A control method for a multi-channel power output control system of an edge computing UPS

Abstract

The present application discloses a control method for a multi-power output control system of an edge computing UPS, which relates to the field of intelligent control. The method comprises: controlling the monitoring unit to monitor the real-time power input parameters of the line and the real-time load information corresponding to different output paths in real time; controlling the input prediction model in the edge computing unit to obtain the line predicted power input parameters according to the real-time power input parameters of the line and the historical power input parameters of the line; controlling the load prediction model in the edge computing unit to obtain the predicted load information according to the real-time load information and the historical load information; controlling the working state of the backup power supply based on the predicted power input parameters; and controlling the output power of different output ports of the multi-power output unit based on the predicted load information and the load priority information.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control, and more specifically, to a control method for a multi-power output control system of an edge computing UPS. Background Art

[0002] In modern power systems, with the continuous increase in power demand and the increasing complexity of power equipment, how to ensure the continuous power supply of key equipment has become an important research direction. Traditional power management systems usually rely on simple redundant designs, that is, to improve the reliability of power supply by equipping important loads with backup power supplies. However, this redundant design has many shortcomings in terms of power switching efficiency, power distribution accuracy, and energy utilization effectiveness.

[0003] At present, common power management methods mainly include the following:

[0004] 1. Static redundancy design: This method usually configures a backup power supply in advance, such as an uninterruptible power supply (UPS) or a diesel generator set, and automatically switches to the backup power supply when the main power supply fails. However, this method often cannot efficiently switch back to the main power supply after the main power supply is restored, and cannot respond in time to unforeseen load fluctuations or failures in the system, resulting in unstable power supply.

[0005] 2. Rule-based control: A simple threshold judgment method is used to set a fixed threshold based on basic electrical parameters such as voltage or current. When the parameter exceeds the set range, the backup power supply is triggered. However, since this method lacks dynamic perception of the overall system status, it often leads to untimely or erroneous power switching, and cannot achieve accurate matching of load demand.

[0006] 3. Single prediction method: Some advanced power management systems try to introduce load prediction and power input parameter prediction methods, but they usually rely on a single algorithm or model (such as simple time series analysis), which is difficult to cope with the complex and changing power demand and dynamic changes in load priority. In addition, such systems lack comprehensive management and intelligent scheduling of multiple output ports, which may lead to uneven power distribution or waste of resources when load demand changes suddenly. Summary of the invention

[0007] A series of simplified concepts are introduced in the Summary of the Invention section, which will be further described in detail in the Detailed Description of the Invention section. The Summary of the Invention section of this application does not mean to attempt to define the key features and essential technical features of the claimed technical solution, nor does it mean to attempt to determine the scope of protection of the claimed technical solution.

[0008] The present application proposes a control method for a multi-power output control system of an edge computing UPS, wherein the multi-power supply of the edge computing UPS includes a multi-power output unit, an edge computing unit, a monitoring unit, and a backup power supply unit.

[0009] The above methods include:

[0010] Control the monitoring unit to monitor the real-time power input parameters of the line and the real-time load information corresponding to different output paths in real time;

[0011] Control the input prediction model in the edge computing unit to obtain the line prediction power input parameter according to the line real-time power input parameter and the line historical power input parameter;

[0012] Control the load prediction model in the edge computing unit to obtain predicted load information according to the real-time load information and historical load information;

[0013] Controlling the working state of the backup power supply based on the predicted power input parameter;

[0014] The output power of different output ports of the multi-way power output unit is controlled based on the predicted load information and the load priority information.

[0015] In a feasible implementation manner, the power input parameters include voltage information, current information, frequency information, phase angle information, power factor information, total harmonic distortion information, instantaneous fluctuation information and power failure frequency information;

[0016] The above-mentioned input prediction model includes an input layer, a feature extraction module, a prediction module, a fusion module and an output layer;

[0017] The feature extraction module includes a time series feature extractor and a statistical feature extractor. The time feature extractor extracts time series features based on a sliding window method, a differential operation method, and a frequency domain analysis method. The statistical features extracted by the statistical feature extractor include the mean, variance, maximum value, and minimum value.

[0018] The above prediction module includes a target long short-term memory network and a target support vector regression network;

[0019] The fusion module is used to perform weighted fusion on the first output of the target long short-term memory network and the second output of the target support vector regression network to obtain the line prediction power input parameters.

[0020] In a feasible implementation, it also includes:

[0021] Inputting the time series features into the target long short-term memory network to obtain the first output;

[0022] The time series features and the statistical features are input into a target support vector regression network to obtain the second output, wherein the target support vector regression network is an SVR model of a radial basis function kernel.

[0023] In a feasible implementation manner, the specific steps of determining the first weight corresponding to the first output and the second weight corresponding to the second output include:

[0024] Each particle is represented as a weighted combination of a first weight and a second weight, wherein the particle swarm is randomly initialized in the weight space;

[0025] In each iteration, the velocity and position of the particles are updated so that the particles tend to the global optimal solution and the individual optimal solution;

[0026] Update the fitness of particles according to the prediction error after weighted fusion, and gradually find the global optimal weight combination;

[0027] In the global optimal weight combination, a sliding window is defined, and the prediction error of each model in the window is calculated to update the weight, and the first weight and the second weight updated in real time are obtained;

[0028] The update formula for the first weight w1 and the second weight w2 is:

[0029]

[0030] w1(t) and w2(t) are the first and second weights at time t, w1(t+1) and w2((t+1)) are the new weights at time t+1, and α is the learning rate MSE SVR and MSE LSTM is the prediction error of the target support vector regression network and the target long short-term memory network, MSE SVR is the error value of the SVR model at the current moment, MSE LSTM is the error value of the LSTM model at the current moment.

[0031] In a feasible implementation manner, the load information includes power demand, current demand, load type, load change rate and load working time; the load prediction model includes an input layer, a feature extraction layer, a convolution layer, a prediction recognition layer, a fully connected layer and an output layer;

[0032] The output of the above convolutional layer is:

[0033] C t =ReLU(W c *x t +b c )

[0034] Among them, * represents the convolution operation, ReLU is the activation function, Ct is the output feature of the convolution layer, x t is the feature data extracted by the feature extraction layer, W c is the convolution kernel weight, b c is bias;

[0035] The calculation formula of the prediction recognition layer is:

[0036] i t =σ(W i C t +U i h t-1 +b i )

[0037] f t =σ(W f C t +U f h t-1 +b f )

[0038] o t =σ(W o C t +U o h t-1 +b o )

[0039] c t =f t ⊙c t-1 +i t ⊙tanh(W c C t +U c h t-1 +b c )

[0040] h t =o t ⊙tanh(c t )

[0041] Among them, i t 、f t , o t They are the states of the input gate, forget gate, and output gate;

[0042] c t is the cell state, h is the hidden state;

[0043] W i , W f , W o , W c is the weight matrix, U i , U f , Uo , U c is the weight matrix associated with the hidden state at the previous moment, b i , b f , b o , b c is the bias term;

[0044] σ represents the sigmoid function, ⊙ represents element-by-element multiplication;

[0045] The calculation formula of the fully connected layer is:

[0046] y t =ReLU(W d h t +b d )

[0047] W d is the weight matrix of the fully connected layer, b d is the bias term, yt is the output feature of the fully connected layer;

[0048] The predicted load information obtained by the output layer is

[0049]

[0050] Among them, W o is the weight matrix of the output layer, b o is the bias term, For predicted load information.

[0051] In a feasible implementation manner, the controlling the working state of the backup power supply based on the predicted power input parameter includes:

[0052] Determine the optimization objective function;

[0053] Randomly initialize the particle swarm, where each particle represents a possible solution, and each possible solution includes position and velocity;

[0054] Encode the problem’s decision variables as particle positions to represent the state of the solution in the search space;

[0055] Update the velocity and position of each particle based on the following formula:

[0056] v i (t+1))=w·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(g(t)-x i (t))

[0057] xi (t+1)=x i (t)+v i (t+1)

[0058] Among them, w is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers used to control the movement and search behavior of particles.

[0059] Predict the fitness of particles based on the target deep learning model;

[0060] The first k particles with the highest fitness are retained to enter the next generation, and the remaining particles are reinitialized or mutated according to probability;

[0061] Calculate the objective function value for each particle's position and evaluate its quality through the fitness function;

[0062] When the maximum number of iterations is reached or the fitness function converges, the algorithm is terminated and the backup power supply control parameters are output;

[0063] The working state of the backup power supply is controlled based on the backup power supply control parameter.

[0064] In a feasible implementation, the above optimization objective function is:

[0065] f total (P, T, E, S) = α1f1 (P, T) + α2f2 (P, U) + α3f3 (P, E) - α4f4 (P, S)

[0066] Among them, α1, α2, α3, α4 are weight coefficients, f1(P, T) is the objective function of minimizing the switching frequency of the backup power supply, f2(P, U) is the objective function of minimizing the system instability time after the power supply is switched, f3(P, E) is the objective function of minimizing the consumption of the backup power supply, and f4(P, S) is the objective function of maximizing the stability and safety of power supply;

[0067] In a feasible implementation manner, the above-mentioned controlling the output power of different output ports of the above-mentioned multi-way power output unit based on the above-mentioned predicted load information and load priority information includes:

[0068] Define the state St of the system:

[0069] St = {Lt, Pt}

[0070] Wherein, Lt = [L1, L2, ..., Ln] is the load information of each output port at the current time t, Lt = [L1, L2, ..., Ln] is the current output power, n is the number of output ports, the output power of each output port is P, and the load priority of each output port is Prio i, the predicted load information is L i ;

[0071] Define an action set A, where the action set includes: increasing the power of a certain output port, reducing the power of a certain output port, and keeping the current power unchanged;

[0072] Design reward function R t , where the reward function includes load matching reward, priority matching reward and energy consumption penalty;

[0073] The target policy network is optimized based on the deep Q network, where the target policy network is π(A t |S t ), the above target strategy network is in state S t Next select Action A t The probability distribution of

[0074] Each action A t and the corresponding reward R t Stored in the experience replay pool, samples are randomly drawn from it for batch update;

[0075] Update the Q value using the Bellman equation;

[0076] Initialize the above-mentioned deep Q network and target policy network, continuously optimize the strategy through interaction with the environment, select actions according to the current strategy in each iteration, observe new states and rewards after execution, update the network weights, and control the output power of different output ports of the above-mentioned multi-channel power output unit through the learned strategy.

[0077] In a feasible implementation, the above reward function is:

[0078] R t =α1R1+α2R2+α3R3

[0079] The above load matching rewards are:

[0080]

[0081] The above priority matching rewards are:

[0082]

[0083] The above energy consumption penalty is:

[0084]

[0085] Among them, P i is the actual output power of the ith output port, L i is the load demand corresponding to the ith output port, Prioi is the load priority of the i-th output port, R1 is the load matching reward, R2 is the priority matching reward, R3 is the energy consumption penalty, α1, α2, α3 are the weight parameters of each part in the comprehensive reward function, R t is the comprehensive reward function.

[0086] In a feasible implementation manner, the backup power supply unit has multiple power modules, and different power modules use different energy supply methods.

[0087] In summary, the multi-channel power output control system of the edge computing UPS proposed in this application combines real-time monitoring, intelligent prediction, multi-channel power output control and load priority management, and solves many problems existing in traditional power management systems. This application can dynamically monitor the power system status by collecting line power input parameters and real-time load information of each output path in real time through the monitoring unit. Compared with the traditional static redundancy design, this solution can understand the operating status of the system in real time, and make more accurate predictions and responses based on this information, ensuring that it can respond quickly when power fluctuations or anomalies occur, and avoid the impact of power outages. Through the input prediction model and load prediction model in the edge computing unit, this application can perform multi-dimensional intelligent predictions based on line power input parameters and load information respectively. Unlike traditional methods that rely on a single algorithm, this solution uses a variety of advanced machine learning algorithms (such as long short-term memory network LSTM and support vector regression SVR) to improve the accuracy and robustness of the prediction, so that possible power anomalies and load demand changes can be predicted in advance, further improving the stability and reliability of the system. This application can not only predict abnormal conditions of power input in advance, but also intelligently control the working state of the backup power supply based on the prediction results. Compared with the traditional rule-based control method, this solution optimizes the startup and switching timing of the backup power supply through an intelligent algorithm, reduces unnecessary energy consumption, and ensures that the backup power supply can smoothly and efficiently take over the power supply task when the main power supply fails. This application introduces the concept of load priority in the power distribution process, combines load prediction information and priority information, and realizes precise control of multiple power output units through intelligent algorithms (such as reinforcement learning or ant colony optimization algorithm). Compared with the traditional simple power distribution method, this solution can be dynamically adjusted according to the importance and demand of the load, ensuring that the critical load has priority to obtain a stable power supply and maximizing energy utilization efficiency. This application realizes the full process control of power input, load prediction, power distribution and backup power management through the efficient integration of edge computing units. Its modular design makes the system have good scalability, and can flexibly adjust algorithms and control strategies according to actual needs to adapt to complex power requirements in different application scenarios. Compared with traditional power management technology, this solution not only improves the reliability and stability of power supply, but also optimizes the efficiency of energy use, reduces unnecessary power waste, and provides a more intelligent and efficient solution for modern power systems.

[0088] The control method of the multi-power output control system of the edge computing UPS proposed in this application and other advantages, objectives and features of this application will be partially reflected through the following description, and will also be partially understood by technical personnel in this field through research and practice of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present specification. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0090] Figure 1 A flowchart of a control method for a multi-power output control system of an edge computing UPS provided in an embodiment of the present application. DETAILED DESCRIPTION

[0091] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0092] See also Figure 1 , is a flow chart of a control method for a multi-power output control system of an edge computing UPS provided in an embodiment of the present application. The present application proposes a control method for a multi-power output control system of an edge computing UPS. The multi-power supply of the edge computing UPS includes a multi-power output unit, an edge computing unit, a monitoring unit and a backup power supply unit. The method may specifically include:

[0093] The above methods include:

[0094] S110, controlling the monitoring unit to monitor the real-time power input parameters of the line and the real-time load information corresponding to different output paths in real time;

[0095] S120, controlling the input prediction model in the edge computing unit to obtain the line prediction power input parameter according to the line real-time power input parameter and the line historical power input parameter;

[0096] S130, controlling the load prediction model in the edge computing unit to obtain predicted load information according to the real-time load information and historical load information;

[0097] S140, controlling the working state of the backup power supply based on the predicted power input parameter;

[0098] S150. Control the output power of different output ports of the multi-way power output unit based on the predicted load information and the load priority information.

[0099] Exemplarily, the system monitors the real-time status of the power line through the monitoring unit. Specifically, the monitoring unit is responsible for collecting and recording the real-time power input parameters of the current line, such as voltage, current, frequency, phase angle, power factor, total harmonic distortion, instantaneous fluctuation and power failure frequency. At the same time, the monitoring unit also monitors the real-time load information of each output channel (i.e., different output ports of the multi-channel power output unit), which may include the power demand, current demand, load type, load change rate and load working time of each output channel.

[0100] The input prediction model in the edge computing unit generates the predicted power input parameters of the line based on the real-time power input parameters provided by the current monitoring unit and the historical power input parameters collected previously. The input prediction model may use time series analysis, statistical methods, or machine learning algorithms (such as long short-term memory network LSTM and support vector regression SVR) to make predictions. These prediction parameters will be used in subsequent steps to ensure that the system can make corresponding adjustments before power anomalies occur. By predicting future power input conditions, data support is provided for the management of backup power supplies to ensure that the system can respond quickly when power fluctuations or anomalies occur.

[0101] The load prediction model in the edge computing unit will use real-time load information and historical load data to predict the future load demand of each output channel. The model may include convolutional neural networks (CNNs), time series analysis methods, etc., to extract and analyze patterns or trends in load information. Ultimately, the predicted load information will provide a basis for the system to allocate output power. Through accurate prediction of load demand, it is ensured that the system can reasonably allocate the power of multiple power output units to avoid unnecessary energy waste or insufficient load.

[0102] The system decides whether to start or adjust the working state of the backup power supply based on the predicted power input parameters generated by S120. The backup power supply can be UPS, uninterruptible power supply, generator set, etc. When it is predicted that the main power supply may have problems or be unstable, the system will start the backup power supply in advance to ensure the continuity and stability of power supply. By predicting and adjusting the backup power supply status in advance, the impact on the system in the event of power outages or anomalies can be prevented, and the normal power supply of the load can be guaranteed.

[0103] The system intelligently distributes the output power of multiple power output units based on the predicted load information obtained by S130 and the priority information of each load. The load priority information is used to determine the importance of each load to ensure that critical loads are given priority to obtain sufficient power supply. At the same time, the system will distribute power through intelligent algorithms (such as reinforcement learning or ant colony optimization algorithms) to maximize system efficiency and reduce energy consumption. Ensure that high-priority loads can obtain sufficient power under any circumstances, and maximize energy utilization efficiency and system stability by optimizing power distribution.

[0104] In summary, the multi-channel power output control system of the edge computing UPS proposed in this application combines real-time monitoring, intelligent prediction, multi-channel power output control and load priority management, and solves many problems existing in traditional power management systems. This application can dynamically monitor the power system status by collecting line power input parameters and real-time load information of each output path in real time through the monitoring unit. Compared with the traditional static redundancy design, this solution can understand the operating status of the system in real time, and make more accurate predictions and responses based on this information, ensuring that it can respond quickly when power fluctuations or anomalies occur, and avoid the impact of power outages. Through the input prediction model and load prediction model in the edge computing unit, this application can perform multi-dimensional intelligent predictions based on line power input parameters and load information respectively. Unlike traditional methods that rely on a single algorithm, this solution uses a variety of advanced machine learning algorithms (such as long short-term memory network LSTM and support vector regression SVR) to improve the accuracy and robustness of the prediction, so that possible power anomalies and load demand changes can be predicted in advance, further improving the stability and reliability of the system. This application can not only predict abnormal conditions of power input in advance, but also intelligently control the working state of the backup power supply based on the prediction results. Compared with the traditional rule-based control method, this solution optimizes the startup and switching timing of the backup power supply through an intelligent algorithm, reduces unnecessary energy consumption, and ensures that the backup power supply can smoothly and efficiently take over the power supply task when the main power supply fails. This application introduces the concept of load priority in the power distribution process, combines load prediction information and priority information, and realizes precise control of multiple power output units through intelligent algorithms (such as reinforcement learning or ant colony optimization algorithm). Compared with the traditional simple power distribution method, this solution can be dynamically adjusted according to the importance and demand of the load, ensuring that the critical load has priority to obtain a stable power supply and maximizing energy utilization efficiency. This application realizes the full process control of power input, load prediction, power distribution and backup power management through the efficient integration of edge computing units. Its modular design makes the system have good scalability, and can flexibly adjust algorithms and control strategies according to actual needs to adapt to complex power requirements in different application scenarios. Compared with traditional power management technology, this solution not only improves the reliability and stability of power supply, but also optimizes the efficiency of energy use, reduces unnecessary power waste, and provides a more intelligent and efficient solution for modern power systems.

[0105] In some examples, the power input parameters include voltage information, current information, frequency information, phase angle information, power factor information, total harmonic distortion information, instantaneous fluctuation information, and power failure frequency information;

[0106] The above-mentioned input prediction model includes an input layer, a feature extraction module, a prediction module, a fusion module and an output layer;

[0107] The feature extraction module includes a time series feature extractor and a statistical feature extractor. The time feature extractor extracts time series features based on a sliding window method, a differential operation method, and a frequency domain analysis method. The statistical features extracted by the statistical feature extractor include the mean, variance, maximum value, and minimum value.

[0108] The above prediction module includes a target long short-term memory network and a target support vector regression network;

[0109] The fusion module is used to perform weighted fusion on the first output of the target long short-term memory network and the second output of the target support vector regression network to obtain the line prediction power input parameters.

[0110] For example, in the example of this solution, the power input parameters refer to a series of key parameters related to the operating state of the power system. These parameters include:

[0111] Voltage information: The voltage value of the line, used to determine the stability of power transmission. Current information: The current value of the line, used to evaluate the power consumption of the load. Frequency information: The power frequency of the line, used to ensure the synchronization and stability of the power system. Phase angle information: The phase angle between the line voltage and current, used to evaluate the power factor and power quality. Power factor information: Indicates the power factor of the line, used to determine the effective utilization rate of electricity. Total harmonic distortion information: Used to evaluate the degree of distortion of the power waveform in the line and reflect the power quality. Instantaneous fluctuation information: The instantaneous changes in line voltage or current, used to identify sudden power fluctuations. Power outage frequency information: The frequency of power outages on the line within a specific time period, used to evaluate the reliability of power supply.

[0112] The input prediction model is designed to predict future line conditions based on power input parameters. The model includes the following core components:

[0113] 1. Input layer

[0114] The input layer is used to receive the raw data of the above power input parameters to ensure that the model can obtain comprehensive power system status information.

[0115] 2. Feature Extraction Module

[0116] The feature extraction module is a key component of the input prediction model, which is used to extract useful feature information from the raw input data. This module is further divided into two parts:

[0117] Time series feature extractor: Use sliding window method, differential operation and frequency domain analysis method to process time series data, so as to extract time series features that reflect the trend of power input parameters. These features help to identify and predict dynamic changes in the power system.

[0118] Statistical feature extractor: Extracts statistical features from power input parameters, including mean, variance, maximum and minimum values. These statistical features provide overall distribution information of power input parameters, which helps the prediction model capture the steady-state characteristics in the data.

[0119] 3. Prediction Module

[0120] The prediction module is responsible for predicting the line status based on the extracted features. This module consists of two parts:

[0121] Target Long Short-Term Memory Network (LSTM): LSTM network can process long time series data, is good at capturing the long-term dependencies of power input parameters, and generates the first output.

[0122] Target Support Vector Regression Network (SVR): The SVR model is based on a radial basis function kernel and is used to perform accurate regression prediction of power input parameters to generate a second output.

[0123] 4. Fusion Module

[0124] The fusion module is used to perform weighted fusion of the first output of the target long short-term memory network and the second output of the target support vector regression network. This fusion method can make full use of the respective advantages of LSTM and SVR to generate more accurate line prediction power input parameters.

[0125] 5. Output Layer

[0126] The output layer receives the fused results and generates the final line prediction power input parameters. These parameters will be used to control the working state of the backup power supply and the output power of the multi-channel power output unit.

[0127] Through the collaborative work of the above modules, this solution can accurately predict the line status based on the power input parameters, providing strong support for subsequent power control. Compared with traditional prediction methods, the input prediction model in the method proposed in the embodiment of the present application can not only extract deep feature information from the original data, but also improve the prediction accuracy through multi-model fusion, thereby better ensuring the stability and reliability of the power system.

[0128] In some examples, this also includes:

[0129] Inputting the time series features into the target long short-term memory network to obtain the first output;

[0130] The time series features and the statistical features are input into a target support vector regression network to obtain the second output, wherein the target support vector regression network is an SVR model of a radial basis function kernel.

[0131] For example, the first thing to be processed is the time series features extracted from the power input parameters. These time series features reflect the trends and laws of the power input parameters changing over time, such as voltage fluctuations, current changes, frequency changes, etc.

[0132] These time series features are directly input to the target long short-term memory network (LSTM). The LSTM network is a deep learning model that can process and predict time series data. It can effectively capture and learn long-term and short-term dependencies in time series data through built-in memory units and gating mechanisms. The time series features extracted by the time series feature extractor in the input prediction model, such as feature data processed by the sliding window method, differential operation algorithm and frequency domain analysis method, are directly passed to the target long short-term memory network. After receiving these features, the LSTM network processes the information through its hierarchical structure and finally generates the first output. The output of the target long short-term memory network, that is, the first output, represents the prediction result of the power input parameters based on the time series features.

[0133] Unlike the target LSTM network, which processes time series features, the target SVR network receives both time series features and statistical features. The SVR model is particularly suitable for regression problems and can make accurate numerical predictions based on the input feature data. The time series features and the statistical features (including mean, variance, maximum and minimum) extracted by the statistical feature extractor are input to the target SVR network. The target SVR network uses the radial basis function kernel (RBF), which can effectively handle nonlinear regression problems and enable the model to capture complex patterns in the data. The target SVR network generates a second output based on the input time series features and statistical features, which represents the prediction results of the power input parameters after comprehensively considering the time series and statistical features.

[0134] The two outputs from the above process, namely the first output (from the target LSTM network) and the second output (from the target SVR network), are passed to the fusion module.

[0135] The function of the fusion module is to perform weighted fusion of the first output and the second output to obtain more accurate and reliable line prediction power input parameters. In the fusion module, the first output and the second output are weighted according to the preset weight coefficient. In this way, the fusion module can combine the time series processing capability of the LSTM network and the precise regression capability of the SVR model to generate more accurate prediction results. The fused result is passed to the output layer as the final line prediction power input parameter, which is used for the control of the working status of the backup power supply and the power distribution of the multi-way power output unit in the subsequent steps.

[0136] By inputting the time series features into the target long short-term memory network and the target support vector regression network respectively, and weightedly fusing the outputs of the two networks in the subsequent steps, the method proposed in the embodiment of the present application can fully utilize the advantages of the two different models to generate more accurate and reliable power input parameter prediction results. This processing flow ensures that the system can make accurate predictions and adjustments in a timely manner when facing a complex and changeable power input environment, thereby improving the stability and reliability of the system.

[0137] In some examples, the specific steps of determining the first weight corresponding to the first output and the second weight corresponding to the second output include:

[0138] Each particle is represented as a weighted combination of a first weight and a second weight, wherein the particle swarm is randomly initialized in the weight space;

[0139] In each iteration, the velocity and position of the particles are updated so that the particles tend to the global optimal solution and the individual optimal solution;

[0140] Update the fitness of particles according to the prediction error after weighted fusion, and gradually find the global optimal weight combination;

[0141] In the global optimal weight combination, a sliding window is defined, and the prediction error of each model in the window is calculated to update the weight, and the first weight and the second weight updated in real time are obtained;

[0142] The update formula for the first weight w1 and the second weight w2 is:

[0143]

[0144] w1(t) and w2(t) are the first and second weights at time t, w1(t+1) and w2(t+1) are the new weights at time t+1, and α is the learning rate MSE SVR and MSE LSTM is the prediction error of the target support vector regression network and the target long short-term memory network, MSE SVR is the error value of the SVR model at the current moment, MSE LSTM is the error value of the LSTM model at the current moment.

[0145] Exemplarily, first, each particle is represented as a weight combination, that is, the state of the particle is composed of a first weight w1 and a second weight w2. The particle swarm is randomly initialized in the weight space, that is, the initial position of the particle swarm (i.e., the initial weight combination) is randomly selected. The initial velocity of each particle is also randomly set. Random initialization ensures that the particle swarm can search extensively in the weight space, is not limited by the initial conditions, and can cover more possible solutions.

[0146] In each iteration, the particle's velocity and position (i.e. weights w1 and w2) are updated. The particle's update follows the following rules:

[0147] v i (t+1)=w·v i (t)+c1·r1·(p i (t)-x i (t))+C2·r2·(g(t)-x i (t))

[0148] x i (t+1)=x i (t)+v i (t+1)

[0149] Among them, vi(t) is the velocity of particle i at the tth iteration, xi(t) is the position of particle i at the tth iteration (i.e., the current weight combination w1 and w2), pi(t) is the individual optimal solution found by particle i during the iteration, and g(t) is the current global optimal solution. Through the above update rules, the particle swarm will gradually approach the global optimal solution (i.e., the best weight combination). The updated particle position represents the candidate value of weights w1 and w2 at the next moment.

[0150] After each iteration, the fitness of each particle is evaluated based on the prediction error after weighted fusion. Specifically, the first output of the target long short-term memory network (LSTM) and the second output of the target support vector regression network (SVR) are fused according to the current weight combination w1 and w2, and the prediction error after fusion is calculated. The fitness of the particle is updated based on the error value, and the particle swarm will gradually find the optimal weight combination. Through error feedback, the particle swarm optimization algorithm can continuously adjust the weight combination, reduce the prediction error, and improve the overall prediction performance of the model.

[0151] After the particle swarm finds the global optimal weight combination, the sliding window mechanism is introduced on this basis. A sliding window is defined, the prediction error of each model in the window is calculated, and the weight is updated accordingly. The update formula of the new weights w1(t+1) and w2(t+1) is as follows:

[0152]

[0153] w1(t) and w2(t) are the first and second weights at time t, w1(t+1) and w2(t+1) are the new weights at time t+1, and α is the learning rate MSE SVR and MSE LSTM is the prediction error of the target support vector regression network and the target long short-term memory network, MSE SVR is the error value of the SVR model at the current moment, MSE LSTM is the error value of the LSTM model at the current moment.

[0154] Through the sliding window mechanism, the weights w1 and w2 can be adjusted in real time, so that the model can adapt to the dynamically changing environment and continuously optimize the prediction accuracy.

[0155] The method proposed in the embodiment of the present application uses a particle swarm optimization algorithm to dynamically determine the weighted fusion weights of the target long short-term memory network and the target support vector regression network, making the final prediction result more accurate. Combined with the sliding window mechanism, the scheme can adjust the weights in real time to ensure that the model can still maintain high-precision prediction capabilities when the power system environment changes. This weight determination method not only improves the robustness of the system, but also enhances the adaptability and flexibility of the scheme in different application scenarios.

[0156] In some examples, the load information includes power demand, current demand, load type, load change rate, and load working time; the load prediction model includes an input layer, a feature extraction layer, a convolution layer, a prediction recognition layer, a fully connected layer, and an output layer;

[0157] The output of the above convolutional layer is:

[0158] C t =ReLU(W c *x t +b c )

[0159] Among them, * represents the convolution operation, ReLU is the activation function, Ct is the output feature of the convolution layer, x t is the feature data extracted by the feature extraction layer, W c is the convolution kernel weight, b c is bias;

[0160] The calculation formula of the prediction recognition layer is:

[0161] i t =σ(W i C t +U i h t-1 +b i )

[0162] ft =σ(W f C t +U f h t-1 +b f )

[0163] o t =σ(W o C t +U o h t-1 +b o )

[0164] c t =f t ⊙c t-1 +i t ⊙tanh(W c C t +U c h t-1 +b c )

[0165] h t =o t ⊙tanh(c t )

[0166] Among them, i t 、f t , o t They are the states of the input gate, forget gate, and output gate;

[0167] c t is the cell state, h is the hidden state;

[0168] W i , W f , W o , W c is the weight matrix, U i , U f , U o , U c is the weight matrix associated with the hidden state at the previous moment, b i , b f , b o , b c is the bias term;

[0169] σ represents the sigmoid function, ⊙ represents element-by-element multiplication;

[0170] The calculation formula of the fully connected layer is:

[0171] y t =ReLU(W d h t +b d )

[0172] W d is the weight matrix of the fully connected layer, b d is the bias term, y t is the output feature of the fully connected layer;

[0173] The predicted load information obtained by the output layer is

[0174]

[0175] Among them, W o is the weight matrix of the output layer, b o is the bias term, For predicted load information.

[0176] Exemplarily, in the present solution, load information refers to key parameters related to the electrical load in the system, which are used to characterize the specific characteristics and requirements of the load, including: Power demand: indicates the electrical power required by the load. Current demand: indicates the amount of current required by the load. Load type: used to distinguish different types of loads, such as inductive loads, capacitive loads, or resistive loads. Load change rate: indicates the speed or magnitude of change in load demand. Load operating time: indicates the duration or periodicity of load operation. The load prediction model is a multi-layer deep learning model that is used to process the above load information and generate prediction results.

[0177] The model includes the following core components:

[0178] 1. Input layer

[0179] The input layer is used to receive the raw data of load information, including power demand, current demand, load type, load change rate, load working time, etc. The input layer ensures that the model can obtain comprehensive load information for subsequent processing.

[0180] 2. Feature extraction layer

[0181] The feature extraction layer is responsible for extracting key features that can represent the load characteristics from the input data. These features may include some transformations or combinations of the original load information to capture important patterns or trends in the data.

[0182] 3. Convolutional Layer

[0183] The convolution layer is used to perform convolution operations on the features generated by the feature extraction layer to extract high-level feature representations. The convolution operation can identify local patterns in load information, such as the changing pattern of power demand.

[0184] Calculation formula: The output of the convolutional layer is calculated by the following formula:

[0185] Ct =ReLU(W c *x t +b c )

[0186] Among them, * represents the convolution operation, ReLU is the activation function, Ct is the output feature of the convolution layer, x t is the feature data extracted by the feature extraction layer, W c is the convolution kernel weight, b c is bias;

[0187] 4. Prediction and recognition layer

[0188] Prediction and recognition layer The state update formula of the prediction and recognition layer is:

[0189] i t =σ((W i C t +U i h t-1 +b i )

[0190] f t =σ(W f C t +U f h t-1 +b f )

[0191] o t =σ(W o C t +U o h t-1 +b o )

[0192] c t =f t ⊙c t-1 +i t ⊙tanh(W c C t +U c h t-1 +b c )

[0193] h t =o t ⊙tanh(c t )

[0194] Among them, i t 、f t , o t They are the states of the input gate, forget gate, and output gate;

[0195] c t is the cell state, h is the hidden state;

[0196] W i , W f , W o , W c is the weight matrix, U i , U f , U o , U c is the weight matrix associated with the hidden state at the previous moment, b i , b f , b o , b c is the bias term;

[0197] σ represents the sigmoid function, ⊙ represents element-by-element multiplication;

[0198] 5. Fully connected layer

[0199] The fully connected layer is used to further process the output of the prediction and recognition layer, integrate multi-dimensional feature information, and prepare for the final prediction output.

[0200] Calculation formula: The output of the fully connected layer is calculated by the following formula:

[0201] The calculation formula of the fully connected layer is:

[0202] y t =ReLU(W d h t +b d )

[0203] W d is the weight matrix of the fully connected layer, b d is the bias term, yt is the output feature of the fully connected layer;

[0204] 6. Output Layer

[0205] The output layer receives the results of the fully connected layer and generates the final load prediction information. The output represents the prediction results of key indicators such as load power demand and current demand.

[0206] Calculation formula: The prediction result of the output layer is calculated by the following formula: The predicted load information of the output layer is

[0207]

[0208] Among them, W o is the weight matrix of the output layer, b o is the bias term, For predicted load information.

[0209] The method proposed in the embodiment of the present application can accurately extract features from load information and make predictions by utilizing deep learning models, especially convolutional layers and long short-term memory networks. The convolutional layer is used to extract local features, while the prediction recognition layer captures long-term and short-term dependencies in time series through the LSTM structure, thereby generating more accurate load prediction results. The design of the fully connected layer and the output layer further ensures the accuracy and applicability of the prediction information, and ultimately achieves comprehensive prediction and intelligent management of system load demand.

[0210] In some examples, controlling the working state of the backup power supply based on the predicted power input parameter includes:

[0211] Determine the optimization objective function;

[0212] Randomly initialize the particle swarm, where each particle represents a possible solution, and each possible solution includes position and velocity;

[0213] Encode the problem’s decision variables as particle positions to represent the state of the solution in the search space;

[0214] Update the velocity and position of each particle based on the following formula:

[0215] v i (t+1)=w·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(g(t)-x i (t))

[0216] x i (t+1)=x i (t)+v i (t+1)

[0217] Among them, w is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers used to control the movement and search behavior of particles.

[0218] Predict the fitness of particles based on the target deep learning model;

[0219] The first k particles with the highest fitness are retained to enter the next generation, and the remaining particles are reinitialized or mutated according to probability;

[0220] Calculate the objective function value for each particle's position and evaluate its quality through the fitness function;

[0221] When the maximum number of iterations is reached or the fitness function converges, the algorithm is terminated and the backup power supply control parameters are output;

[0222] The working state of the backup power supply is controlled based on the backup power supply control parameter.

[0223] For example, the optimization objective function is the core of the entire control strategy, defining multiple optimization objectives that the system needs to achieve. The optimization objective function usually contains several different sub-objectives, such as minimizing the frequency of backup power switching, minimizing the system instability time after power switching, minimizing the energy consumption of the backup power supply, and maximizing the stability and safety of power supply. The weighted combination of these sub-objectives forms the final optimization objective function.

[0224] In the initial stage of the algorithm, the particle swarm is randomly initialized. Each particle represents a possible solution, and each solution consists of the particle's position and velocity. The particle's position represents the current solution, i.e., the backup power control parameter, while the velocity represents the particle's movement direction and amplitude in the search space. Through random initialization, the particle swarm can cover the entire search space, ensuring that the algorithm starts searching from a variety of solutions, and increasing the probability of finding the global optimal solution.

[0225] The decision variables of the problem are encoded as the positions of particles. Specifically, the control parameters of the backup power supply (such as switching time, output power, etc.) are encoded as the positions of particles. This allows the position of the particle to intuitively represent the corresponding solution state in the search space. This encoding method can transform complex control problems into particle position searches in a multidimensional search space, making it easier to use the PSO algorithm for optimization.

[0226] In each iteration, the velocity and position of each particle can be updated based on the following formula:

[0227] v i (t+1)=w·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(g(t)-x i (t))

[0228] x i (t+1)=x i (t)+v i (t+1)

[0229] Among them: w is the inertia weight, which controls the moving inertia of the particle.

[0230] c1 and c2 are acceleration constants, which adjust the speed of particles moving to the individual optimal position and the global optimal position respectively.

[0231] r1 and r2 are random numbers used to introduce randomness so that the particle swarm can widely explore different solution spaces during the search process.

[0232] Vi (t) is the velocity of particle i in the tth iteration.

[0233] x i (t) is the position of particle i in the tth iteration.

[0234] By updating the speed and position, particles gradually tend to the individual optimal solution and the global optimal solution, thereby improving the search efficiency.

[0235] The fitness of each particle is predicted based on the target deep learning model. The fitness function is used to evaluate the quality of the solution represented by the particle. The higher the fitness value, the better the quality of the solution. Through fitness evaluation, the algorithm can identify and retain the optimal solution, while eliminating suboptimal solutions, so that the particle swarm gradually tends to the global optimum.

[0236] The first k particles with the highest fitness are retained to enter the next generation, and the remaining particles are reinitialized or mutated according to a certain probability. This operation can generate new solutions for subsequent searches by randomly adjusting the position or speed. Through screening and mutation, the algorithm can avoid falling into local optimality and improve global search capabilities.

[0237] The objective function value is calculated for the position of each particle, and its quality is evaluated by the fitness function. When the maximum number of iterations is reached or the fitness function converges, the algorithm is terminated and the optimal backup power control parameters are output. The termination condition of the method ensures that the search process is completed within a reasonable time and provides a feasible optimal control solution.

[0238] Based on the optimized backup power control parameters, the working state of the backup power is actually controlled. This includes when to switch to the backup power supply, how to adjust the output power of the backup power supply, etc., to achieve the optimal operation of the power system. Through precise control operations, stable and efficient power supply is ensured in a complex power system environment, minimizing the power switching frequency, unstable time and energy consumption, while ensuring the safety of power supply.

[0239] In some examples, the above optimization objective function is:

[0240] f total (P, T, E, S) = α1f1 (P, T) + α2f2 (P, U) + α3f3 (P, E) - α4F4 (P, S)

[0241] Among them, α1, α2, α3, α4 are weight coefficients, f1(P, T) is the objective function to minimize the switching frequency of the backup power supply, and f 2( f3(P, E) is the objective function for minimizing the system instability time after power switching, f4(P, S) is the objective function for maximizing power supply stability and safety.

[0242] In some examples, controlling the output power of different output ports of the multi-way power output unit based on the predicted load information and the load priority information includes:

[0243] Define the state St of the system:

[0244] St = {Lt, Pt}

[0245] Wherein, Lt = [L1, L2, ..., Ln] is the load information of each output port at the current time t, Lt = [L1, L2, ..., Ln] is the current output power, n is the number of output ports, the output power of each output port is P, and the load priority of each output port is Prio i , the predicted load information is L i ;

[0246] Define an action set A, where the action set includes: increasing the power of a certain output port, reducing the power of a certain output port, and keeping the current power unchanged;

[0247] Design reward function R t , where the reward function includes load matching reward, priority matching reward and energy consumption penalty;

[0248] The target policy network is optimized based on the deep Q network, where the target policy network is π(A t ∣S t ), the above target strategy network is in state S t Next select Action A t The probability distribution of

[0249] Each action A t and the corresponding reward R t Stored in the experience replay pool, samples are randomly drawn from it for batch update;

[0250] Update the Q value using the Bellman equation;

[0251] Initialize the above-mentioned deep Q network and target policy network, continuously optimize the strategy through interaction with the environment, select actions according to the current strategy in each iteration, observe new states and rewards after execution, update the network weights, and control the output power of different output ports of the above-mentioned multi-channel power output unit through the learned strategy.

[0252] Exemplarily, the state St of the system is defined as follows: At each time t, the state St of the system consists of the current load information and output power of each output port. Specifically:

[0253] St={Lt,Pt}

[0254] in:

[0255] Lt=[L1, L2, …, Ln] represents the load information of each output port at the current time ttt, and Li represents the load demand of the i-th output port.

[0256] Pt=[P1, P2,…, Pn] represents the current output power of each output port.

[0257] n is the number of output ports, and the output power of each output port is Pi.

[0258] By defining the system state St, the operation status of the system can be accurately described at each moment and provide basic information for subsequent decision-making processes.

[0259] Define action set A: Action set A contains all possible operations that the system can perform at each time t. Specifically, the action set includes the following three operations: 1. Increase the power of a certain output port; 2. Reduce the power of a certain output port; 3. Keep the current power unchanged.

[0260] The action set A defines the actions that the system can take in different states, so that decisions can be made based on the current system state St.

[0261] Design reward function Rt: The reward function Rt is used to evaluate the effect of each action At. The design of the reward function combines load matching reward, priority matching reward and energy consumption penalty.

[0262] Optimizing the target policy network based on the deep Q network: The deep Q network (DQN) optimizes the target policy network π(At|St) through interaction with the environment. The network describes the probability distribution of selecting an action At in the state StS_tSt. Through continuous training, DQN can learn the optimal strategy to maximize the probability of selecting the optimal action in a given state.

[0263] The Deep Q Network is optimized in the following ways:

[0264] Each action At and the corresponding reward Rt are stored in the experience replay pool. The experience replay pool records all actions and corresponding feedback in the system history. Randomly extract samples from the experience replay pool for batch update to avoid the correlation between samples affecting the learning effect.

[0265] The Q-value is updated using the Bellman equation to assess the long-term value of each state-action pair:

[0266]

[0267] Among them, γ is the discount factor, which measures the impact of future rewards on current decisions.

[0268] Initialize the deep Q network and target policy network, and continuously optimize the policy through interaction with the environment. In each iteration, the system selects an action At according to the current policy, observes the new state St+1 and reward Rt after execution, and updates the network weights.

[0269] The specific strategy optimization process is as follows:

[0270] The system first selects an action At through the current policy network π(At|St). After executing the action, the system enters the new state St+1 and obtains the corresponding reward Rt.

[0271] This experience (St, At, Rt, St+1) is stored in the experience replay pool. A batch of experiences are randomly selected and used to update the Q network and adjust the parameters of the target policy network to make it closer to the optimal policy. Through this process, the deep Q network gradually learns how to choose the best action in different states to maximize the cumulative reward.

[0272] Through the learned strategy, the system can intelligently control the output power of different output ports of the multi-way power output unit. Specifically, the system will determine the power allocation of each output port based on the current state St and the learned optimal strategy π*(At|St.

[0273] The ultimate goal is to ensure that the system can operate efficiently and stably through the optimal strategy, while meeting the load requirements of each output port, especially the needs of high-priority loads.

[0274] This solution combines the design of deep Q network and reward function to realize the output power control of multiple power sources based on predicted load information and load priority information. Through the definition of system state, the construction of action set and the optimization of strategy network, it can finally achieve accurate and efficient power allocation in complex power environment and meet the multiple objectives of the system.

[0275] In some examples, the above reward function is:

[0276] R t =α1R1+α2R2+α3R3

[0277] The above load matching rewards are:

[0278]

[0279] The above priority matching rewards are:

[0280]

[0281] The above energy consumption penalty is:

[0282]

[0283] Among them, P i is the actual output power of the ith output port, L i is the load demand corresponding to the ith output port, Prio i is the load priority of the i-th output port, R1 is the load matching reward, R2 is the priority matching reward, R3 is the energy consumption penalty, α1, α2, α3 are the weight parameters of each part in the comprehensive reward function, and Rt is the comprehensive reward function.

[0284] Exemplarily, the embodiment of the present application designs a comprehensive reward function R t , systematically balancing the needs of load matching, priority matching, and energy consumption control. The load matching reward R1 ensures that the system meets the load requirements of each output port as much as possible, the priority matching reward R2 prioritizes the needs of high-priority loads, and the energy consumption penalty R3 suppresses unnecessary energy consumption. By adjusting the weight parameters α1, α2, and α3, it can flexibly adapt to the optimization goals in different scenarios and ensure that the system can achieve efficient and stable operation in a complex power environment.

[0285] In some examples, the backup power supply unit has multiple power modules, and different power modules use different energy supply methods.

[0286] Exemplarily, the backup power supply unit is composed of multiple independent power modules, each of which can work independently and work in conjunction with other modules when needed. This modular design improves the flexibility and reliability of the system because different modules can be combined, scheduled or replaced according to needs. The modular design allows the system to select the appropriate power module for power supply according to real-time needs and environmental conditions, thereby improving energy utilization efficiency and enhancing the system's fault tolerance.

[0287] Different power modules can use different energy supply methods to adapt to different power demand scenarios. Specifically, these energy supply methods may include but are not limited to: battery energy storage module, diesel power generation module, solar power generation module, wind power generation module and hydrogen fuel cell module.

[0288] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A control method for a multi-channel power output control system of an edge computing UPS, characterized in that: The multi-way power supply of the edge computing UPS includes a multi-way power supply output unit, an edge computing unit, a monitoring unit and a backup power supply unit. The method comprises: Control the monitoring unit to monitor the real-time power input parameters of the line and the real-time load information corresponding to different output paths in real time; Control the input prediction model in the edge computing unit to obtain the line prediction power input parameter according to the line real-time power input parameter and the line historical power input parameter; Controlling the load prediction model in the edge computing unit to obtain predicted load information according to the real-time load information and the historical load information; controlling the operating state of the backup power supply based on the predicted power input parameter; Controlling the output power of different output ports of the multi-way power output unit based on the predicted load information and the load priority information; The power input parameters include voltage information, current information, frequency information, phase angle information, power factor information, total harmonic distortion information, instantaneous fluctuation information and power failure frequency information; The input prediction model includes an input layer, a feature extraction module, a prediction module, a fusion module and an output layer; The feature extraction module includes a time series feature extractor and a statistical feature extractor. The time feature extractor extracts time series features based on a sliding window method, a differential operation method, and a frequency domain analysis method. The statistical features extracted by the statistical feature extractor include extracting mean, variance, maximum value, and minimum value. The prediction module includes a target long short-term memory network and a target support vector regression network; The fusion module is used to perform weighted fusion on the first output of the target long short-term memory network and the second output of the target support vector regression network to obtain the line prediction power input parameter; The specific steps of determining the first weight corresponding to the first output and the second weight corresponding to the second output include: Each particle is represented as a weighted combination of a first weight and a second weight, wherein the particle swarm is randomly initialized in the weight space; In each iteration, the velocity and position of the particles are updated so that the particles tend to the global optimal solution and the individual optimal solution; Update the fitness of particles according to the prediction error after weighted fusion, and gradually find the global optimal weight combination; In the global optimal weight combination, a sliding window is defined, and the prediction error of each model in the window is calculated to update the weight, and the first weight and the second weight updated in real time are obtained; The update formula for the first weight w1 and the second weight w2 is: w1(t) and w2(t) are the first and second weights at time t, w1(t+1) and w2(t+1) are the new weights at time t+1, and α is the learning rate MSE SVR and MSE LSTM is the prediction error of the target support vector regression network and the target long short-term memory network, MSE SVR is the error value of the SVR model at the current moment, MSE LSTM is the error value of the LSTM model at the current moment.

2. The control method according to claim 1, characterized in that: Also includes: Inputting the time series feature into the target long short-term memory network to obtain the first output; The time series features and the statistical features are input into a target support vector regression network to obtain the second output, wherein the target support vector regression network is an SVR model of a radial basis function kernel.

3. The control method according to claim 1, characterized in that: The load information includes power demand, current demand, load type, load change rate and load working time; the load prediction model includes an input layer, a feature extraction layer, a convolution layer, a prediction recognition layer, a fully connected layer and an output layer; The convolutional layer output is: C t =ReU(W c *x t +b c ) Among them, * represents the convolution operation, ReLU is the activation function, Ct is the output feature of the convolution layer, x t is the feature data extracted by the feature extraction layer, W c is the convolution kernel weight, b c is bias; The calculation formula of the prediction recognition layer is: i t =σ(W i C t +U i h t-1 +b i ) f t =σ(W f C t +U f h t-1 +b f ) o t =σ(W o C t +U o h t-1 +b o ) c t =f t ⊙c t-1 +i i ⊙tanh(W c C t +U c h t-1 +b c ) h t =o t ⊙tanh(c t ) Among them, i t 、f t , o t They are the states of the input gate, forget gate, and output gate; c t is the cell state, h is the hidden state; W i , W f , W o , W c is the weight matrix, U i , U f , U o , U c is the weight matrix associated with the hidden state at the previous moment, b i , b f , b o , b c is the bias term; σ represents the sigmoid function, ⊙ represents element-by-element multiplication; The calculation formula of the fully connected layer is: and t =ReLU(W d h t +b d ) W d is the weight matrix of the fully connected layer, b d is the bias term, y t is the output feature of the fully connected layer; The predicted load information obtained by the output layer is Among them, W o is the weight matrix of the output layer, b o is the bias term, For predicted load information.

4. The control method according to claim 1, characterized in that: The controlling the working state of the backup power supply based on the predicted power input parameter comprises: Determine the optimization objective function; Randomly initialize the particle swarm, where each particle represents a possible solution, and each possible solution includes position and velocity; Encode the problem’s decision variables as particle positions to represent the state of the solution in the search space; Update the velocity and position of each particle based on the following formula: υ i (t+1)=w·υ i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(g(t)-x i (t)) x i (t+1)=x i (t)+υ i (t+1) Among them, w is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers used to control the movement and search behavior of particles; Predict the fitness of particles based on the target deep learning model; The first seven particles with the highest fitness are retained to enter the next generation, and the remaining particles are reinitialized or mutated according to probability; Calculate the objective function value for each particle's position and evaluate its quality through the fitness function; When the maximum number of iterations is reached or the fitness function converges, the algorithm is terminated and the backup power supply control parameters are output; The working state of the backup power supply is controlled based on the backup power supply control parameter.

5. The control method according to claim 4, characterized in that: The optimization objective function is: f total (P,T,E,S)=α1f1(P,T)+α2f2(P,U)+α3f3(P,E)-α4f4P,S) Among them, α1, α2, α3, α4 are weight coefficients, f1(P, T) is the objective function of minimizing the switching frequency of the backup power supply, f2(P, U) is the objective function of minimizing the system instability time after power switching, f3(P, E) is the objective function of minimizing the consumption of the backup power supply, and f4(P, S) is the objective function of maximizing the stability and safety of power supply.

6. The control method according to claim 1, characterized in that: The controlling the output power of different output ports of the multi-way power output unit based on the predicted load information and the load priority information includes: Define the state St of the system: St = {Lt, Pt} Wherein, Lt = [L1, L2, ..., Ln] is the load information of each output port at the current time t, Lt = [L1, L2, ..., Ln] is the current output power, n is the number of output ports, the output power of each output port is P, and the load priority of each output port is Prio i , the predicted load information is L i ; Define an action set A, where the action set includes: increasing the power of a certain output port, reducing the power of a certain output port, and keeping the current power unchanged; Design reward function R t , where the reward function includes load matching reward, priority matching reward and energy consumption penalty; The target policy network is optimized based on the deep Q network, where the target policy network is π(A t |S t ), the target strategy network is in state S t Next select Action A t The probability distribution of Each action A t and the corresponding reward R t Stored in the experience replay pool, samples are randomly drawn from it for batch update; Update the Q value using the Bellman equation; Initialize the deep Q network and the target policy network, continuously optimize the policy through interaction with the environment, select actions according to the current policy in each iteration, observe new states and rewards after execution, update the network weights, and control the output power of different output ports of the multi-channel power output unit through the learned policy.

7. The control method according to claim 6, characterized in that: The reward function is: R t =α1R1+α2R2+α3R3 The load matching reward is: The priority matching reward is: The energy consumption penalty is: Among them, P i is the actual output power of the ith output port, L i is the load demand corresponding to the ith output port, Prio i is the load priority of the i-th output port, R1 is the load matching reward, R2 is the priority matching reward, R3 is the energy consumption penalty, α1, α2, α3 are the weight parameters of each part in the comprehensive reward function, R t is the comprehensive reward function.

8. The control method according to claim 1, characterized in that: The backup power supply unit has multiple power modules, and different power modules use different energy supply methods.

Images