and a compensation control amount is calculated according to the compensation parameter matrix, wherein the compensation control amount includes a line compensation amount, a switch compensation amount and an energy storage compensation amount, including: Normalizing the line loss data in the power loss distribution matrix, calculating the line impedance based on the input voltage and the output current, and obtaining a line loss weight coefficient; Performing frequency domain analysis on the switching loss data in the power loss distribution matrix, calculating the turn-on loss and turn-off loss based on the switching frequency and the current stress, and obtaining a switching loss weight coefficient; Performing a time series analysis on the energy storage unit loss data in the power loss distribution matrix, calculating the energy storage loss based on the charge and discharge state and the internal resistance change, and obtaining the energy storage loss weight coefficient; The line loss weight coefficient, the switch loss weight coefficient and the energy storage loss weight coefficient are combined to construct a weight matrix, and the weight matrix is ​​normalized to obtain a compensation parameter matrix; Calculating a line power compensation value based on the compensation parameter matrix, wherein the line compensation amount is proportional to the product of the input voltage and the output current, and calculating a switch power compensation value based on the compensation parameter matrix, wherein the switch compensation amount is proportional to the product of the switching frequency and the switching current; The energy storage power compensation value is calculated based on the compensation parameter matrix, the energy storage compensation amount is proportional to the product of the charging and discharging current and the equivalent internal resistance, and the line power compensation value, the switch power compensation value and the energy storage power compensation value are linearly superimposed to obtain the compensation control amount.

8. An emergency mobile power supply, characterized in that: A control method for executing an emergency mobile power supply according to any one of claims 1 to 7, comprising: The acquisition module is used to collect the operating parameter data of the emergency mobile power supply and perform adaptive wavelet transform noise reduction to obtain a standardized feature matrix; A feature extraction module, used to construct a multidimensional feature vector using the standardized feature matrix, and input the multidimensional feature vector into an extended convolutional capsule network for hierarchical feature extraction to obtain global power supply features and local power supply features; An establishment module is used to establish a deep deterministic strategy network based on the global power supply characteristics and the local power supply characteristics, wherein the deep deterministic strategy network includes a value assessment module and an action generation module; the value assessment module is used to score the power supply status to obtain a status scoring result, and the status scoring result and the corresponding control action are stored in an experience replay pool; An iteration module, used for inputting the historical data in the experience replay pool into the action generation module, optimizing and iterating the control action, and obtaining a first control strategy; performing online control using the first control strategy, and recording state transition data during the control process; A generation module is used to input the state transfer data into a power loss compensation model to perform power loss compensation control analysis to obtain a compensation control amount, and superimpose the compensation control amount with the first control strategy to obtain a second control strategy, perform PWM modulation based on the second control strategy, and generate corresponding charge and discharge control signals.

Images