Control method and system of multi-source intelligent power manager based on 5G communication

Abstract

The invention discloses a control method and system of a multi-source intelligent power manager based on 5G communication, and the method comprises the steps: obtaining multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology through employing a multi-scale pulse fusion tensor decomposition algorithm, and generating a power state joint tensor with time-space alignment; inputting the power supply state joint tensor into a physically constrained adversarial prediction network, and outputting an energy dynamic balance vector including power supply margin prediction; performing multi-target dynamic game optimization on the energy dynamic balance vector, and outputting an anti-interference multi-source cooperative power supply strategy matrix; and inputting the multi-source cooperative power supply strategy matrix into a distributed cooperative control framework driven by edge computing, and finally outputting a multi-source intelligent control instruction set meeting low delay and high reliability. According to the embodiment of the invention, the flexibility and response speed of power management can be improved, and the stability and reliability of power supply are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of power control, and particularly relates to a control method and system for a multi-source intelligent power manager based on 5G communication. Background Art

[0002] Traditional power management methods often rely on static control strategies and single power scheduling mechanisms, making it difficult to fully utilize the advantages of various power sources in the system and unable to adapt to real-time changing load demands. This leads to problems such as low power utilization efficiency, unstable power supply, and possible frequent power outages. Summary of the Invention

[0003] The object of the present invention is to provide a control method and system for a multi-source intelligent power manager based on 5G communication to solve the deficiencies in the prior art, improve the flexibility and response speed of power management, and ensure the stability and reliability of power supply.

[0004] An embodiment of the present application provides a control method for a multi-source intelligent power manager based on 5G communication, the method comprising: Obtaining multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminating multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generating a spatio-temporally aligned power state joint tensor; Inputting the power state joint tensor into a physically constrained adversarial prediction network, reconstructing a multi-source collaborative power supply current pattern based on a dynamic game framework, separating load mutation and steady-state fluctuation characteristics through a residual-focusing frequency-domain decomposition algorithm, and outputting an energy dynamic balance vector including power supply margin prediction; Performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a candidate set of power supply strategies using a Monte Carlo tree search-driven deep policy network, screening the global optimal solution through a Nash equilibrium solver with manifold projection constraints, and outputting an anti-interference multi-source collaborative power supply strategy matrix; Inputting the multi-source collaborative power supply strategy matrix into an edge computing-driven distributed collaborative control framework, optimizing strategy execution parameters based on a real-time feedback mechanism verified by digital twin, coordinating multi-node control instructions through a 5G multi-hop transmission protocol with dynamic weight allocation, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability.

[0005] Optionally, the obtaining multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminating multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generating a spatio-temporally aligned power state joint tensor includes: Obtain the multi-source heterogeneous power data of the multi-source intelligent power manager. According to the original time-domain signals of the photovoltaic array output waveform and the charge-discharge curve of the energy storage battery in the data, use the dynamic threshold gating of the spiking neural network for pulse coding, convert the continuous waveform into a pulse trigger sequence, and generate a pulse spatio-temporal matrix with time resolution; wherein, the dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism and retains the effective energy fluctuation characteristics; Construct a multi-source signal phase difference metric tensor based on the pulse spatio-temporal matrix and the environmental temperature field distribution data, and perform cross-modal time calibration through a phase synchronization algorithm with pulse timing constraints to eliminate the phase drift caused by photovoltaic transient fluctuations and battery aging, and output a phase-aligned multi-scale pulse tensor; Input the multi-scale pulse tensor into the multi-scale pulse fusion tensor decomposition algorithm, construct a pulse energy density spectrum in the time-frequency hybrid domain, decompose the core feature sub-tensors of photovoltaic, energy storage, and load through a time-frequency constraint mechanism with convolutional kernel parameterization, and suppress time-frequency aliasing interference; Perform pulse energy accumulation and spatial topology mapping on the core feature sub-tensors, construct a multi-source energy dynamic topology map based on a dynamic weight assignment algorithm triggered by pulses, and generate a spatio-temporally aligned power state joint tensor; wherein, the topology map reflects the dynamic coupling relationship between energy nodes in real time through an edge weight update mechanism triggered by pulses.

[0006] Optionally, input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply current pattern based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a frequency-domain decomposition algorithm with residual focusing, and output an energy dynamic balance vector including power supply margin prediction, including: Input the power state joint tensor into a physically constrained adversarial prediction network, construct a latent representation of the multi-source power supply current pattern through a generator network, and impose a frequency-domain sparsity constraint using a discriminator network to generate an adversarial feature vector; wherein, the discriminator uses a Fourier-domain adversarial constraint to force the generator to separate steady-state and transient characteristics; Perform frequency-domain sparse residual focusing processing on the adversarial feature vector, extract the load mutation residual component through a band-pass filter kernel with an adaptive bandwidth, and use a steady-state fluctuation suppression algorithm to eliminate background noise, and output a high-resolution load mutation feature spectrum; Construct a game payoff matrix for load mutation and steady-state fluctuation based on a dynamic game framework, and dynamically allocate the attention weights of the two types of features through an adversarial attention mechanism, wherein the adversarial attention mechanism optimizes the sparsity of the attention mask through a game strategy gradient update mechanism; Perform time-domain integration and energy accumulation on the load mutation feature spectrum, and combine the state-of-charge constraint of the energy storage battery. Generate an energy dynamic balance vector containing the power supply capacity for the next several seconds through a power supply margin prediction model. Among them, the power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

[0007] Optionally, perform multi-objective dynamic game optimization on the energy dynamic balance vector. Use a deep policy network driven by Monte Carlo tree search to generate a candidate set of power supply strategies. Screen the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and output an anti-interference multi-source collaborative power supply strategy matrix, including: According to the energy dynamic balance vector, use a deep policy network driven by Monte Carlo tree search to generate a candidate tree of power supply strategies. Among them, the deep policy network synchronously evaluates the short-term benefits and long-term stability of the strategy through a policy-value dual-head network. Perform a high-dimensional policy space manifold projection on the candidate tree of power supply strategies. Extract the core features of the strategy through a dimensionality reduction algorithm constrained by geodesic distance to generate a low-dimensional compact policy manifold. Among them, the dimensionality reduction algorithm suppresses strategy conflicts through manifold curvature optimization. Construct a multi-objective dynamic game model on the low-dimensional compact policy manifold. Use a Nash equilibrium solver with hybrid strategy optimization to conduct strategy benefit games. Screen the global optimal solution through the shadow price iteration algorithm and output an anti-interference multi-source collaborative power supply strategy matrix. Verify the convergence of the hybrid strategy in dynamic game theory for the multi-source collaborative power supply strategy matrix. Prove the superlinear convergence rate under multi-objective conflicts through Lyapunov stability analysis, and dynamically adjust the strategy search depth based on the verification results.

[0008] Optionally, input the multi-source collaborative power supply strategy matrix into a distributed collaborative control framework driven by edge computing. Optimize the strategy execution parameters based on a real-time feedback mechanism verified by digital twin. Coordinate multi-node control instructions through a 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability, including: Input the multi-source collaborative power supply strategy matrix into a digital twin verification module driven by edge computing. Simulate battery thermal runaway and photovoltaic shading mutations through a fault scenario adversarial distillation technology to generate a strategy robustness evaluation vector. Among them, the digital twin verification module optimizes the simulation parameters through a dynamic weight mirror update mechanism. According to the strategy robustness evaluation vector, construct a dynamic weight allocation model for the 5G multi-hop transmission protocol. Adjust the priority weights of control instructions in real time based on the base station load rate and channel quality, and generate an anti-congestion instruction distribution sequence. Perform multi - base - station collaborative control conflict detection on the instruction distribution sequence, and adopt an implicit gradient compensation algorithm to eliminate policy execution deviation. Among them, the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence through the backpropagation of conflict residuals; Perform timestamp synchronization on the corrected control instructions through the time - delay equalization technology of multi - base - station collaboration, and adopt an instruction buffering mechanism with a sliding - window constraint to ensure that the end - to - end delay is less than 1 ms, and finally output a multi - source intelligent control instruction set with low delay and high reliability.

[0009] Another embodiment of this application provides a control system for a multi - source intelligent power manager based on 5G communication. The system includes: An acquisition module, which is used to acquire multi - source heterogeneous power data of the multi - source intelligent power manager, construct a dynamic energy topology by using a multi - scale pulse fusion tensor decomposition algorithm, eliminate multi - source signal interference through a phase - synchronous time - frequency constraint mechanism, and generate a power - state joint tensor with spatio - temporal alignment; A separation module, which is used to input the power - state joint tensor into a physically - constrained adversarial prediction network, reconstruct a multi - source collaborative power supply manifold based on a dynamic game framework, separate load mutation and steady - state fluctuation characteristics through a residual - focused frequency - domain decomposition algorithm, and output an energy dynamic balance vector including power - supply margin prediction; An optimization module, which is used to perform multi - objective dynamic game optimization on the energy dynamic balance vector, generate a candidate set of power - supply strategies by using a deep policy network driven by Monte Carlo tree search, screen the global optimal solution through a Nash equilibrium solver with manifold - projection constraints, and output an anti - interference multi - source collaborative power - supply strategy matrix; An output module, which is used to input the multi - source collaborative power - supply strategy matrix into a distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on a real - time feedback mechanism verified by digital twin, coordinate multi - node control instructions through a 5G multi - hop transmission protocol with dynamic weight allocation, and finally output a multi - source intelligent control instruction set that meets low delay and high reliability.

[0010] Another embodiment of this application provides a storage medium, in which a computer program is stored. Among them, the computer program is set to execute the method described in any one of the above when running.

[0011] Another embodiment of this application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0012] Compared with the prior art, a control method of a multi-source intelligent power manager based on 5G communication provided by the present invention obtains multi-source heterogeneous power data of the multi-source intelligent power manager, constructs a dynamic energy topology by using a multi-scale pulse fusion tensor decomposition algorithm, and generates a power state joint tensor with spatio-temporal alignment; inputs the power state joint tensor into a physically constrained adversarial prediction network, and outputs an energy dynamic balance vector including power supply margin prediction; performs multi-objective dynamic game optimization on the energy dynamic balance vector, and outputs an anti-interference multi-source collaborative power supply strategy matrix; inputs the multi-source collaborative power supply strategy matrix into an edge computing-driven distributed collaborative control framework, and finally outputs a multi-source intelligent control instruction set that meets low latency and high reliability, so as to improve the flexibility and response speed of power management and ensure the stability and reliability of power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The hardware structure block diagram of a computer terminal for a control method of a multi-source intelligent power manager based on 5G communication provided by an embodiment of the present invention; Figure 2 The flow schematic diagram of a control method of a multi-source intelligent power manager based on 5G communication provided by an embodiment of the present invention; Figure 3 The structure schematic diagram of a control system of a multi-source intelligent power manager based on 5G communication provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be construed as a limitation of the present invention.

[0015] An embodiment of the present invention first provides a control method of a multi-source intelligent power manager based on 5G communication, and this method can be applied to electronic devices, such as computer terminals, specifically ordinary computers, etc.

[0016] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 The hardware structure block diagram of a computer terminal for a control method of a multi-source intelligent power manager based on 5G communication provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0017] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions, and when the program instructions are executed, the processor can execute any control method of a multi-source intelligent power manager based on 5G communication.

[0018] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0019] The internal memory provides an environment for the operation of a computer program in a non-volatile storage medium. When the computer program is executed by the processor, the processor can be caused to execute any control method of a multi-source intelligent power manager based on 5G communication.

[0020] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in

[0021] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0022] See Figure 2 , an embodiment of the present invention provides a control method for a multi-source intelligent power manager based on 5G communication, which may include the following steps: S201, obtain multi-source heterogeneous power data of the multi-source intelligent power manager, construct a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminate multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generate a spatio-temporal aligned power state joint tensor; Specifically, multi-source heterogeneous power data of the multi-source intelligent power manager can be obtained. According to the original time-domain signals of the output waveform of the photovoltaic array and the charge and discharge curve of the energy storage battery in the data, pulse coding is performed using the dynamic threshold gating of the pulse neural network to convert the continuous waveform into a pulse trigger sequence, and a pulse spatio-temporal matrix with time resolution is generated; wherein, the dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism and retains the characteristics of effective energy fluctuations; Multi-source heterogeneous power data includes voltage / current waveforms of photovoltaic arrays, charge / discharge curves of energy storage batteries (such as SOC status, charge / discharge rate), load power demands, etc. To achieve efficient processing, continuous signals are first pulse-coded through a Spiking Neural Network (SNN). The input layer of the SNN consists of multiple neurons, and each neuron corresponds to a power parameter (such as photovoltaic voltage, battery current).

[0023] The core of dynamic threshold gating is to adaptively adjust the pulse triggering threshold to distinguish effective fluctuations from noise. The specific process is as follows: Baseline threshold initialization: Based on historical data statistics, set the initial threshold (such as the photovoltaic voltage threshold is ±5% of the nominal value).

[0024] Real-time threshold adjustment: Use a sliding window (window width 200ms) to calculate the signal standard deviation σ, and dynamically adjust the threshold to μ ± 3σ (μ is the window mean). For example, when the photovoltaic array undergoes transient fluctuations due to cloud cover, the voltage standard deviation increases to 8%, and the threshold is automatically relaxed to ±24% to avoid false triggering.

[0025] Pulse triggering rule: When the signal change rate exceeds the threshold, generate a pulse (encoded as 1), otherwise remain silent (encoded as 0). For example, when the battery current suddenly increases from 10A to 50A within 2 seconds (change rate 20A / s), a high-frequency pulse sequence is triggered.

[0026] The generated pulse spatio-temporal matrix has dimensions N × T, where N is the number of power parameters (such as 3 parameters for each of photovoltaic, battery, and load), and T is the time step (resolution 1ms). For example, during photovoltaic transient fluctuations, the pulse density in the corresponding row of the matrix increases significantly, reflecting the energy fluctuation characteristics.

[0027] According to the pulse spatio-temporal matrix and the environmental temperature field distribution data, construct a multi-source signal phase difference metric tensor, and perform cross-modal time calibration through a phase synchronization algorithm with pulse timing constraints to eliminate phase drift caused by photovoltaic transient fluctuations and battery aging, and output a phase-aligned multi-scale pulse tensor; Environmental temperature field data is collected by distributed temperature sensors (such as DS18B20), with a spatial resolution of 0.5m × 0.5m grid and a time synchronization accuracy of ±10ms. The steps for constructing the phase difference metric tensor are as follows: Multi-modal timestamp alignment: Use the precise clock synchronization protocol of the 5G network (such as IEEE 1588 PTP) to align the timestamps of temperature data and the power pulse matrix to the microsecond level.

[0028] Phase difference calculation: Mutual information analysis is performed on the correlation between each power parameter and the temperature field to construct a three-dimensional tensor (parameter × temperature grid × time lag). For example, the mutual information between the photovoltaic voltage and the temperature field reaches a peak at a time lag of 300 ms, indicating the delayed effect of temperature changes on photovoltaic output.

[0029] Pulse timing constraint: Align pulse sequences of different modalities through the dynamic time warping (DTW) algorithm. For example, battery aging causes an increase in its response delay by 50 ms, and DTW eliminates the phase difference by stretching / compressing the time axis.

[0030] The phase synchronization algorithm uses wavelet coherence analysis: Wavelet transform: Perform continuous wavelet transform (using the Mexican Hat wavelet) on the pulse matrix and temperature field data respectively to generate the time-frequency energy spectrum.

[0031] Coherence detection: Calculate the cross-wavelet coherence spectrum to identify significant coherence regions (such as the coherence between photovoltaic voltage and temperature in the 1 Hz frequency band > 0.8).

[0032] Phase correction: Linearly interpolate and compensate for the phase difference in the coherence region. For example, if it is detected that the temperature field change leads the photovoltaic fluctuation by 20 ms, the entire photovoltaic pulse sequence is shifted forward by 20 ms.

[0033] The output multi-scale pulse tensor has dimensions N × S × T, where S is the scale parameter (such as three frequency bands: 0.1 Hz, 1 Hz, 10 Hz), and T is the synchronized time step. In this tensor, the pulse events of photovoltaic, battery, and load are strictly aligned in the time-frequency domain, providing a consistent benchmark for subsequent analysis.

[0034] Input the multi-scale pulse tensor into the multi-scale pulse fusion tensor decomposition algorithm to construct the pulse energy density spectrum in the time-frequency hybrid domain. Decompose the core feature sub-tensors of photovoltaic, energy storage, and load through the time-frequency constraint mechanism parameterized by the convolution kernel, and suppress time-frequency aliasing interference; The multi-scale pulse fusion tensor decomposition algorithm (MPFTD) combines the advantages of tensor decomposition and convolutional neural networks. The specific steps are as follows: Construction of pulse energy density spectrum: Perform sliding window integration on the pulse matrix of each scale (frequency band) (window width 100 ms, step size 10 ms), and calculate the pulse density (number of pulses / window) as an energy index. For example, in the 10 Hz frequency band, the photovoltaic pulse density reflects the fluctuations caused by the series connection effect of the strings.

[0035] Convolution kernel parameterization: Design three groups of learnable convolution kernels, corresponding to the characteristic patterns of photovoltaic, battery, and load respectively: Photovoltaic kernel: Size 3×3 (time × scale), capturing the fluctuation patterns caused by cloud movement; Battery core: Size 5×1, extracting the slow-varying features during charge and discharge processes; Load core: Size 1×5, identifying sudden load demands.

[0036] Feature separation and reconstruction: Extracting the core features of each power source through convolution operations and using non-negative matrix factorization (NMF) to constrain the non-negativity of the feature sub-tensors. For example, in the photovoltaic sub-tensor, high-frequency components (>5Hz) are suppressed, and the minute-level fluctuation trend is retained.

[0037] Time-frequency aliasing suppression is achieved through frequency-domain masking: Aliasing detection: Calculating the cross-power spectral density (CPSD) between scales to identify aliasing regions (such as harmonic interference in the 1Hz and 10Hz frequency bands).

[0038] Mask generation: Applying an exponentially decaying weight (decay factor 0.8) to the aliasing region to reduce its contribution. For example, during the battery charge and discharge process, the 1Hz current ripple is aliased with the 10Hz PWM switching noise, and the mask attenuates its energy by 64%.

[0039] The decomposed core feature sub-tensors include: Photovoltaic sub-tensor: Dimension N_pv×S_pv×T, reflecting irradiance changes and string mismatch; Battery sub-tensor: Dimension N_bat×S_bat×T, characterizing SOC changes and internal resistance aging; Load sub-tensor: Dimension N_load×S_load×T, describing load steps and harmonic distortion.

[0040] ‌1. Photovoltaic system parameters N_pv (dimension of photovoltaic parameters) Meaning: The number of characteristic parameters of the photovoltaic array, usually including: output voltage (V_pv), output current (I_pv), power (P_pv), string mismatch rate (%).

[0041] Example: If 3 types of parameters (voltage, current, power) are monitored, then N_pv = 3.

[0042] S_pv (photovoltaic frequency band scale) Meaning: The number of frequency band divisions for photovoltaic feature analysis, used to capture fluctuations at different time scales: Low frequency (0.1Hz): Reflecting minute-level fluctuations caused by cloud movement; Medium frequency (1Hz): Second-level fluctuations caused by string cascading effects; High frequency (10Hz): Millisecond-level disturbances such as inverter switching noise.

[0043] Example: If 3 frequency bands are divided, then S_pv = 3.

[0044] 2. Energy storage battery parameters N_bat (Battery parameter dimension) Meaning: Key monitoring parameters of the battery system, such as: State of Charge (SOC, %), charge and discharge current (A), internal resistance (mΩ), temperature (°C).

[0045] Example: If three items of SOC, current, and temperature are selected, then N_bat = 3.

[0046] S_bat (Battery frequency band scale) Meaning: Frequency band division for battery characteristic analysis, focusing on different dynamic processes: Ultra-low frequency (0.01 Hz): Slow change of SOC (hour level); Low frequency (0.1 Hz): Charge and discharge cycle (minute level); High frequency (1 Hz): Current ripple (second level).

[0047] Example: If two frequency bands (SOC and current ripple) are analyzed, then S_bat = 2.

[0048] 3. Load parameters N_load (Load parameter dimension) Meaning: Core parameters at the load end, such as: active power (kW), reactive power (kVar), total harmonic distortion (THD%), sudden load addition / removal flag (Boolean value).

[0049] Example: If power and harmonics are monitored, then N_load = 2.

[0050] S_load (Load frequency band scale) Meaning: Frequency band division of load dynamic behavior: Fundamental frequency (50 / 60 Hz): Steady-state power demand; Medium frequency (100 - 500 Hz): Motor starting transient; High frequency (>1 kHz): Harmonics of switching devices.

[0051] Example: If the fundamental frequency and harmonics are concerned, then S_load = 2.

[0052] Perform pulse energy accumulation and spatial topology mapping on the core feature sub-tensor, construct a multi-source energy dynamic topology map based on the pulse-triggered dynamic weight allocation algorithm, and generate a spatio-temporally aligned power state joint tensor; among them, the topology map reflects the dynamic coupling relationship between energy nodes in real time through the pulse-triggered edge weight update mechanism.

[0053] Pulse energy accumulation adopts time-exponentially weighted moving average (EWMA): Attenuation factor setting: Select the attenuation speed according to the characteristic frequency band. For low-frequency characteristics (such as battery SOC), the attenuation factor λ = 0.9 (slow attenuation), and for high-frequency characteristics (such as load mutation), λ = 0.5 (fast attenuation).

[0054] Energy integration: Perform weighted summation on each characteristic subtensor in the time dimension. For example, the energy integration of the photovoltaic subtensor reflects the cumulative irradiance within 15 minutes.

[0055] The spatial topology mapping is based on the physical connection relationship of power supply nodes: Node definition: Abstract the photovoltaic array, battery pack, and load center as topological nodes. Each node contains position coordinates (such as GPS data) and electrical parameters (such as rated power).

[0056] Edge weight initialization: Set the initial weight according to the electrical distance (such as cable impedance). The lower the impedance, the higher the weight.

[0057] The dynamic weight allocation algorithm updates the edge weight through pulse events: Pulse trigger condition: When the characteristic energy of a certain node exceeds the threshold (such as the energy of the photovoltaic node > 100 kW·s), trigger the weight update.

[0058] Weight update rule: Positive coupling: If the pulse event of node A causes the energy of node B to increase (such as increased photovoltaic power generation promoting battery charging), the edge weight increases by Δw = 0.1; Negative coupling: If the pulse of node A suppresses the energy of node B (such as a sudden increase in load causing accelerated battery discharge), the edge weight decreases by Δw = 0.05.

[0059] The generated joint tensor of power supply states has dimensions N×S×T×C, where C is the coupling relationship channel (such as photovoltaic-battery, battery-load, etc.). For example, when the load suddenly increases, the coupling strength of the battery-load channel rises from 0.7 to 0.9, triggering an adjustment of the coordinated power supply strategy.

[0060] Example application: In a microgrid scenario, when the output of the photovoltaic array drops suddenly due to cloud cover, the pulse trigger mechanism detects in real time that the weight of the photovoltaic-battery edge drops from 0.6 to 0.4, and the system immediately starts battery discharge compensation to ensure the continuity of load power supply.

[0061] S202, input the joint tensor of the power supply states into the adversarial prediction network with physical constraints, reconstruct the multi-source coordinated power supply manifold based on the dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through the residual-focusing frequency domain decomposition algorithm, and output the energy dynamic balance vector including the prediction of the power supply margin; Specifically, the power supply state joint tensor can be input into the adversarial prediction network with physical constraints. The potential representation of the multi-source power supply manifold is constructed through the generator network, and the frequency-domain sparsity constraint is imposed by the discriminator network to generate adversarial feature vectors. Among them, the discriminator uses Fourier-domain adversarial constraints to force the generator to separate steady-state and transient features. The adversarial prediction network consists of a generator and a discriminator, which work together through physical constraints and a frequency-domain adversarial mechanism. The generator adopts a deep residual network (ResNet) architecture. The input is the power supply state joint tensor (dimension: time step × number of energy nodes × feature dimension, such as 100×5×64), and the output is the potential power supply manifold representation. The network structure contains 4 residual blocks, each block consists of a 1D convolutional layer (kernel size 3, stride 1), batch normalization (BatchNorm), and a gated linear unit (GLU). The residual skip connection alleviates the problem of gradient disappearance. For example, the transient fluctuation characteristics of a photovoltaic array (such as a sudden power drop caused by cloud occlusion) capture high-frequency changes through the residual block, while the slow charge and discharge characteristics of an energy storage battery retain low-frequency information through the skip connection.

[0062] The discriminator is designed as a frequency-domain and time-domain dual-channel structure: Time-domain channel: A 3-layer convolutional network (kernel size 5, number of channels 32 / 64 / 128) extracts time-domain features. Frequency-domain channel: The fast Fourier transform (FFT) is performed on the input data to convert the time-domain signal into a frequency-domain energy spectrum. Key frequency bands are extracted through band-pass filtering (0.1Hz~10Hz), and then processed through 3 layers of convolution.

[0063] The Fourier-domain adversarial constraint is achieved through gradient backpropagation in the frequency-domain channel. The discriminator imposes a sparsity constraint on the frequency-domain energy spectrum output by the generator: The steady-state components (such as the daily power fluctuation of a photovoltaic) are concentrated in the low-frequency band (0.1Hz~1Hz) and need to be kept smooth; the transient components (such as load mutations) are distributed in the high-frequency band (1Hz~10Hz) and need to highlight the spike features. For example, if the generator fails to separate the 10Hz high-frequency component of a certain load mutation, the discriminator will force the generator to correct it through frequency-domain gradient penalty (such as increasing the loss weight in the high-frequency band).

[0064] The physical constraint is achieved through the energy conservation loss function, which calculates the mean square error (MSE) between the total power of the power supply manifold output by the generator and the actual power supply input. For example, if the predicted power of a photovoltaic is 5kW and the energy storage output is 3kW at a certain moment, the total power prediction value of the generator must be strictly equal to 8kW (load demand), and an error exceeding 5% will trigger a loss penalty.

[0065] The finally generated adversarial feature vector (dimension 256) contains a decoupled representation of the steady-state baseline and transient perturbations. For example, the first 128 dimensions of the vector encode the steady-state coupling relationship of the photovoltaic-storage system, and the last 128 dimensions capture the spatio-temporal patterns of load mutations.

[0066] Perform frequency-domain sparse residual focusing processing on the adversarial feature vector, extract the load mutation residual components through a band-pass filtering kernel with an adaptive bandwidth, and use a steady-state fluctuation suppression algorithm to eliminate background noise, and output a high-resolution load mutation feature spectrum; The goal of frequency-domain sparse residual focusing is to separate the load mutation signal from the adversarial feature vector. The specific process is as follows: Frequency-domain transformation: Perform a short-time Fourier transform (STFT) on the feature vector with a window length of 256 ms and an overlap rate of 75% to generate a time-frequency matrix (time × frequency × energy, such as 100 × 50 × 256).

[0067] Adaptive band-pass filtering: Design a Gaussian filtering kernel with an adjustable bandwidth according to the historical load mutation frequency distribution (e.g., 90% of the mutations are concentrated in 2 Hz - 8 Hz). The center frequency of the filtering kernel is initially 5 Hz and the bandwidth is 2 Hz, and it is dynamically adjusted according to the real-time spectrum energy: If a sudden increase in energy is detected in a certain frequency band (e.g., the energy at 3 Hz increases by 30%), the center frequency shifts towards that location and the bandwidth is narrowed to 1 Hz to improve the resolution. For example, when a 3.5 Hz load spike caused by the simultaneous startup of air conditioners is detected, the filtering kernel automatically adjusts to 3.5 Hz ± 0.5 Hz.

[0068] Residual extraction: Take the difference between the filtered time-frequency matrix and the original time-frequency matrix to obtain the residual components. The steady-state fluctuations (such as the change in photovoltaic power caused by the gradual change of sunlight) are suppressed outside the frequency band, while the load mutation components are retained due to the enhancement within the frequency band.

[0069] Background noise suppression: Use moving average filtering (window length 1 second) to eliminate the random noise in the residuals, and further denoise through threshold gating (e.g., setting the residuals with energy lower than the mean by 2σ to zero).

[0070] The steady-state fluctuation suppression algorithm is based on the multi-scale analysis of wavelet transform: Perform 5-layer wavelet decomposition (the mother wavelet is db4) on the original feature vector, and extract the approximation coefficients (low-frequency steady-state components) and detail coefficients (high-frequency noise); Set to zero the high-frequency noise (such as sensor quantization error) in the detail coefficients that is irrelevant to the load mutation, reconstruct the signal and subtract it from the residuals to improve the signal-to-noise ratio.

[0071] In the finally output load mutation feature spectrum (dimension 100×50), each time-frequency unit is labeled with a mutation intensity (0~1). For example, an intensity of 0.8 in the 5Hz frequency band at a certain moment indicates that a load surge event lasting 200ms has been detected.

[0072] Construct a game payoff matrix for load mutation and steady-state fluctuation based on a dynamic game framework, and dynamically allocate the attention weights of the two types of features through an adversarial attention mechanism. Among them, the adversarial attention mechanism optimizes the sparsity of the attention mask through a game strategy gradient update mechanism; The dynamic game framework models load mutation (the attacking side) and steady-state fluctuation (the defending side) as two players in the game, and the goal is to balance their impacts in power supply margin prediction.

[0073] Payoff matrix construction: The payoff function of the load mutation side is the positive correlation between the mutation intensity and the prediction error (the stronger the mutation, the larger the prediction error, and the higher the payoff); The payoff function of the steady-state fluctuation side is the negative correlation between the prediction result and the smoothing benchmark (the smoother the fluctuation, the smaller the prediction error, and the higher the payoff).

[0074] The dimension of the payoff matrix is the mutation intensity level (5 levels) × the fluctuation smoothing level (5 levels). For example, when the mutation intensity is level 3 (medium) and the fluctuation smoothing level is level 2 (mild fluctuation), the payoff of the mutation side is 0.6 and the payoff of the fluctuation side is 0.4.

[0075] Adversarial attention mechanism: The attention mask (dimension 100) is the weight distribution on the time series, and the initial value is generated by the time-domain integration of the load mutation feature spectrum; Optimize the mask sparsity through the policy gradient algorithm (PPO): During training, if the prediction error in the high attention weight region (such as the moment when mutation occurs) decreases, increase the weight of this region; conversely, if the error in the low weight region increases, decrease the weight. The learning rate is set to 0.001, and the KL divergence threshold of 0.02 prevents over-adjustment.

[0076] Example of dynamic weight allocation: When a continuous load mutation is detected (such as the startup of a data center server cluster), the weight of the attention mask rises to 0.9 during the mutation period (t = 10s~15s), suppressing the impact of steady-state fluctuation; During the stable operation stage (t = 30s~60s), the weight drops to 0.2, giving priority to the slow fluctuations of photovoltaic / storage.

[0077] During the training process, both sides of the game explore the strategy space through Monte Carlo simulation. After 1000 iterations, the sparsity of the attention mask (the proportion of non-zero weights) is optimized from the initial 80% to 30%, significantly enhancing the detection weights of key mutation events.

[0078] Perform time-domain integration and energy accumulation on the load mutation feature spectrum, and combine with the state of charge constraint of the energy storage battery. Generate an energy dynamic balance vector containing the power supply capacity in the next few seconds through the power supply margin prediction model; among them, the power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

[0079] The power supply margin prediction model adopts a hybrid architecture of a temporal convolutional network (TCN) and a long short-term memory network (LSTM): Time-domain integration: Integrate the load mutation feature spectrum according to a time window (such as 1 second) to generate a mutation energy sequence (dimension 100×1); Energy accumulation: Superimpose the predicted power of photovoltaic, the real-time output of energy storage, and the mutation energy to obtain a net power supply capacity sequence; State of charge (SOC) constraint: When the SOC of the energy storage is lower than 20%, limit the discharge power to 50% of the rated value; when the SOC is higher than 80%, limit the charging power to 30%.

[0080] Adversarial sample enhancement injects two types of perturbations into the training data: Load mutation perturbation: Randomly insert false mutations with an amplitude of ±20% and a duration of 0.1 - 1 second; Sensor noise: Add Gaussian noise (μ = 0, σ = 0.05) to simulate signal acquisition errors.

[0081] The model training adopts the Curriculum Learning strategy: Primary stage: Train with pure data, learning rate 0.001, batch size 64; Advanced stage: Gradually increase the proportion of adversarial samples to 50%, reduce the learning rate to 0.0001, and batch size 32.

[0082] The prediction output is an energy dynamic balance vector (dimension 10), and each element corresponds to the power supply margin in the next 1 second (unit kW). For example, the vector [5.2, 4.8, 4.5,..., 3.0] represents the additional power margin that can be provided per second within the next 10 seconds. When the margin in a certain second is lower than the threshold (such as 1kW), trigger emergency discharge of the energy storage or load shedding.

[0083] Example application: Scenario 1: The sudden increase in load is 8 kW, the current PV output is 5 kW, the available power of the energy storage is 3 kW, and the margin vector shows that the margins in the next 3 seconds are 0 kW, -2 kW, and -5 kW respectively. The system immediately starts the standby diesel generator; Scenario 2: The steady-state load fluctuates by ±1 kW, and the margin vector is stable at 2 kW - 3 kW. The system maintains the current strategy.

[0084] After adversarial training, the prediction error of the model in a noisy environment is reduced from 15% to 6%, and the decision-making stability in the critical state of SOC (20% - 30%) is improved by 40%.

[0085] S203. Perform multi-objective dynamic game optimization on the energy dynamic balance vector, use a deep policy network driven by Monte Carlo tree search to generate a candidate set of power supply strategies, and screen the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and output an anti-interference multi-source collaborative power supply strategy matrix; Specifically, according to the energy dynamic balance vector, a deep policy network driven by Monte Carlo tree search can be used to generate a candidate tree of power supply strategies. Among them, the deep policy network synchronously evaluates the short-term benefits and long-term stability of the strategy through a dual-head network of policy value; The deep policy network driven by Monte Carlo tree search (MCTS) aims to generate a candidate tree of power supply strategies that takes into account both short-term benefits (such as immediate power supply stability) and long-term stability (such as equipment life balance). The network architecture adopts a dual-head design: Policy head: Output the probability distribution of each power supply strategy in the current state (for example, the probability of choosing PV power supply is 0.6, the probability of energy storage power supply is 0.3, and the probability of mains supplement is 0.1). The dimension is the same as the action space (assuming 3 power supply modes); Value head: Predict the long-term value of the strategy (the value range is [-1, 1], a positive value indicates benefit, and a negative value indicates potential risk).

[0086] MCTS process: Selection stage: Starting from the root node, select child nodes through the UCB1 (Upper Confidence Bound) formula to balance exploration and exploitation. For example, the UCB value of a certain node = Q (average value) + 2√(ln(N) / n), where N is the number of visits to the parent node and n is the number of visits to the child node.

[0087] Expansion stage: When encountering an unvisited node, expand a new node. For example, when the state of charge (SOC) of the energy storage battery is lower than 20%, expand the node of "switching to the mains".

[0088] Simulation stage: Use a random strategy to simulate to the end state (such as predicting the power supply state in the next 30 seconds), and calculate the simulation reward (such as the number of power outages, battery loss rate).

[0089] Backtracking stage: Update the access times and value means of each node on the path. For example, a certain simulation shows that switching to photovoltaic causes a sudden voltage drop, then the value of the corresponding node decreases by 0.2.

[0090] Training parameters: The tree search depth is set to 5 layers (corresponding to 5 decision steps), and each round of iteration is 100 times; The input dimension of the policy network is 128 (the feature dimension of the energy dynamic balance vector), and the hidden layers are 256-128-64; The learning rate is 0.001, the batch size is 32, and the Adam optimizer is used.

[0091] Example: In a scenario where the photovoltaic output suddenly drops, MCTS generates the following candidate policy branches: Branch 1: Immediately start energy storage power supply (short-term benefit +0.8, long-term value -0.3, due to the risk of deep battery discharge); Branch 2: Reduce the load by 10% and call the mains power (short-term benefit +0.5, long-term value +0.6); Branch 3: Enable the standby diesel generator (short-term benefit +0.2, long-term value -0.7, due to the high carbon emission cost).

[0092] Perform a high-dimensional policy space manifold projection on the candidate tree of the power supply strategy, extract the core features of the strategy through a dimensionality reduction algorithm constrained by geodesic distance, and generate a low-dimensional compact policy manifold; wherein, the dimensionality reduction algorithm suppresses policy conflicts through manifold curvature optimization; The manifold projection of the high-dimensional policy space (such as 100-dimensional) aims to extract the core features of the strategy (such as the power supply mode switching frequency, resource consumption rate) for subsequent optimization. The specific process is as follows: Manifold construction: Use the UMAP (Uniform Manifold Approximation and Projection) algorithm, set n_neighbors = 15 (control the local structure sensitivity), min_dist = 0.1 (ensure that the points are closely distributed after projection); The input is the policy feature vector (including parameters such as power supply mode, power distribution ratio, switching delay, etc.), with a dimension of 100 → reduced to a 3D visualization space.

[0093] Geodesic distance constraint: Calculate the geodesic distance (the shortest path on the manifold) between policy points through the Floyd-Warshall algorithm, and replace the Euclidean distance to maintain the topological structure; For example, the geodesic distance between Strategy A (photovoltaic priority) and Strategy B (energy storage priority) is 1.2, while the distances between Strategy C (mains power dependence) and the former two are 2.3 and 1.8 respectively, indicating that Strategies A and B are more similar.

[0094] Curvature optimization: Detect the local curvature of the manifold (using the Ricci flow algorithm), and impose a penalty term on the high-curvature region (strategy conflict area); If the curvature radius of a certain area < 0.5 (indicating intensive strategy conflicts), then avoid strategy confusion by increasing the projection interval (such as forcing the strategy spacing ≥ 0.3).

[0095] Example: After projection, three clusters are formed in the strategy manifold: Cluster 1: Aggressive strategies (high photovoltaic utilization rate, low battery SOC protection); Cluster 2: Conservative strategies (mains power dependence, high battery redundancy); Cluster 3: Hybrid strategies (dynamic switching, medium risk).

[0096] Through curvature optimization, the boundary between Cluster 1 and Cluster 2 is clear, and the strategy conflicts are reduced by 30%.

[0097] Construct a multi-objective dynamic game model on the low-dimensional compact strategy manifold, use a Nash equilibrium solver with hybrid strategy optimization for strategy revenue games, screen the global optimal solution through the shadow price iteration algorithm, and output an anti-interference multi-source collaborative power supply strategy matrix; The multi-objective dynamic game model abstracts the optimization objectives of power supply strategies into the interest competition of multiple game players (such as "stability player", "economy player", "environmental protection player"), and finds the optimal compromise point through Nash equilibrium.

[0098] Revenue matrix construction: Revenue of the stability player: The reciprocal of the power supply interruption time (in seconds), with a weight of 0.5; Revenue of the economy player: The negative value of the total power supply cost (in yuan), with a weight of 0.3; Revenue of the environmental protection player: The negative value of the carbon emission (in kg), with a weight of 0.2.

[0099] For example, the revenue vector of Strategy A is (0.8, -500, -200), and that of Strategy B is (0.6, -300, -50).

[0100] Solving the mixed strategy Nash equilibrium: Use the Lemke-Howson algorithm to solve the equilibrium point, which supports asymmetric game matrices; The shadow price iteration algorithm is used to handle constraints (such as battery SOC ≥ 20%), and dynamically adjusts the resource allocation price through Lagrange multipliers. For example, when the SOC approaches 20%, the shadow price of energy storage power supply rises from 0.5 yuan / kWh to 1.2 yuan / kWh to inhibit overuse.

[0101] Global optimal solution screening: Pareto front generation: Screen the non-dominated solution set through the NSGA-II algorithm, with a population size of 200 and 50 iterations; Decision maker preference injection: Use TOPSIS (technique for order preference by similarity to an ideal solution) combined with weight preferences to select the final solution. For example, if stability is preferred, then select the strategy closest to the ideal solution (interruption time = 0, cost = 0, carbon emissions = 0). Example output: The optimal strategy matrix includes the information shown in Table 1: Table 1

[0102] Verify the convergence of the mixed strategy in dynamic game theory for the multi-source collaborative power supply strategy matrix, prove the superlinear convergence rate under multi-objective conflicts through Lyapunov stability analysis, and dynamically adjust the strategy search depth based on the verification results.

[0103] Lyapunov stability analysis is used to verify the convergence of the strategy in a dynamic environment to ensure that strategy adjustments do not cause oscillations or divergence.

[0104] Lyapunov function construction: Define the system energy function V(x)=∑(objective deviation)^2, such as V = 0.5*(interruption time - time objective)^2 + 0.3*(cost - cost objective)^2 + 0.2*(carbon emissions - emissions objective)^2; When V decreases over time, the system is stable.

[0105] Convergence verification: Calculate dV / dt through numerical differentiation. If dV / dt < -k*V (k > 0), then superlinear convergence is satisfied; For example, after a certain strategy is executed, V decreases from the initial value of 1.2 to 0.3 (Δt = 5 seconds), dV / dt = -0.18 < -0.1*0.3 = -0.03, satisfying the superlinear convergence condition.

[0106] Strategy search depth adjustment: If the convergence rate is lower than the threshold (such as Δt > 10 seconds), then increase the MCTS search depth (from 5 levels to 7 levels); If oscillations are detected (the sign of dV / dt alternates), then enable the backtracking mechanism and force a switch to the historical optimal strategy. Example: The initial policy search depth is 5 layers, and the convergence time is 8 seconds; It is detected that after a certain policy execution, dV / dt = -0.05, which is lower than the threshold of -0.1. The search depth is automatically adjusted to 7 layers, and the convergence time is shortened to 6 seconds; After the adjustment, the system stability is improved, and the power supply interruption volatility is reduced from ±15% to ±5%.

[0107] Monte Carlo tree search: The search depth is 5 layers, iterated 100 times, and 30 candidate policies are generated; The policy evaluation time consumption ≤ 50 ms (based on the NVIDIA Jetson AGX Xavier edge computing platform).

[0108] Manifold projection: After UMAP dimensionality reduction, the policy features are 3-dimensional, and the clustering purity ≥ 85%; The calculation error of the geodesic distance < 5% (compared with the Euclidean distance).

[0109] Nash equilibrium solution: The solution time of the Lemke-Howson algorithm ≤ 20 ms; The shadow price update frequency is 1 Hz, adapting to real-time resource fluctuations.

[0110] Lyapunov verification: The superlinear convergence rate threshold k = 0.1; The response time for dynamically adjusting the search depth ≤ 100 ms.

[0111] Through the above implementation, the multi-source collaborative power supply policy matrix reduces the power supply interruption rate to 0.5% / month in actual measurement, saves 15% in comprehensive cost, and reduces carbon emissions by 20%, significantly superior to the traditional rule control method.

[0112] S204, input the multi-source collaborative power supply policy matrix into the edge computing-driven distributed collaborative control framework, optimize the policy execution parameters based on the real-time feedback mechanism verified by digital twin, and coordinate the multi-node control instructions through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability.

[0113] Specifically, the multi-source collaborative power supply policy matrix can be input into the digital twin verification module driven by edge computing. Through the fault scenario adversarial distillation technology, the battery thermal runaway and photovoltaic shading mutation are simulated to generate a policy robustness evaluation vector; among them, the digital twin verification module optimizes the simulation parameters through the dynamic weight mirror update mechanism; The core of the digital twin verification module is to verify the robustness of the power supply strategy in extreme scenarios through a high-fidelity simulation model. First, based on the multi-source collaborative power supply strategy matrix (such as the photovoltaic power output allocation ratio, energy storage charge and discharge thresholds, etc.), digital mirrors of the battery and photovoltaic modules are constructed. The battery thermal runaway scenario is simulated through an electrochemical-thermal coupling model: when the strategy requires the battery to discharge at a 2C rate (i.e., 2 times the capacity current), the twin model calculates the internal temperature distribution in real time. If the temperature of a certain battery cell exceeds 60°C (the thermal runaway critical value), an alarm is triggered. The sudden change scenario of photovoltaic shading is simulated using a ray tracing algorithm to simulate cloud occlusion. For example, the photovoltaic output power is suddenly reduced from 5kW to 1kW within 10 seconds to verify whether the strategy can quickly switch to energy storage power supply.

[0114] The fault scenario adversarial distillation technology uses a generative adversarial network (GAN) to generate extreme working condition data: Generator: The input is historical normal working condition data (such as battery SOC = 80%, ambient temperature 25°C), and the output is simulated thermal runaway time series data (such as a 10-second curve of temperature rising from 25°C to 80°C); Discriminator: The input is real thermal runaway data and generated data, and the output is the authenticity probability. During training, the generator optimizes the distribution matching degree between the generated data and the real data through the Wasserstein loss function, and the gradient penalty coefficient λ = 10.

[0115] The dynamic weight mirror update mechanism dynamically adjusts the simulation parameters through reinforcement learning: Parameter space: includes battery internal resistance (0.05~0.2Ω), photovoltaic conversion efficiency (15%~22%); Reward function: the matching degree between the simulation result and the physical constraint (such as a reward of +1 when the temperature prediction error < ±3°C); Policy update: The PPO algorithm is used (learning rate 0.0003, batch size 64). In each round of iteration, 100 groups of parameter combinations are generated, and the top 10 groups with the highest rewards are selected to update the mirror weights.

[0116] The finally generated policy robustness evaluation vector includes the following dimensions: Thermal runaway suppression score (0~1, based on the proportion of the duration of temperature exceeding the standard); Photovoltaic switching delay (unit: ms, the speed of the strategy responding to the sudden change of shading); Energy storage cycle life loss (percentage, the impact of the strategy on battery aging). For example, a certain strategy score is [0.92, 45ms, 0.8%], indicating excellent performance in the thermal runaway scenario, but the switching delay needs to be optimized.

[0117] Construct a dynamic weight allocation model for the 5G multi-hop transmission protocol based on the policy robustness evaluation vector, and adjust the priority weights of control instructions in real time based on the base station load rate and channel quality to generate an instruction distribution sequence that resists congestion; The goal of the dynamic weight allocation model is to optimize the transmission priority of control instructions in the 5G multi-hop network to ensure that critical instructions (such as emergency shutdown of energy storage) are transmitted first. The weight allocation rule is based on the following real-time parameters: Base station load rate (CPU utilization rate, threshold 80%): When the load rate > 80%, reduce the weight of non-critical instructions (such as fine-tuning of photovoltaic power); Channel quality (SNR, unit dB): When SNR < 15dB, increase the weight of retransmission instructions; Instruction urgency (divided into levels 1 to 5, level 5 is the highest).

[0118] The dynamic weight calculation uses the Q-learning algorithm: State space: 3D vector (load rate, SNR, urgency); Action space: Weight adjustment range (-0.3 to +0.3, step size 0.1); Reward function:

[0119] During training, explore the action space through the ε-greedy strategy (ε = 0.1), learning rate α = 0.01, and discount factor γ = 0.9.

[0120] Generation of the instruction distribution sequence that resists congestion: Priority sorting: Sort instructions in descending order of weight, for example: Instruction A (energy storage shutdown, weight 0.95); Instruction B (photovoltaic switching, weight 0.85); Resource reservation: Reserve 20% of the transmission bandwidth for high-weight instructions; Dynamic backoff: If the channel is congested (packet loss rate > 5%), low-weight instructions are delayed by 1 to 3 time slots for transmission.

[0121] For example, when the base station load rate reaches 85% and SNR = 12dB, the policy increases the weight of the energy storage shutdown instruction from 0.9 to 0.95 and allocates dedicated time slots to ensure that the end-to-end delay < 50ms.

[0122] Perform multi-base station cooperative control conflict detection on the instruction distribution sequence, and use the implicit gradient compensation algorithm to eliminate the policy execution deviation. Among them, the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence through the backpropagation of the conflict residual; Multi-base station cooperative conflict detection is achieved through the distributed consensus algorithm: Collision detection: Each base station broadcasts the hash value (SHA-256) of the local instruction queue periodically (every 100 ms). If the hashes are inconsistent, a collision is determined. Collision localization: The binary search method is used to trace back the last 10 instructions to identify the source of the collision (e.g., base station A and base station B send photovoltaic load reduction instructions simultaneously).

[0123] The core of the implicit gradient compensation algorithm is to correct the instruction parameters through backpropagation: Residual calculation: Measure the deviation Δ = 3 kW between the actual execution effect (e.g., the actual photovoltaic power drops by 5 kW) and the expected target (drop by 8 kW). Gradient backpropagation: Calculate the contribution degree of each link layer by layer along the control link (base station → edge controller → inverter). For example: The contribution of the base station instruction coding error is 30%; The contribution of packet loss caused by channel noise is 50%; The contribution of the inverter response delay is 20%; Parameter correction: Adjust the instruction parameters according to the contribution degree ratio. For example, correct the target power from 8 kW to 8 kW + 3 kW × 0.3 = 8.9 kW.

[0124] Example of dynamic correction: A certain collision causes an amplitude deviation of 10 A in the energy storage charging instruction. The algorithm reduces the deviation to 0.5 A through 3 iterations (learning rate 0.1), and at the same time adjusts the phase synchronization error from 15° to 2°.

[0125] Timestamp synchronization of the corrected control instructions is performed through the time delay equalization technology of multi-base station cooperation. An instruction buffering mechanism with a sliding window constraint is adopted to ensure that the end-to-end delay is less than 1 ms, and finally a multi-source intelligent control instruction set with low delay and high reliability is output.

[0126] The time delay equalization technology is based on the IEEE 1588 Precision Time Protocol (PTP) to achieve microsecond-level synchronization: Timestamp synchronization: The master base station periodically sends synchronization messages, and the slave base stations calculate the link delay (such as transmission delay ± 0.1 ms) and dynamically compensate for the clock offset; Time delay equalization: The least mean square (LMS) filter is used to predict the link jitter. For example, the time delay fluctuation range of a certain link is 0.5~1.2 ms, and the filter stabilizes it to 0.8 ± 0.05 ms.

[0127] Sliding window buffering mechanism: Window size: 10 ms, accommodating 5 instruction frames (each frame is 2 ms); Dynamic adjustment: If the time delay of 3 consecutive frames > 0.9 ms, the window is reduced to 8 ms to accelerate processing; Redundancy elimination: Directly discard obsolete instructions (such as latency > 1ms) to ensure real-time performance.

[0128] Final instruction set generation: Timestamp alignment: Attach a unified clock label (such as GPS time source) to all instructions; Priority execution: High-weight instructions are preferentially allocated to the first half of the window; Reliability verification: Verify the integrity of instructions through CRC-32 checksum, with an error rate < 1e-6.

[0129] For example, a certain control instruction set contains 20 instructions. After latency equalization, the end-to-end latency distribution is 0.8ms ± 0.1ms, and the packet loss rate is 0.05%, meeting the requirements of low latency and high reliability. All instructions are sent through the 5G URLLC (Ultra-Reliable Low-Latency Communication) channel to ensure the real-time and stable operation of the power grid.

[0130] It can be seen that by obtaining the multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using the multi-scale pulse fusion tensor decomposition algorithm, and generating a spatio-temporally aligned power state joint tensor; inputting the power state joint tensor into a physically constrained adversarial prediction network, outputting an energy dynamic balance vector containing power supply margin prediction; performing multi-objective dynamic game optimization on the energy dynamic balance vector, outputting an anti-interference multi-source collaborative power supply strategy matrix; inputting the multi-source collaborative power supply strategy matrix into an edge computing-driven distributed collaborative control framework, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0131] Another embodiment of the present invention provides a control system for a multi-source intelligent power manager based on 5G communication. Refer to Figure 3 , the system may include: An acquisition module 301, configured to acquire multi-source heterogeneous power data of the multi-source intelligent power manager, construct a dynamic energy topology using the multi-scale pulse fusion tensor decomposition algorithm, eliminate multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generate a spatio-temporally aligned power state joint tensor; A separation module 302, configured to input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply manifold based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a residual focusing frequency-domain decomposition algorithm, and output an energy dynamic balance vector containing power supply margin prediction; An optimization module 303 is configured to perform multi-objective dynamic game optimization on the energy dynamic balance vector, generate a candidate set of power supply strategies using a deep policy network driven by Monte Carlo tree search, screen for the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and output an anti-interference multi-source collaborative power supply strategy matrix; An output module 304 is configured to input the multi-source collaborative power supply strategy matrix into a distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on a real-time feedback mechanism verified by digital twins, coordinate multi-node control instructions through a 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability.

[0132] It can be seen that obtaining multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, generating a jointly aligned power state tensor in space and time; inputting the jointly aligned power state tensor into a physically constrained adversarial prediction network, outputting an energy dynamic balance vector containing power supply margin prediction; performing multi-objective dynamic game optimization on the energy dynamic balance vector, outputting an anti-interference multi-source collaborative power supply strategy matrix; inputting the multi-source collaborative power supply strategy matrix into a distributed collaborative control framework driven by edge computing, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0133] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0134] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps: S201, obtaining multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminating multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generating a jointly aligned power state tensor in space and time; S202, inputting the jointly aligned power state tensor into a physically constrained adversarial prediction network, reconstructing a multi-source collaborative power supply manifold based on a dynamic game framework, separating load mutation and steady-state fluctuation characteristics through a residual-focusing frequency-domain decomposition algorithm, and outputting an energy dynamic balance vector containing power supply margin prediction; S203, performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a candidate set of power supply strategies using a deep policy network driven by Monte Carlo tree search, screening for the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and outputting an anti-interference multi-source collaborative power supply strategy matrix; S204. Input the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twin, and coordinate the multi-node control instructions through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability.

[0135] It can be seen that obtaining the multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology by using the multi-scale pulse fusion tensor decomposition algorithm, and generating a joint tensor of power states with spatio-temporal alignment; inputting the joint tensor of power states into the physically constrained adversarial prediction network, and outputting an energy dynamic balance vector including the prediction of power supply margin; performing multi-objective dynamic game optimization on the energy dynamic balance vector, outputting an anti-interference multi-source collaborative power supply strategy matrix; inputting the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability, so as to improve the flexibility and response speed of power management and ensure the stability and reliability of power supply.

[0136] An embodiment of the present invention also provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0137] Specifically, the above electronic device may further include a transmission device and an input / output device, where the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0138] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201. Obtain the multi-source heterogeneous power data of the multi-source intelligent power manager, construct a dynamic energy topology by using the multi-scale pulse fusion tensor decomposition algorithm, eliminate multi-source signal interference through the phase synchronization time-frequency constraint mechanism, and generate a joint tensor of power states with spatio-temporal alignment; S202. Input the joint tensor of power states into the physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply manifold based on the dynamic game framework, and separate the load mutation and steady-state fluctuation characteristics through the residual focusing frequency-domain decomposition algorithm, and output an energy dynamic balance vector including the prediction of power supply margin; S203. Perform multi-objective dynamic game optimization on the energy dynamic balance vector, generate a candidate set of power supply strategies by using a deep policy network driven by Monte Carlo tree search, and screen the global optimal solution through a Nash equilibrium solver with manifold projection constraints, and output an anti-interference multi-source collaborative power supply strategy matrix; S204. Input the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twin, and coordinate the multi-node control instructions through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability.

[0139] It can be seen that by obtaining the multi-source heterogeneous power data of the multi-source intelligent power manager, constructing a dynamic energy topology using the multi-scale pulse fusion tensor decomposition algorithm, and generating a joint tensor of power states with spatio-temporal alignment; inputting the joint tensor of power states into the physically constrained adversarial prediction network to output an energy dynamic balance vector containing power supply margin prediction; performing multi-objective dynamic game optimization on the energy dynamic balance vector to output an anti-interference multi-source collaborative power supply strategy matrix; inputting the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0140] The above has detailed the structure, features and effects of the present invention according to the illustrated embodiments. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, still within the spirit covered by the specification and drawings, shall be within the protection scope of the present invention.

Claims

1. A control method for a multi-source intelligent power manager based on 5G communication, characterized in that, The method includes: Obtaining multi-source heterogeneous power data of a multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminating multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generating a spatially and temporally aligned power state joint tensor; Inputting the power state joint tensor into a physically constrained adversarial prediction network, reconstructing a multi-source collaborative power supply current pattern based on a dynamic game framework, separating load mutation and steady-state fluctuation characteristics through a residual focusing frequency-domain decomposition algorithm, and outputting an energy dynamic balance vector containing power supply margin prediction; Performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a power supply strategy candidate set using a deep policy network driven by Monte Carlo tree search, screening the global optimal solution through a Nash equilibrium solver with manifold projection constraints, and outputting an anti-interference multi-source collaborative power supply strategy matrix; Inputting the multi-source collaborative power supply strategy matrix into a distributed collaborative control framework driven by edge computing, optimizing strategy execution parameters based on a real-time feedback mechanism verified by digital twin, coordinating multi-node control instructions through a 5G multi-hop transmission protocol with dynamic weight allocation, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability.

2. The method according to claim 1, characterized in that The obtaining of multi-source heterogeneous power data of a multi-source intelligent power manager, constructing a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminating multi-source signal interference through a phase synchronization time-frequency constraint mechanism, and generating a spatially and temporally aligned power state joint tensor includes: Obtaining multi-source heterogeneous power data of a multi-source intelligent power manager, performing pulse coding on the original time-domain signals of the photovoltaic array output waveform and the energy storage battery charge and discharge curve in the data using the dynamic threshold gating of a pulse neural network, converting the continuous waveform into a pulse trigger sequence, and generating a pulse spatio-temporal matrix with time resolution; wherein, the dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism and retains effective energy fluctuation characteristics; According to the pulse spatio-temporal matrix and the environmental temperature field distribution data, constructing a multi-source signal phase difference metric tensor, performing cross-modal time calibration through a phase synchronization algorithm with pulse timing constraints, eliminating phase drift caused by photovoltaic transient fluctuations and battery aging, and outputting a phase-aligned multi-scale pulse tensor; Inputting the multi-scale pulse tensor into a multi-scale pulse fusion tensor decomposition algorithm, constructing a pulse energy density spectrum in the time-frequency hybrid domain, decomposing the core feature sub-tensors of photovoltaic, energy storage, and load through a time-frequency constraint mechanism with convolutional kernel parameterization, and suppressing time-frequency aliasing interference; Performing pulse energy accumulation and spatial topology mapping on the core feature sub-tensors, constructing a multi-source energy dynamic topology map based on a pulse-triggered dynamic weight allocation algorithm, and generating a spatially and temporally aligned power state joint tensor; wherein, the topology map reflects the dynamic coupling relationship between energy nodes in real time through a pulse-triggered edge weight update mechanism.

3. The method according to claim 2, wherein The above-mentioned step of inputting the power state joint tensor into the physically constrained adversarial prediction network, reconstructing the multi-source collaborative power supply manifold based on the dynamic game framework, separating the load mutation and steady-state fluctuation characteristics through the residual-focusing frequency-domain decomposition algorithm, and outputting the energy dynamic balance vector including the power supply margin prediction, includes: Input the power state joint tensor into the physically constrained adversarial prediction network, construct the latent representation of the multi-source power supply manifold through the generator network, and utilize the discriminator network to impose frequency-domain sparsity constraints to generate adversarial feature vectors; wherein, the discriminator uses Fourier-domain adversarial constraints to force the generator to separate steady-state and transient characteristics; Perform frequency-domain sparse residual focusing processing on the adversarial feature vectors, extract the load mutation residual components through the band-pass filter kernel with an adaptive bandwidth, and use the steady-state fluctuation suppression algorithm to eliminate background noise, and output the high-resolution load mutation feature spectrum; Construct the game payoff matrix of load mutation and steady-state fluctuation based on the dynamic game framework, and dynamically allocate the attention weights of the two types of characteristics through the adversarial attention mechanism, wherein the adversarial attention mechanism optimizes the sparsity of the attention mask through the game strategy gradient update mechanism; Perform time-domain integration and energy accumulation on the load mutation feature spectrum, combine the energy storage battery state of charge constraint, and generate the energy dynamic balance vector including the power supply capacity in the next few seconds through the power supply margin prediction model; wherein, the power supply margin prediction model improves the prediction robustness through the adversarial sample enhancement technology.

4. The method according to claim 3, characterized in that, The above-mentioned step of performing multi-objective dynamic game optimization on the energy dynamic balance vector, using the Monte Carlo tree search-driven deep policy network to generate a candidate set of power supply strategies, and screening the global optimal solution through the Nash equilibrium solver with manifold projection constraints, and outputting the anti-interference multi-source collaborative power supply strategy matrix, includes: According to the energy dynamic balance vector, use the Monte Carlo tree search-driven deep policy network to generate a candidate tree of power supply strategies, wherein the deep policy network synchronously evaluates the short-term benefits and long-term stability of the strategies through the policy value dual-head network; Perform a high-dimensional policy space manifold projection on the candidate tree of power supply strategies, extract the core features of the strategies through the dimensionality reduction algorithm with geodesic distance constraints, and generate a low-dimensional compact policy manifold; wherein, the dimensionality reduction algorithm suppresses strategy conflicts through manifold curvature optimization; Construct a multi-objective dynamic game model on the low-dimensional compact policy manifold, perform strategy payoff games using the Nash equilibrium solver with mixed strategy optimization, and screen the global optimal solution through the shadow price iteration algorithm, and output the anti-interference multi-source collaborative power supply strategy matrix; Perform the convergence verification of the mixed strategy in dynamic game theory on the multi-source collaborative power supply strategy matrix, prove the superlinear convergence rate under multi-objective conflicts through Lyapunov stability analysis, and dynamically adjust the strategy search depth based on the verification results.

5. The method according to claim 4, wherein Input the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twin, coordinate the multi-node control instructions through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability, including: Input the multi-source collaborative power supply strategy matrix into the digital twin verification module driven by edge computing, simulate battery thermal runaway and photovoltaic shading mutations through the fault scenario adversarial distillation technology, and generate a strategy robustness evaluation vector; wherein, the digital twin verification module optimizes the simulation parameters through the dynamic weight mirror update mechanism; According to the strategy robustness evaluation vector, construct a dynamic weight allocation model for the 5G multi-hop transmission protocol, adjust the priority weight of the control instructions in real time based on the base station load rate and channel quality, and generate an anti-congestion instruction distribution sequence; Perform multi-base station collaborative control conflict detection on the instruction distribution sequence, and adopt the implicit gradient compensation algorithm to eliminate the strategy execution deviation, wherein the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence through the backpropagation of the conflict residuals; Synchronize the timestamps of the corrected control instructions through the multi-base station collaborative delay equalization technology, and adopt the instruction buffering mechanism with sliding window constraints to ensure that the end-to-end delay is less than 1ms, and finally output a multi-source intelligent control instruction set with low latency and high reliability.

6. A control system for a multi-source intelligent power manager based on 5G communication, characterized in that, The system includes: An acquisition module, which is used to acquire the multi-source heterogeneous power data of the multi-source intelligent power manager, construct a dynamic energy topology by using the multi-scale pulse fusion tensor decomposition algorithm, eliminate multi-source signal interference through the phase synchronization time-frequency constraint mechanism, and generate a spatio-temporally aligned power state joint tensor; A separation module, which is used to input the power state joint tensor into the physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply manifold based on the dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through the residual focusing frequency domain decomposition algorithm, and output an energy dynamic balance vector including power supply margin prediction; An optimization module, which is used to perform multi-objective dynamic game optimization on the energy dynamic balance vector, generate a power supply strategy candidate set by using the deep policy network driven by Monte Carlo tree search, and screen the global optimal solution through the Nash equilibrium solver with manifold projection constraints, and output an anti-interference multi-source collaborative power supply strategy matrix; An output module, which is used to input the multi-source collaborative power supply strategy matrix into the edge computing-driven distributed collaborative control framework, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twin, coordinate the multi-node control instructions through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally output a multi-source intelligent control instruction set that meets low latency and high reliability.

7. The system according to claim 6, characterized in that The acquisition module is specifically used for: Obtain the multi-source heterogeneous power data of the multi-source intelligent power manager. According to the original time-domain signals of the photovoltaic array output waveform and the charge-discharge curve of the energy storage battery in the data, use the dynamic threshold gating of the spiking neural network for pulse coding, convert the continuous waveform into a pulse trigger sequence, and generate a pulse spatio-temporal matrix with time resolution; wherein, the dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism and retains the effective energy fluctuation characteristics; According to the pulse spatio-temporal matrix and the environmental temperature field distribution data, construct a multi-source signal phase difference metric tensor, and perform cross-modal time calibration through a phase synchronization algorithm with pulse timing constraints to eliminate the phase drift caused by photovoltaic transient fluctuations and battery aging, and output a phase-aligned multi-scale pulse tensor; Input the multi-scale pulse tensor into the multi-scale pulse fusion tensor decomposition algorithm, construct a pulse energy density spectrum in the time-frequency hybrid domain, and decompose the core feature sub-tensors of photovoltaic, energy storage, and load through a time-frequency constraint mechanism with parameterized convolution kernels, and suppress time-frequency aliasing interference; Perform pulse energy accumulation and spatial topology mapping on the core feature sub-tensors, construct a multi-source energy dynamic topology map based on the dynamic weight allocation algorithm triggered by pulses, and generate a spatio-temporally aligned power state joint tensor; wherein, the topology map reflects the dynamic coupling relationship between energy nodes in real time through the edge weight update mechanism triggered by pulses.

8. The system according to claim 7, wherein The separation module is specifically used for: Input the power state joint tensor into a physically constrained adversarial prediction network, construct a latent representation of the multi-source power supply manifold through the generator network, and use the discriminator network to impose frequency-domain sparsity constraints to generate an adversarial feature vector; wherein, the discriminator uses Fourier-domain adversarial constraints to force the generator to separate steady-state and transient features; Perform frequency-domain sparse residual focusing processing on the adversarial feature vector, extract the load mutation residual component through a band-pass filter kernel with an adaptive bandwidth, and use a steady-state fluctuation suppression algorithm to eliminate background noise, and output a high-resolution load mutation feature spectrum; Construct a game payoff matrix for load mutation and steady-state fluctuation based on a dynamic game framework, and dynamically allocate the attention weights of the two types of features through an adversarial attention mechanism, wherein the adversarial attention mechanism optimizes the sparsity of the attention mask through a game strategy gradient update mechanism; Perform time-domain integration and energy accumulation on the load mutation feature spectrum, and combine the energy storage battery state of charge constraint, and generate an energy dynamic balance vector including the power supply capacity in the next few seconds through a power supply margin prediction model; wherein, the power supply margin prediction model improves the prediction robustness through an adversarial sample enhancement technique.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is set to execute the method according to any one of claims 1-5 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1-5.

Images