A control method and system for a multi-source intelligent power manager based on 5G communication

Abstract

The present invention discloses a control method and system for a multi-source intelligent power manager based on 5G communication. The method includes: 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, and generating a time-space aligned power state joint tensor; inputting the power state joint tensor into a physically constrained adversarial prediction network, and outputting an energy dynamic balance vector containing a power supply margin prediction; performing multi-objective dynamic game optimization on the energy dynamic balance vector, 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, and finally outputting a multi-source intelligent control instruction set that meets low latency and high reliability. Utilizing the embodiments of the present invention, the flexibility and response speed of power management can be improved, and the stability and reliability of power supply can be ensured.

Description

Technical Field

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

[0002] Traditional power management methods often rely on static control strategies and a single power scheduling mechanism. These methods fail to fully leverage the strengths of various power sources in the system and are unable to adapt to real-time changes in load demand. This leads to problems such as inefficient power utilization, unstable power supply, and potentially frequent power outages. Summary of the Invention

[0003] The purpose of the present invention is to provide a control method and system for a multi-source intelligent power manager based on 5G communication to address 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:

[0005] Acquire multi-source heterogeneous power data from a multi-source intelligent power manager, use a multi-scale pulse fusion tensor decomposition algorithm to build a dynamic energy topology, eliminate multi-source signal interference through a phase-synchronized time-frequency constraint mechanism, and generate a space-time aligned power state joint tensor;

[0006] The power state joint tensor is input into a physically constrained adversarial prediction network, and the multi-source collaborative power supply shape is reconstructed based on a dynamic game framework. The load mutation and steady-state fluctuation characteristics are separated by a residual-focused frequency domain decomposition algorithm, and an energy dynamic balance vector including power supply margin prediction is output;

[0007] A multi-objective dynamic game optimization is performed on the energy dynamic balance vector. A deep strategy network driven by Monte Carlo tree search is used to generate a candidate set of power supply strategies. A Nash equilibrium solver with manifold projection constraints is used to screen the global optimal solution, and an interference-resistant multi-source collaborative power supply strategy matrix is output.

[0008] The multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing. The strategy execution parameters are optimized based on the real-time feedback mechanism verified by digital twins. The multi-node control instructions are coordinated through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally the multi-source intelligent control instruction set that meets the requirements of low latency and high reliability is output.

[0009] Optionally, the method of acquiring 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-synchronized time-frequency constraint mechanism, and generating a spatiotemporally aligned power state joint tensor includes:

[0010] The multi-source heterogeneous power data from the multi-source intelligent power manager is acquired. Based on the original time domain signals of the photovoltaic array output waveform and the energy storage battery charge and discharge curve in the data, pulse encoding is performed using dynamic threshold gating of a pulse neural network. The continuous waveform is converted into a pulse trigger sequence, generating a pulse space-time matrix with time resolution. The dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism, preserving the effective energy fluctuation characteristics.

[0011] Based on the pulse spatiotemporal matrix and the ambient temperature field distribution data, a multi-source signal phase difference metric tensor is constructed. A phase synchronization algorithm with pulse timing constraints is used to perform cross-modal time calibration, eliminating phase drift caused by photovoltaic transient fluctuations and battery aging, and outputting a phase-aligned multi-scale pulse tensor.

[0012] The multi-scale pulse tensor is input into the multi-scale pulse fusion tensor decomposition algorithm to construct the pulse energy density spectrum in the time-frequency hybrid domain. The core characteristic sub-tensors of photovoltaic, energy storage and load are decomposed through the time-frequency constraint mechanism parameterized by the convolution kernel, and the time-frequency aliasing interference is suppressed.

[0013] Pulse energy accumulation and spatial topological mapping are performed on the core characteristic sub-tensor, and a multi-source energy dynamic topological map is constructed based on a pulse-triggered dynamic weight allocation algorithm to generate a spatiotemporally aligned power state joint tensor; wherein, the topological map reflects the dynamic coupling relationship between energy nodes in real time through a pulse-triggered edge weight update mechanism.

[0014] Optionally, the power state joint tensor is input into a physically constrained adversarial prediction network, a multi-source collaborative power supply shape is reconstructed based on a dynamic game framework, a frequency domain decomposition algorithm with residual focus is used to separate load mutation and steady-state fluctuation characteristics, and an energy dynamic balance vector including a power supply margin prediction is output, including:

[0015] The power state joint tensor is input into a physically constrained adversarial prediction network, a potential representation of the multi-source power supply shape is constructed through a generator network, and a frequency domain sparsity constraint is imposed using a discriminator network 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;

[0016] Performing frequency-domain sparse residual focusing processing on the adversarial feature vector, extracting the load mutation residual component through a bandpass filter kernel with adaptive bandwidth, and using a steady-state fluctuation suppression algorithm to eliminate background noise, outputting a high-resolution load mutation feature spectrum;

[0017] Based on the dynamic game framework, a game payoff matrix of load mutation and steady-state fluctuation is constructed. The attention weights of the two types of features are dynamically allocated through an adversarial attention mechanism. The adversarial attention mechanism optimizes the sparsity of the attention mask through the game strategy gradient update mechanism.

[0018] The load mutation characteristic spectrum is integrated in the time domain and energy is accumulated. Combined with the state of charge constraint of the energy storage battery, a power supply margin prediction model is used to generate an energy dynamic balance vector including the power supply capacity for several seconds in the future. The power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

[0019] Optionally, the energy dynamic balance vector is subjected to multi-objective dynamic game optimization, a deep strategy network driven by Monte Carlo tree search is used to generate a candidate set of power supply strategies, a global optimal solution is screened by a Nash equilibrium solver constrained by manifold projection, and an anti-interference multi-source collaborative power supply strategy matrix is output, including:

[0020] Based on the energy dynamic balance vector, a deep policy network driven by Monte Carlo tree search is used to generate a power supply strategy candidate tree, wherein the deep policy network synchronously evaluates the short-term benefits and long-term stability of the strategy through a policy value dual-headed network;

[0021] Performing high-dimensional strategy space manifold projection on the power supply strategy candidate tree, extracting strategy core features through a geodesic distance-constrained dimensionality reduction algorithm, and generating a low-dimensional compact strategy manifold; wherein the dimensionality reduction algorithm suppresses strategy conflicts through manifold curvature optimization;

[0022] A multi-objective dynamic game model is constructed on the low-dimensional compact strategy manifold, a Nash equilibrium solver optimized by a hybrid strategy is used to conduct a strategy-benefit game, a shadow price iterative algorithm is used to screen the global optimal solution, and an anti-interference multi-source collaborative power supply strategy matrix is output;

[0023] The hybrid strategy convergence in dynamic game theory is verified for the multi-source collaborative power supply strategy matrix, and the superlinear convergence rate under multi-objective conflict is proved through Lyapunov stability analysis. The strategy search depth is dynamically adjusted based on the verification results.

[0024] Optionally, the multi-source collaborative power supply strategy matrix is input into a distributed collaborative control framework driven by edge computing, the strategy execution parameters are optimized based on a real-time feedback mechanism verified by digital twins, and multi-node control instructions are coordinated through a 5G multi-hop transmission protocol with dynamic weight allocation, ultimately outputting a multi-source intelligent control instruction set that meets low latency and high reliability, including:

[0025] The multi-source collaborative power supply strategy matrix is input into the edge computing-driven digital twin verification module, and the battery thermal runaway and photovoltaic shading mutation are simulated through the fault scenario adversarial distillation technology to generate a strategy robustness evaluation vector; wherein, the digital twin verification module optimizes the simulation parameters through a dynamic weight mirror update mechanism;

[0026] Based on the strategy robustness evaluation vector, a dynamic weight allocation model for the 5G multi-hop transmission protocol is constructed. The priority weights of control instructions are adjusted in real time based on the base station load rate and channel quality to generate an anti-congestion instruction distribution sequence.

[0027] Performing multi-base station coordinated control conflict detection on the instruction distribution sequence and using an implicit gradient compensation algorithm to eliminate policy execution deviations, wherein the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence by backpropagation of the conflict residual;

[0028] The corrected control instructions are timestamped and synchronized through multi-base station collaborative delay balancing technology, and a sliding window-constrained instruction buffer mechanism is used to ensure end-to-end delay is less than 1ms, ultimately outputting a multi-source intelligent control instruction set with low latency and high reliability.

[0029] Another embodiment of the present application provides a control system for a multi-source intelligent power manager based on 5G communication, the system comprising:

[0030] The acquisition module is used to obtain multi-source heterogeneous power data from the multi-source intelligent power manager, build a dynamic energy topology using a multi-scale pulse fusion tensor decomposition algorithm, eliminate multi-source signal interference through a phase-synchronized time-frequency constraint mechanism, and generate a time-space aligned power state joint tensor;

[0031] a separation module for inputting the power state joint tensor into a physically constrained adversarial prediction network, reconstructing the multi-source collaborative power supply shape based on a dynamic game framework, separating load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and outputting an energy dynamic balance vector including a power supply margin prediction;

[0032] An optimization module is used to perform multi-objective dynamic game optimization on the energy dynamic balance vector, generate a candidate set of power supply strategies using a deep strategy network driven by Monte Carlo tree search, screen the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and output an interference-resistant multi-source collaborative power supply strategy matrix;

[0033] The output module is used to input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

[0034] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0035] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0036] Compared with the prior art, the present invention provides a control method for a multi-source intelligent power manager based on 5G communication, which obtains multi-source heterogeneous power data of the multi-source intelligent power manager, adopts a multi-scale pulse fusion tensor decomposition algorithm to construct a dynamic energy topology, and generates a time-space aligned power state joint tensor; the power state joint tensor is input into a physically constrained adversarial prediction network, and an energy dynamic balance vector containing a power supply margin prediction is output; the energy dynamic balance vector is optimized through multi-objective dynamic game, and an anti-interference multi-source collaborative power supply strategy matrix is output; the multi-source collaborative power supply strategy matrix is input into a distributed collaborative control framework driven by edge computing, and finally a multi-source intelligent control instruction set that meets low latency and high reliability is output, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A hardware block diagram of a computer terminal that implements a control method for a multi-source intelligent power manager based on 5G communication according to an embodiment of the present invention;

[0038] Figure 2 A flowchart of a control method for a multi-source intelligent power manager based on 5G communication provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of the structure of a control system of a multi-source intelligent power manager based on 5G communication provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0041] An embodiment of the present invention first provides a control method for a multi-source intelligent power manager based on 5G communication. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0042] The following describes it in detail by taking running on a computer terminal as an example. 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 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0043] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to execute any one of the control methods for a multi-source intelligent power manager based on 5G communication.

[0044] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0045] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any control method of the multi-source intelligent power manager based on 5G communication.

[0046] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0047] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0048] See also 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:

[0049] S201, acquires multi-source heterogeneous power data from the multi-source intelligent power manager, uses a multi-scale pulse fusion tensor decomposition algorithm to build a dynamic energy topology, eliminates multi-source signal interference through a phase-synchronized time-frequency constraint mechanism, and generates a time-space aligned power state joint tensor;

[0050] Specifically, the multi-source heterogeneous power data of the multi-source intelligent power manager can be obtained. Based on the original time domain signals of the photovoltaic array output waveform and the energy storage battery charge and discharge curve in the data, pulse encoding is performed using dynamic threshold gating of a pulse neural network to convert the continuous waveform into a pulse trigger sequence, generating a pulse space-time matrix with time resolution. The dynamic threshold gating suppresses high-frequency noise pulses through an adaptive threshold adjustment mechanism, preserving the effective energy fluctuation characteristics.

[0051] Multi-source heterogeneous power data includes the voltage and current waveforms of the photovoltaic array, the charge and discharge curves of the energy storage battery (such as the state of charge and discharge rate), and the load power requirements. To achieve efficient processing, the continuous signals are first pulse-encoded using a spiking neural network (SNN). The input layer of the SNN consists of multiple neurons, each corresponding to a power parameter (such as photovoltaic voltage and battery current).

[0052] The core of dynamic threshold gating is to adaptively adjust the pulse trigger threshold to distinguish effective fluctuations from noise. The specific process is as follows:

[0053] Baseline threshold initialization: Based on historical data statistics, the initial threshold is set (for example, the PV voltage threshold is ±5% of the nominal value).

[0054] Real-time threshold adjustment: A sliding window (200ms width) is used to calculate the signal standard deviation, σ, and dynamically adjust the threshold to μ±3σ (μ is the window mean). For example, if the PV array experiences transient fluctuations due to cloud cover, the voltage standard deviation increases to 8%, and the threshold is automatically relaxed to ±24% to prevent false triggering.

[0055] Pulse triggering rule: When the signal change rate exceeds the threshold, a pulse is generated (coded as 1); otherwise, silence is generated (coded as 0). For example, if the battery current suddenly increases from 10A to 50A within 2 seconds (change rate of 20A / s), a high-frequency pulse sequence is triggered.

[0056] The resulting pulse space-time matrix has dimensions N × T, where N is the number of power source parameters (e.g., three each for photovoltaic, battery, and load) and T is the time step (with a resolution of 1 ms). For example, during a transient PV fluctuation, the pulse density in the corresponding row of the matrix increases significantly, reflecting the energy fluctuation characteristic.

[0057] Based on the pulse spatiotemporal matrix and the ambient temperature field distribution data, a multi-source signal phase difference metric tensor is constructed. A phase synchronization algorithm with pulse timing constraints is used to perform cross-modal time calibration, eliminating phase drift caused by photovoltaic transient fluctuations and battery aging, and outputting a phase-aligned multi-scale pulse tensor.

[0058] The ambient 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:

[0059] Multimodal timestamp alignment: Leveraging the 5G network’s precise clock synchronization protocols (such as IEEE 1588 PTP), the temperature data and the power pulse matrix’s timestamps are aligned to the microsecond level.

[0060] Phase difference calculation: Mutual information analysis is performed on the correlation between each power supply parameter and the temperature field, constructing 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 lag of 300ms, indicating the delayed impact of temperature changes on photovoltaic output.

[0061] Pulse timing constraints: Dynamic time warping (DTW) is used to align pulse trains of different modalities. For example, battery aging causes its response delay to increase by 50ms. DTW eliminates phase differences by stretching or compressing the time axis.

[0062] The phase synchronization algorithm uses wavelet coherence analysis:

[0063] Wavelet transform: Continuous wavelet transform (Mexican Hat wavelet) is performed on the pulse matrix and temperature field data to generate the time-frequency energy spectrum.

[0064] Coherence detection: Calculate the cross-wavelet coherence spectrum and identify significant coherence regions (e.g., the coherence between photovoltaic voltage and temperature in the 1 Hz frequency band is > 0.8).

[0065] Phase correction: Linear interpolation is used to compensate for phase differences in the coherent region. For example, if the temperature field change is detected to be 20ms ahead of the photovoltaic fluctuation, the photovoltaic pulse sequence will be shifted forward by 20ms.

[0066] The output multi-scale pulse tensor has dimensions N × S × T, where S is the scale parameter (e.g., 0.1 Hz, 1 Hz, and 10 Hz frequency bands) and T is the synchronized time step. In this tensor, the pulse events of the photovoltaic, battery, and load are strictly aligned in the time-frequency domain, providing a consistent benchmark for subsequent analysis.

[0067] The multi-scale pulse tensor is input into the multi-scale pulse fusion tensor decomposition algorithm to construct the pulse energy density spectrum in the time-frequency hybrid domain. The core characteristic sub-tensors of photovoltaic, energy storage and load are decomposed through the time-frequency constraint mechanism parameterized by the convolution kernel, and the time-frequency aliasing interference is suppressed.

[0068] The Multiscale Pulse Fusion Tensor Decomposition (MPFTD) algorithm combines the advantages of tensor decomposition and convolutional neural networks. The specific steps are as follows:

[0069] Pulse energy density spectrum construction: A sliding window integration (window width 100ms, step size 10ms) is performed on the pulse matrix at each scale (frequency band), and the pulse density (number of pulses / window) is calculated as an energy indicator. For example, in the 10Hz frequency band, the PV pulse density reflects the fluctuations caused by the string cascade effect.

[0070] Convolution kernel parameterization: Design three sets of learnable convolution kernels, corresponding to the characteristic patterns of photovoltaic, battery, and load respectively:

[0071] Photovoltaic core: size 3×3 (time × scale), capturing wave patterns caused by cloud movement;

[0072] Battery core: size 5×1, extracts the slow-changing features of the charge and discharge process;

[0073] Load core: size 1×5, identifies sudden load demands.

[0074] Feature separation and reconstruction: Convolution operations are used to extract the core features of each power source, and non-negative matrix factorization (NMF) is used to constrain the non-negativity of the feature sub-tensors. For example, in the photovoltaic sub-tensor, high-frequency components (>5Hz) are suppressed, preserving minute-level fluctuation trends.

[0075] Time-frequency aliasing suppression is achieved through frequency domain masking:

[0076] Aliasing detection: Calculate the cross power spectral density (CPSD) between each scale to identify aliasing areas (such as harmonic interference in the 1Hz and 10Hz frequency bands).

[0077] Mask generation: An exponential decay weight (attenuation factor 0.8) is applied to aliased regions to reduce their contribution. For example, during battery charging and discharging, a 1Hz current ripple aliases with a 10Hz PWM switching noise. The mask attenuates its energy by 64%.

[0078] The decomposed core feature sub-tensors include:

[0079] Photovoltaic sub-tensor: dimension N_pv×S_pv×T, reflecting irradiance changes and string mismatch;

[0080] Battery sub-tensor: dimension N_bat×S_bat×T, representing SOC change and internal resistance aging;

[0081] Load sub-tensor: Dimension N_load × S_load × T, describes the load step and harmonic distortion.

[0082] 1. Photovoltaic system parameters

[0083] N_pv (photovoltaic parameter dimension)

[0084] Meaning: The number of characteristic parameters of the photovoltaic array, usually including: output voltage (V_pv), output current (I_pv), power (P_pv), and string mismatch rate (%).

[0085] Example: If three types of parameters (voltage, current, and power) are monitored, then N_pv=3.

[0086] S_pv (photovoltaic band scale)

[0087] Meaning: The number of frequency bands used for photovoltaic characteristic analysis to capture fluctuations at different time scales:

[0088] Low frequency (0.1 Hz): reflects minute-level fluctuations caused by cloud movement;

[0089] Medium frequency (1Hz): Second-level fluctuations caused by the cascade effect of series connection;

[0090] High frequency (10 Hz): millisecond-level disturbances such as inverter switching noise.

[0091] Example: If three frequency bands are divided, S_pv=3.

[0092] 2. Energy storage battery parameters

[0093] N_bat (battery parameter dimension)

[0094] Meaning: Key monitoring parameters of the battery system, such as state of charge (SOC, %), charge and discharge current (A), internal resistance (mΩ), and temperature (°C).

[0095] Example: If SOC, current, and temperature are selected, then N_bat=3.

[0096] S_bat (battery band scale)

[0097] Meaning: Frequency band division for battery characteristic analysis, focusing on different dynamic processes:

[0098] Ultra-low frequency (0.01Hz): SOC changes slowly (hourly);

[0099] Low frequency (0.1Hz): charge and discharge cycle (minutes);

[0100] High frequency (1Hz): current ripple (seconds).

[0101] Example: If two frequency bands (SOC and current ripple) are analyzed, S_bat=2.

[0102] 3. Load parameters

[0103] N_load (load parameter dimension)

[0104] Meaning: Core parameters of the load end, such as active power (kW), reactive power (kVar), harmonic distortion (THD%), and sudden load increase / reduction flag (Boolean value).

[0105] Example: If power and harmonics are monitored, N_load = 2.

[0106] S_load (load band scale)

[0107] Meaning: Frequency band division of load dynamic behavior:

[0108] Fundamental frequency (50 / 60Hz): steady-state power demand;

[0109] Medium frequency (100-500Hz): motor starting transient;

[0110] High frequency (>1kHz): switching device harmonics.

[0111] Example: If you are concerned about the fundamental frequency and harmonics, then S_load = 2.

[0112] Pulse energy accumulation and spatial topological mapping are performed on the core characteristic sub-tensor, and a multi-source energy dynamic topological map is constructed based on a pulse-triggered dynamic weight allocation algorithm to generate a spatiotemporally aligned power state joint tensor; wherein, the topological map reflects the dynamic coupling relationship between energy nodes in real time through a pulse-triggered edge weight update mechanism.

[0113] Pulse energy accumulation uses time exponentially weighted moving average (EWMA):

[0114] Attenuation factor setting: Select the attenuation speed according to the characteristic frequency band. The attenuation factor of low-frequency characteristics (such as battery SOC) is λ=0.9 (slow attenuation), and the attenuation factor of high-frequency characteristics (such as load mutation) is λ=0.5 (fast attenuation).

[0115] Energy integration: Performs a weighted summation of each feature sub-tensor along the time dimension. For example, the energy integral of the photovoltaic sub-tensor reflects the cumulative irradiation over 15 minutes.

[0116] The spatial topology mapping is based on the physical connection relationship of the power nodes:

[0117] Node definition: Photovoltaic arrays, battery packs, and load centers are abstracted into topological nodes. Each node contains location coordinates (such as GPS data) and electrical parameters (such as rated power).

[0118] Edge weight initialization: Set the initial weight based on the electrical distance (such as cable impedance). The lower the impedance, the higher the weight.

[0119] The dynamic weight assignment algorithm drives edge weight updates through pulse events:

[0120] Pulse trigger condition: When the characteristic energy of a node exceeds the threshold (such as photovoltaic node energy > 100kW·s), the weight update is triggered.

[0121] Weight update rule:

[0122] Forward coupling: If a pulse event at node A causes the energy of node B to increase (e.g., increased photovoltaic power generation prompts battery charging), the edge weight increases by Δw=0.1;

[0123] Negative coupling: If the pulse of node A suppresses the energy of node B (for example, a sudden increase in load causes accelerated battery discharge), the edge weight is reduced by Δw = 0.05.

[0124] The resulting power state joint tensor has dimensions of N × S × T × C, where C represents 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 increases from 0.7 to 0.9, triggering an adjustment to the coordinated power supply strategy.

[0125] Example application: In a microgrid scenario, the photovoltaic array output suddenly drops 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. The system immediately initiates battery discharge compensation to ensure continuous load power supply.

[0126] S202: Input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply shape based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and output an energy dynamic balance vector including a power supply margin prediction;

[0127] Specifically, the power state joint tensor can be input into a physically constrained adversarial prediction network, a potential representation of the multi-source power supply shape can be constructed through a generator network, and a frequency domain sparsity constraint can be imposed using a discriminator network 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;

[0128] 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 uses a deep residual network (ResNet) architecture, taking as input a joint tensor of power states (dimensions: time steps × number of energy nodes × feature dimensions, e.g., 100 × 5 × 64), and outputting a potential power supply shape representation. The network structure consists of four residual blocks, each consisting of a 1D convolutional layer (kernel size 3, stride 1), batch normalization (BatchNorm), and a gated linear unit (GLU). Residual skip connections mitigate the vanishing gradient problem. For example, transient fluctuations in photovoltaic arrays (such as power drops caused by cloud cover) are captured by residual blocks, capturing high-frequency variations. The slow charging and discharging characteristics of energy storage batteries are preserved through skip connections, preserving low-frequency information.

[0129] The discriminator is designed as a frequency domain-time domain dual-channel structure:

[0130] Time domain channel: 3-layer convolutional network (kernel size 5, number of channels 32 / 64 / 128) to extract time domain features;

[0131] Frequency domain channel: Perform fast Fourier transform (FFT) on the input data to convert the time domain signal into a frequency domain energy spectrum. The key frequency bands are extracted through bandpass filtering (0.1Hz~10Hz), and then three layers of convolution are performed.

[0132] Fourier domain adversarial constraints are implemented through gradient backpropagation in the frequency domain. The discriminator imposes sparsity constraints on the frequency domain energy spectrum output by the generator: steady-state components (such as daily average photovoltaic power fluctuations) are concentrated in the low-frequency band (0.1Hz-1Hz) and must be smoothed; transient components (such as sudden load changes) are distributed in the high-frequency band (1Hz-10Hz) and must highlight peak features. For example, if the generator fails to isolate the 10Hz high-frequency component of a sudden load change, the discriminator will force the generator to correct the problem through frequency domain gradient penalties (such as increasing the loss weight of the high-frequency band).

[0133] Physical constraints are implemented through an energy conservation loss function, which calculates the mean square error (MSE) between the total power output of the generator and the actual power input. For example, if the predicted photovoltaic power at a certain moment is 5kW and the energy storage output is 3kW, the total power prediction of the generator must strictly equal 8kW (the load demand). If the error exceeds 5%, a penalty is triggered.

[0134] The final generated adversarial feature vector (dimension 256) contains a decoupled representation of the steady-state baseline and transient disturbances. For example, the first 128 dimensions of the vector encode the steady-state coupling relationship between photovoltaics and energy storage, and the last 128 dimensions capture the spatiotemporal pattern of load mutations.

[0135] Performing frequency-domain sparse residual focusing processing on the adversarial feature vector, extracting the load mutation residual component through a bandpass filter kernel with adaptive bandwidth, and using a steady-state fluctuation suppression algorithm to eliminate background noise, outputting a high-resolution load mutation feature spectrum;

[0136] 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:

[0137] Frequency domain transformation: Perform 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, for example, 100 × 50 × 256).

[0138] Adaptive bandpass filtering: Based on the historical frequency distribution of load mutations (e.g., 90% of mutations are concentrated between 2Hz and 8Hz), a Gaussian filter kernel with adjustable bandwidth is designed. The filter kernel's initial center frequency is 5Hz, and its bandwidth is 2Hz. This kernel dynamically adjusts based on real-time spectrum energy. If a sudden energy increase is detected in a certain frequency band (e.g., a 30% increase in energy at 3Hz), the center frequency is shifted toward that frequency, and the bandwidth is narrowed to 1Hz to improve resolution. For example, if a 3.5Hz load spike caused by the activation of a cluster of air conditioners is detected, the filter kernel automatically adjusts to 3.5Hz±0.5Hz.

[0139] Residual Extraction: The filtered time-frequency matrix is differentiated from the original time-frequency matrix to obtain the residual component. Steady-state fluctuations (such as PV power changes caused by gradually changing sunlight) are suppressed due to out-of-band fluctuations, while sudden load fluctuations are retained due to in-band enhancements.

[0140] Background noise suppression: Moving average filtering (window length 1 second) is used to eliminate random noise in the residuals, and further denoising is performed through threshold gating (for example, residuals with energy lower than 2σ of the mean are set to zero).

[0141] Steady-state fluctuation suppression algorithm is based on multi-scale analysis of wavelet transform:

[0142] Perform 5-layer wavelet decomposition on the original eigenvector (the mother wavelet is db4) to extract the approximate coefficients (low-frequency steady-state components) and detail coefficients (high-frequency noise);

[0143] The high-frequency noise in the detail coefficient that is not related to the load mutation (such as sensor quantization error) is set to zero, subtracted from the residual after reconstructing the signal, and improves the signal-to-noise ratio.

[0144] In the final output load mutation feature spectrum (dimensions 100×50), each time-frequency unit is labeled with the mutation intensity (0–1). For example, an intensity of 0.8 in the 5Hz frequency band at a certain moment indicates the detection of a load surge event lasting 200ms.

[0145] Based on the dynamic game framework, a game payoff matrix of load mutation and steady-state fluctuation is constructed. The attention weights of the two types of features are dynamically allocated through an adversarial attention mechanism. The adversarial attention mechanism optimizes the sparsity of the attention mask through the game strategy gradient update mechanism.

[0146] The dynamic game framework models load mutation (offensive party) and steady-state fluctuation (defensive party) as two sides of the game, and the goal is to balance the impact of the two in power margin prediction.

[0147] Profit matrix construction:

[0148] The profit function of the load mutation party is a positive correlation between the mutation intensity and the prediction error (the stronger the mutation, the larger the prediction error and the higher the profit);

[0149] The payoff function for the steady-state volatility side is the negative correlation between the forecast result and the smoothed benchmark (the smoother the volatility, the smaller the forecast error and the higher the payoff).

[0150] The dimensions of the profit matrix are mutation intensity level (5 levels) × volatility smoothing level (5 levels). For example, when the mutation intensity is level 3 (medium) and the volatility smoothing level is level 2 (mild volatility), the mutation side’s profit is 0.6 and the volatility side’s profit is 0.4.

[0151] Adversarial Attention Mechanism:

[0152] 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;

[0153] Mask sparsity is optimized using a policy gradient algorithm (PPO). During training, if the prediction error in a region with high attention weight (e.g., at the moment of a mutation) decreases, the weight of that region is increased; conversely, if the error in a region with low attention weight increases, the weight is decreased. The learning rate is set to 0.001, and the KL divergence threshold is 0.02 to prevent over-tuning.

[0154] Dynamic weight assignment example:

[0155] When a continuous load mutation is detected (such as the startup of a data center server cluster), the attention mask weight is increased to 0.9 during the mutation period (t = 10s-15s) to suppress the influence of steady-state fluctuations;

[0156] During the stable operation phase (t=30s~60s), the weight is reduced to 0.2, giving priority to the slow fluctuations of photovoltaic / energy storage.

[0157] During the training process, the two players explored the strategy space through Monte Carlo simulation. After 1,000 iterations, the sparsity of the attention mask (the proportion of non-zero weights) was optimized from the initial 80% to 30%, significantly improving the detection weight of key mutation events.

[0158] The load mutation characteristic spectrum is integrated in the time domain and energy is accumulated. Combined with the state of charge constraint of the energy storage battery, a power supply margin prediction model is used to generate an energy dynamic balance vector including the power supply capacity for several seconds in the future. The power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

[0159] The power margin prediction model uses a hybrid architecture of a temporal convolutional network (TCN) and a long short-term memory network (LSTM):

[0160] Time domain integration: Integrate the load mutation characteristic spectrum according to the time window (such as 1 second) to generate a mutation energy sequence (dimension 100×1);

[0161] Energy accumulation: Superimpose the predicted photovoltaic power, real-time energy storage output and sudden energy to obtain the net power supply capacity sequence;

[0162] State of charge (SOC) constraint: When the energy storage SOC is lower than 20%, the discharge power is limited to 50% of the rated value; when the SOC is higher than 80%, the charging power is limited to 30%.

[0163] Adversarial example augmentation works by injecting two types of perturbations into the training data:

[0164] Load mutation disturbance: randomly insert false mutations with an amplitude of ±20% and a duration of 0.1 to 1 second;

[0165] Sensor noise: Gaussian noise (μ=0, σ=0.05) is added to simulate signal acquisition error.

[0166] Model training adopts the Curriculum Learning strategy:

[0167] Primary stage: Use clean data for training, learning rate 0.001, batch size 64;

[0168] Advanced stage: gradually increase the proportion of adversarial samples to 50%, reduce the learning rate to 0.0001, and use a batch size of 32.

[0169] The prediction output is an energy dynamic balance vector (dimension 10), with each element corresponding to the power margin (in kW) for the next 1 second. For example, the vector [5.2, 4.8, 4.5, ..., 3.0] represents the additional power margin available per second for the next 10 seconds. When the margin falls below a threshold (e.g., 1 kW) in a given second, emergency energy storage discharge or load shedding is triggered.

[0170] Sample Application:

[0171] Scenario 1: The load suddenly increases by 8kW. The current PV output is 5kW, and the available energy storage power is 3kW. The margin vector shows that the margins for the next three seconds will be 0kW, -2kW, and -5kW, respectively. The system immediately starts the backup diesel generator.

[0172] Scenario 2: The steady-state load fluctuates by ±1 kW, the margin vector stabilizes at 2 kW to 3 kW, and the system maintains the current strategy.

[0173] After adversarial training, the model's prediction error in a noisy environment was reduced from 15% to 6%, and the decision-making stability under the critical SOC state (20%~30%) was improved by 40%.

[0174] S203, performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a power supply strategy candidate set using a deep strategy network driven by Monte Carlo tree search, screening a global optimal solution using a manifold projection constrained Nash equilibrium solver, and outputting an interference-resistant multi-source collaborative power supply strategy matrix;

[0175] Specifically, a power supply strategy candidate tree can be generated based on the energy dynamic balance vector using a deep strategy network driven by Monte Carlo tree search, wherein the deep strategy network simultaneously evaluates the short-term benefits and long-term stability of the strategy through a strategy value dual-headed network;

[0176] The deep policy network driven by Monte Carlo Tree Search (MCTS) aims to generate a candidate power supply policy tree that balances short-term benefits (such as immediate power supply stability) with long-term stability (such as balanced equipment life). The network architecture adopts a dual-head design:

[0177] Strategy head: Outputs the probability distribution of each power supply strategy under the current state (e.g., the probability of selecting photovoltaic power supply is 0.6, energy storage power supply is 0.3, and mains power supplement is 0.1). The dimension is consistent with the action space (assuming three power supply modes);

[0178] Value Head: Predicts the long-term value of the strategy (value range [-1,1], positive values indicate benefits, negative values indicate potential risks).

[0179] MCTS Process:

[0180] Selection phase: Starting from the root node, child nodes are selected using the UCB1 (Upper Confidence Bound) formula to balance exploration and exploitation. For example, the UCB value of a 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.

[0181] Expansion phase: When encountering unexplored nodes, new nodes are expanded. For example, when the state of charge (SOC) of the energy storage battery is less than 20%, the "Switch to Mains" node is expanded.

[0182] Simulation phase: Use random strategies to simulate to the end (e.g., predict the power supply status in the next 30 seconds) and calculate simulation rewards (e.g., number of power outages, battery loss rate).

[0183] Backtracking phase: Updates the number of visits and mean value of each node on the path. For example, if a simulation shows that switching to photovoltaics causes a voltage drop, the corresponding node value will be reduced by 0.2.

[0184] Training parameters:

[0185] The tree search depth is set to 5 layers (corresponding to 5 decision steps), and each round of iteration is 100 times;

[0186] The input dimension of the policy network is 128 (the feature dimension of the energy dynamic balance vector), and the hidden layer is 256-128-64;

[0187] The learning rate is 0.001, the batch size is 32, and the Adam optimizer is used.

[0188] Example: In a scenario where photovoltaic power output suddenly drops, MCTS generates the following candidate strategy branches:

[0189] 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);

[0190] Branch 2: Reduce load by 10% and use the mains (short-term benefit +0.5, long-term value +0.6);

[0191] Scenario 3: Enable backup diesel generators (short-term benefit +0.2, long-term value -0.7 due to high carbon emission costs).

[0192] Performing high-dimensional strategy space manifold projection on the power supply strategy candidate tree, extracting strategy core features through a geodesic distance-constrained dimensionality reduction algorithm, and generating a low-dimensional compact strategy manifold; wherein the dimensionality reduction algorithm suppresses strategy conflicts through manifold curvature optimization;

[0193] The manifold projection of a high-dimensional strategy space (e.g., 100 dimensions) aims to extract the core features of the strategy (e.g., power mode switching frequency, resource consumption rate) to facilitate subsequent optimization. The specific process is as follows:

[0194] Manifold construction:

[0195] The UMAP (Uniform Manifold Approximation and Projection) algorithm was used, with n_neighbors = 15 (to control sensitivity to local structure) and min_dist = 0.1 (to ensure that the points are tightly distributed after projection).

[0196] The input is a policy feature vector (including parameters such as power supply mode, power allocation ratio, and switching delay), with a dimension of 100. The dimension is reduced to a 3D visualization space.

[0197] Geodesic distance constraint:

[0198] The geodesic distance (shortest path on the manifold) between strategic points is calculated using the Floyd-Warshall algorithm, replacing the Euclidean distance to maintain the topological structure;

[0199] For example, the geodesic distance between strategy A (PV priority) and strategy B (storage priority) is 1.2, while the distances between strategy C (mains power dependence) and the two are 2.3 and 1.8, respectively, reflecting that strategies A and B are more similar.

[0200] Curvature Optimization:

[0201] Detect local curvature of the manifold (via the Ricci flow algorithm) and impose penalties on areas of high curvature (strategy conflict areas);

[0202] If the curvature radius of a region is < 0.5 (indicating dense strategy conflicts), strategy confusion is avoided by increasing the projection interval (e.g., enforcing the strategy spacing ≥ 0.3).

[0203] Example:

[0204] After projection, three clusters are formed in the policy manifold:

[0205] Cluster 1: Aggressive strategy (high PV utilization, low battery SOC protection);

[0206] Cluster 2: Conservative strategy (mains power dependence, high battery redundancy);

[0207] Cluster 3: Hybrid strategy (dynamic switching, medium risk).

[0208] Through curvature optimization, the boundary between cluster 1 and cluster 2 is clear, and the policy conflict is reduced by 30%.

[0209] A multi-objective dynamic game model is constructed on the low-dimensional compact strategy manifold, a Nash equilibrium solver optimized by a hybrid strategy is used to conduct a strategy-benefit game, a shadow price iterative algorithm is used to screen the global optimal solution, and an anti-interference multi-source collaborative power supply strategy matrix is output;

[0210] The multi-objective dynamic game model abstracts the optimization goal of the power supply strategy into the competition of interests among multiple game parties (such as "stability party", "economic party" and "environmental protection party"), and finds the optimal compromise point through Nash equilibrium.

[0211] Profit matrix construction:

[0212] Stability benefit: the reciprocal of the power outage time (in seconds), with a weight of 0.5;

[0213] Economic benefit: negative value of total power supply cost (yuan), weight 0.3;

[0214] Environmental benefit: negative value of carbon emissions (kg), weight 0.2.

[0215] For example, the payoff vector of strategy A is (0.8, -500, -200), and that of strategy B is (0.6, -300, -50).

[0216] Mixed strategy Nash equilibrium solution:

[0217] Use Lemke-Howson algorithm to solve equilibrium points and support asymmetric game matrices;

[0218] The shadow price iteration algorithm is used to address constraints (such as battery SOC ≥ 20%), dynamically adjusting resource allocation prices through Lagrange multipliers. For example, when the SOC approaches 20%, the shadow price of energy storage power supply increases from 0.5 yuan / kWh to 1.2 yuan / kWh, thus discouraging overuse.

[0219] Global optimal solution screening:

[0220] Pareto front generation: The non-dominated solution set is screened using the NSGA-II algorithm, with a population size of 200 and 50 iterations;

[0221] Decision maker preference injection: TOPSIS (Top-Inferior Solution Distance Method) combined with weighted preferences is used to select the final solution. For example, if stability is preferred, the strategy closest to the ideal solution (interruption time = 0, cost = 0, carbon emissions = 0) is selected. Example output:

[0222] The optimal strategy matrix includes the information shown in Table 1:

[0223] Table 1

[0224]

[0225] The hybrid strategy convergence in dynamic game theory is verified for the multi-source collaborative power supply strategy matrix, and the superlinear convergence rate under multi-objective conflict is proved through Lyapunov stability analysis. The strategy search depth is dynamically adjusted based on the verification results.

[0226] Lyapunov stability analysis is used to verify the convergence of strategies in dynamic environments and ensure that strategy adjustments do not cause oscillations or divergence.

[0227] Lyapunov function construction:

[0228] Define the system energy function V(x) = ∑(target deviation)^2, for example, V = 0.5(interruption time - time target)^2 + 0.3(cost - cost target)^2 + 0.2*(carbon emissions - emission target)^2;

[0229] When V decreases over time, the system is stable.

[0230] Convergence verification:

[0231] Calculate dV / dt by numerical differentiation. If dV / dt<-k*V (k>0), superlinear convergence is satisfied.

[0232] For example, after a certain strategy is executed, V drops from the initial value of 1.2 to 0.3 (Δt=5 seconds), dV / dt=-0.18<-0.1*0.3=-0.03, which meets the superlinear convergence condition.

[0233] Strategy search depth adjustment:

[0234] If the convergence rate is lower than the threshold (e.g., Δt > 10 seconds), increase the MCTS search depth (from 5 layers to 7 layers);

[0235] If oscillation is detected (dV / dt signs are alternating), the backtracking mechanism is enabled to force a switch to the historical optimal strategy. Example:

[0236] The initial strategy search depth is 5 layers, and the convergence time is 8 seconds;

[0237] After detecting that dV / dt = -0.05 after a certain strategy execution 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;

[0238] After the adjustment, system stability was improved and the power supply interruption fluctuation rate was reduced from ±15% to ±5%.

[0239] Monte Carlo Tree Search:

[0240] The search depth is 5 layers, the iterations are 100 times, and 30 candidate strategies are generated;

[0241] Policy evaluation takes ≤50ms (based on the NVIDIA Jetson AGX Xavier edge computing platform).

[0242] Manifold Projection:

[0243] After UMAP dimensionality reduction, the strategy features are 3-dimensional, and the cluster purity is ≥85%;

[0244] The error of geodesic distance calculation is <5% (compared to Euclidean distance).

[0245] Nash equilibrium solution:

[0246] Lemke-Howson algorithm solution time ≤ 20ms;

[0247] The shadow price is updated at a frequency of 1Hz to adapt to real-time resource fluctuations.

[0248] Lyapunov verification:

[0249] Superlinear convergence rate threshold k=0.1;

[0250] The response time for dynamically adjusting the search depth is ≤ 100ms.

[0251] Through the above implementation, the multi-source collaborative power supply strategy matrix has reduced the power supply interruption rate to 0.5% / month in actual measurements, saved 15% in comprehensive costs, and reduced carbon emissions by 20%, which is significantly better than traditional rule control methods.

[0252] S204, input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

[0253] Specifically, the multi-source collaborative power supply strategy matrix can be input into the edge computing-driven digital twin verification module, and the battery thermal runaway and photovoltaic shading mutation can be simulated through the fault scenario adversarial distillation technology to generate a strategy robustness evaluation vector; wherein, the digital twin verification module optimizes the simulation parameters through the dynamic weight mirror update mechanism;

[0254] The core of the digital twin verification module is to verify the robustness of the power supply strategy in extreme scenarios through high-fidelity simulation models. First, based on the multi-source coordinated power supply strategy matrix (such as the photovoltaic output allocation ratio, energy storage charging and discharging thresholds, etc.), a digital mirror of the battery and photovoltaic components is constructed. The battery thermal runaway scenario is simulated using an electrochemical-thermal coupling model: when the strategy requires the battery to discharge at a 2C rate (i.e., twice the capacity current), the twin model calculates the internal temperature distribution in real time. If the temperature of a battery cell exceeds 60°C (the thermal runaway threshold), an early warning is triggered. The photovoltaic shading mutation scenario uses a ray tracing algorithm to simulate cloud cover. 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.

[0255] Fault scenario adversarial distillation technology uses a generative adversarial network (GAN) to generate extreme operating condition data:

[0256] Generator: Input is historical normal operating condition data (such as battery SOC = 80%, ambient temperature 25°C), and output is simulated thermal runaway time series data (such as a 10-second curve of temperature rising from 25°C to 80°C);

[0257] The discriminator takes real thermal runaway data and generated data as input and outputs a probability of authenticity. During training, the generator optimizes the distributional match between generated and real data using the Wasserstein loss function, with a gradient penalty coefficient of λ = 10.

[0258] The dynamic weight mirror update mechanism dynamically adjusts simulation parameters through reinforcement learning:

[0259] Parameter space: including battery internal resistance (0.05~0.2Ω), photovoltaic conversion efficiency (15%~22%);

[0260] Reward function: the degree to which the simulation results match the physical constraints (e.g., reward +1 if the temperature prediction error is <±3°C);

[0261] Strategy update: The PPO algorithm (learning rate 0.0003, batch size 64) is used. 100 parameter combinations are generated in each iteration, and the top 10 with the highest rewards are selected to update the image weights.

[0262] The final generated policy robustness evaluation vector contains the following dimensions:

[0263] Thermal runaway suppression score (0-1, based on the percentage of time the temperature exceeds the limit);

[0264] PV switching delay (unit: ms), the speed at which the strategy responds to sudden changes in shading;

[0265] Energy storage cycle life loss (percentage, the impact of the strategy on battery aging). For example, a strategy with a score of [0.92, 45ms, 0.8%] indicates excellent performance in thermal runaway scenarios, but switching latency needs to be optimized.

[0266] Based on the strategy robustness evaluation vector, a dynamic weight allocation model for the 5G multi-hop transmission protocol is constructed. The priority weights of control instructions are adjusted in real time based on the base station load rate and channel quality to generate an anti-congestion instruction distribution sequence.

[0267] The goal of the dynamic weight allocation model is to optimize the transmission priority of control instructions in 5G multi-hop networks, ensuring that critical instructions (such as energy storage emergency shutdown) are transmitted first. The weight allocation rules are based on the following real-time parameters:

[0268] Base station load rate (CPU utilization, threshold 80%): When the load rate is >80%, the weight of non-critical instructions (such as photovoltaic power fine-tuning) is reduced;

[0269] Channel quality (SNR, in dB): When SNR < 15dB, increase the retransmission command weight;

[0270] The urgency of the command (level 1 to 5, with level 5 being the highest).

[0271] Dynamic weight calculation uses Q-learning algorithm:

[0272] State space: 3D vector (load rate, SNR, urgency);

[0273] Action space: weight adjustment range (-0.3~+0.3, step size 0.1);

[0274] Reward function:

[0275]

[0276] During training, the action space is explored using an ε-greedy strategy (ε=0.1), with a learning rate α=0.01 and a discount factor γ=0.9.

[0277] Anti-congestion instruction distribution sequence generation:

[0278] Prioritization: Arrange instructions in descending order of weight, for example:

[0279] Instruction A (energy storage shutdown, weight 0.95);

[0280] Instruction B (photovoltaic switching, weight 0.85);

[0281] Resource reservation: 20% of the transmission bandwidth is reserved for high-weight instructions;

[0282] Dynamic backoff: If the channel is congested (packet loss rate > 5%), low-weight commands are delayed by 1 to 3 time slots.

[0283] For example, when the base station load rate reaches 85% and the SNR = 12dB, the strategy increases the weight of the energy storage shutdown instruction from 0.9 to 0.95 and allocates a dedicated time slot to ensure end-to-end latency < 50ms.

[0284] Performing multi-base station coordinated control conflict detection on the instruction distribution sequence and using an implicit gradient compensation algorithm to eliminate policy execution deviations, wherein the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence by backpropagation of the conflict residual;

[0285] Multi-base station collaborative conflict detection is achieved through a distributed consensus algorithm:

[0286] Conflict detection: Each base station periodically (every 100ms) broadcasts the hash value (SHA-256) of the local command queue. If the hashes are inconsistent, a conflict is determined.

[0287] Conflict location: Use a binary search method to trace the last 10 commands and identify the source of the conflict (for example, base station A and base station B send PV load reduction commands at the same time).

[0288] The core of the implicit gradient compensation algorithm is to correct the instruction parameters through back propagation:

[0289] Residual calculation: Measure the deviation Δ=3kW between the actual performance (e.g., a 5kW drop in PV power) and the expected target (a drop of 8kW);

[0290] Gradient backpropagation: Calculate the contribution of each link layer by layer along the control link (base station → edge controller → inverter), for example:

[0291] Base station command coding error contributes 30%;

[0292] Channel noise contributes 50% to packet loss;

[0293] Inverter response delay contributes 20%;

[0294] Parameter correction: Adjust the command parameters according to the contribution ratio. For example, correct the target power from 8kW to 8kW + 3kW × 0.3 = 8.9kW.

[0295] Dynamic correction example: A conflict caused the energy storage charging instruction amplitude to deviate by 10A. The algorithm reduced the deviation to 0.5A through three iterations (learning rate 0.1) and simultaneously adjusted the phase synchronization error from 15° to 2°.

[0296] The corrected control instructions are timestamped and synchronized through multi-base station collaborative delay balancing technology, and a sliding window-constrained instruction buffer mechanism is used to ensure end-to-end delay is less than 1ms, ultimately outputting a multi-source intelligent control instruction set with low latency and high reliability.

[0297] Delay equalization technology achieves microsecond-level synchronization based on the IEEE 1588 Precision Time Protocol (PTP):

[0298] Timestamp synchronization: The master base station periodically sends synchronization messages, and the slave base station calculates the link delay (such as transmission delay ±0.1ms) and dynamically compensates for clock offset;

[0299] Delay equalization: A least mean square (LMS) filter is used to predict link jitter. For example, if the delay of a link fluctuates between 0.5 and 1.2 ms, the filter will stabilize it to 0.8 ± 0.05 ms.

[0300] Sliding window buffer mechanism:

[0301] Window size: 10ms, accommodating 5 instruction frames (2ms per frame);

[0302] Dynamic adjustment: If the delay of three consecutive frames is greater than 0.9ms, the window is reduced to 8ms to speed up processing;

[0303] Redundancy elimination: Outdated instructions (such as those with a delay greater than 1ms) are directly discarded to ensure real-time performance.

[0304] Final instruction set generation:

[0305] Timestamp alignment: All instructions are tagged with a unified clock (such as GPS time source);

[0306] Priority execution: high-weight instructions are assigned to the front half of the window first;

[0307] Reliability verification: Instruction integrity is verified through CRC-32 checksum, with an error rate of <1e-6.

[0308] For example, a control instruction set containing 20 instructions achieved end-to-end latency distribution of 0.8ms ± 0.1ms after latency balancing, with a packet loss rate of 0.05%, meeting low latency and high reliability requirements. All instructions were issued via the 5G URLLC (Ultra-Reliable Low Latency Communication) channel, ensuring real-time and stable operation of the power grid.

[0309] It can be seen that the multi-source heterogeneous power supply data of the multi-source intelligent power manager is obtained, and the dynamic energy topology is constructed using the multi-scale pulse fusion tensor decomposition algorithm to generate a time-space aligned power supply state joint tensor; the power supply state joint tensor is input into the physically constrained adversarial prediction network, and an energy dynamic balance vector containing the power supply margin prediction is output; the energy dynamic balance vector is optimized by multi-objective dynamic game, and an anti-interference multi-source collaborative power supply strategy matrix is output; the multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing, and finally a multi-source intelligent control instruction set that meets low latency and high reliability is output, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0310] Another embodiment of the present invention provides a control system for a multi-source intelligent power manager based on 5G communication, see Figure 3 , the system may include:

[0311] Acquisition module 301 is used to acquire multi-source heterogeneous power data from a 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-synchronized time-frequency constraint mechanism, and generate a spatiotemporally aligned power state joint tensor;

[0312] Separation module 302 is configured to input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply shape based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and output an energy dynamic balance vector including a power supply margin prediction;

[0313] 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 strategy network driven by Monte Carlo tree search, screen the global optimal solution using a Nash equilibrium solver constrained by manifold projection, and output an interference-resistant multi-source collaborative power supply strategy matrix;

[0314] The output module 304 is used to input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

[0315] It can be seen that the multi-source heterogeneous power supply data of the multi-source intelligent power manager is obtained, and the dynamic energy topology is constructed using the multi-scale pulse fusion tensor decomposition algorithm to generate a time-space aligned power supply state joint tensor; the power supply state joint tensor is input into the physically constrained adversarial prediction network, and an energy dynamic balance vector containing the power supply margin prediction is output; the energy dynamic balance vector is optimized by multi-objective dynamic game, and an anti-interference multi-source collaborative power supply strategy matrix is output; the multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing, and finally a multi-source intelligent control instruction set that meets low latency and high reliability is output, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0316] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0317] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0318] S201, acquires multi-source heterogeneous power data from the multi-source intelligent power manager, uses a multi-scale pulse fusion tensor decomposition algorithm to build a dynamic energy topology, eliminates multi-source signal interference through a phase-synchronized time-frequency constraint mechanism, and generates a time-space aligned power state joint tensor;

[0319] S202: Input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply shape based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and output an energy dynamic balance vector including a power supply margin prediction;

[0320] S203, performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a power supply strategy candidate set using a deep strategy network driven by Monte Carlo tree search, screening a global optimal solution using a manifold projection constrained Nash equilibrium solver, and outputting an interference-resistant multi-source collaborative power supply strategy matrix;

[0321] S204, input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

[0322] It can be seen that the multi-source heterogeneous power supply data of the multi-source intelligent power manager is obtained, and the dynamic energy topology is constructed using the multi-scale pulse fusion tensor decomposition algorithm to generate a time-space aligned power supply state joint tensor; the power supply state joint tensor is input into the physically constrained adversarial prediction network, and an energy dynamic balance vector containing the power supply margin prediction is output; the energy dynamic balance vector is optimized by multi-objective dynamic game, and an anti-interference multi-source collaborative power supply strategy matrix is output; the multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing, and finally a multi-source intelligent control instruction set that meets low latency and high reliability is output, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0323] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0324] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0325] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0326] S201, acquires multi-source heterogeneous power data from the multi-source intelligent power manager, uses a multi-scale pulse fusion tensor decomposition algorithm to build a dynamic energy topology, eliminates multi-source signal interference through a phase-synchronized time-frequency constraint mechanism, and generates a time-space aligned power state joint tensor;

[0327] S202: Input the power state joint tensor into a physically constrained adversarial prediction network, reconstruct the multi-source collaborative power supply shape based on a dynamic game framework, separate the load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and output an energy dynamic balance vector including a power supply margin prediction;

[0328] S203, performing multi-objective dynamic game optimization on the energy dynamic balance vector, generating a power supply strategy candidate set using a deep strategy network driven by Monte Carlo tree search, screening a global optimal solution using a manifold projection constrained Nash equilibrium solver, and outputting an interference-resistant multi-source collaborative power supply strategy matrix;

[0329] S204, input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

[0330] It can be seen that the multi-source heterogeneous power supply data of the multi-source intelligent power manager is obtained, and the dynamic energy topology is constructed using the multi-scale pulse fusion tensor decomposition algorithm to generate a time-space aligned power supply state joint tensor; the power supply state joint tensor is input into the physically constrained adversarial prediction network, and an energy dynamic balance vector containing the power supply margin prediction is output; the energy dynamic balance vector is optimized by multi-objective dynamic game, and an anti-interference multi-source collaborative power supply strategy matrix is output; the multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing, and finally a multi-source intelligent control instruction set that meets low latency and high reliability is output, thereby improving the flexibility and response speed of power management and ensuring the stability and reliability of power supply.

[0331] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection 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 comprises: Acquire multi-source heterogeneous power data of a 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-synchronized time-frequency constraint mechanism, and generate a space-time aligned power state joint tensor; wherein, acquire multi-source heterogeneous power data of a multi-source intelligent power manager, and according to the original time domain signals of the photovoltaic array output waveform and the energy storage battery charge and discharge curve in the data, use the dynamic threshold gating of a pulse neural network for pulse encoding, convert the continuous waveform into a pulse trigger sequence, and generate a pulse space-time 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 measurement tensor based on the pulse space-time matrix and the ambient temperature field distribution data, perform cross-modal time calibration through a pulse timing constrained phase synchronization algorithm, eliminate the phase drift caused by photovoltaic transient fluctuations and battery aging, and output a phase-aligned multi-scale pulse tensor; The multi-scale pulse tensor is input into the multi-scale pulse fusion tensor decomposition algorithm, and the pulse energy density spectrum is constructed in the time-frequency hybrid domain. The core characteristic sub-tensors of photovoltaic, energy storage, and load are decomposed through the time-frequency constraint mechanism parameterized by the convolution kernel, and the time-frequency aliasing interference is suppressed. The core characteristic sub-tensors are subjected to pulse energy accumulation and spatial topological mapping. A multi-source energy dynamic topological map is constructed based on the pulse-triggered dynamic weight allocation algorithm to generate a spatiotemporally aligned power state joint tensor. The topological map reflects the dynamic coupling relationship between energy nodes in real time through the pulse-triggered edge weight update mechanism. The power state joint tensor is input into a physically constrained adversarial prediction network, and the multi-source collaborative power supply shape is reconstructed based on a dynamic game framework. The load mutation and steady-state fluctuation characteristics are separated by a residual-focused frequency domain decomposition algorithm, and an energy dynamic balance vector including power supply margin prediction is output; A multi-objective dynamic game optimization is performed on the energy dynamic balance vector. A deep strategy network driven by Monte Carlo tree search is used to generate a candidate set of power supply strategies. A Nash equilibrium solver with manifold projection constraints is used to screen the global optimal solution, and an interference-resistant multi-source collaborative power supply strategy matrix is output. The multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing. The strategy execution parameters are optimized based on the real-time feedback mechanism verified by digital twins. The multi-node control instructions are coordinated through the 5G multi-hop transmission protocol with dynamic weight allocation, and finally the multi-source intelligent control instruction set that meets the requirements of low latency and high reliability is output.

2. The method according to claim 1, characterized in that The power state joint tensor is input into the physically constrained adversarial prediction network, and the multi-source collaborative power supply shape is reconstructed based on a dynamic game framework. The load mutation and steady-state fluctuation characteristics are separated by a residual-focused frequency domain decomposition algorithm, and an energy dynamic balance vector containing a power supply margin prediction is output, including: The power state joint tensor is input into a physically constrained adversarial prediction network, a potential representation of the multi-source power supply shape is constructed through a generator network, and a frequency domain sparsity constraint is imposed using a discriminator network 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; Performing frequency-domain sparse residual focusing processing on the adversarial feature vector, extracting the load mutation residual component through a bandpass filter kernel with adaptive bandwidth, and using a steady-state fluctuation suppression algorithm to eliminate background noise, outputting a high-resolution load mutation feature spectrum; Based on the dynamic game framework, a game payoff matrix of load mutation and steady-state fluctuation is constructed. The attention weights of the two types of features are dynamically allocated through an adversarial attention mechanism. The adversarial attention mechanism optimizes the sparsity of the attention mask through the game strategy gradient update mechanism. The load mutation characteristic spectrum is integrated in the time domain and energy is accumulated. Combined with the state of charge constraint of the energy storage battery, a power supply margin prediction model is used to generate an energy dynamic balance vector including the power supply capacity for several seconds in the future. The power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

3. The method according to claim 2, characterized in that The energy dynamic balance vector is optimized through multi-objective dynamic game, a deep strategy network driven by Monte Carlo tree search is used to generate a candidate set of power supply strategies, a Nash equilibrium solver with manifold projection constraints is used to screen the global optimal solution, and an anti-interference multi-source collaborative power supply strategy matrix is output, including: Based on the energy dynamic balance vector, a deep policy network driven by Monte Carlo tree search is used to generate a power supply strategy candidate tree, wherein the deep policy network synchronously evaluates the short-term benefits and long-term stability of the strategy through a policy value dual-headed network; Performing high-dimensional strategy space manifold projection on the power supply strategy candidate tree, extracting strategy core features through a geodesic distance-constrained dimensionality reduction algorithm, and generating a low-dimensional compact strategy manifold; wherein the dimensionality reduction algorithm suppresses strategy conflicts through manifold curvature optimization; A multi-objective dynamic game model is constructed on the low-dimensional compact strategy manifold, a Nash equilibrium solver optimized by a hybrid strategy is used to conduct a strategy-benefit game, a shadow price iterative algorithm is used to screen the global optimal solution, and an anti-interference multi-source collaborative power supply strategy matrix is output; The hybrid strategy convergence in dynamic game theory is verified for the multi-source collaborative power supply strategy matrix, and the superlinear convergence rate under multi-objective conflict is proved through Lyapunov stability analysis. The strategy search depth is dynamically adjusted based on the verification results.

4. The method according to claim 3, characterized in that The multi-source collaborative power supply strategy matrix is input into the distributed collaborative control framework driven by edge computing. The strategy execution parameters are optimized based on the real-time feedback mechanism verified by digital twins. Multi-node control instructions are coordinated through the 5G multi-hop transmission protocol with dynamic weight allocation. Finally, a multi-source intelligent control instruction set that meets low latency and high reliability is output, including: The multi-source collaborative power supply strategy matrix is input into the edge computing-driven digital twin verification module, and the battery thermal runaway and photovoltaic shading mutation are simulated through the fault scenario adversarial distillation technology to generate a strategy robustness evaluation vector; wherein, the digital twin verification module optimizes the simulation parameters through a dynamic weight mirror update mechanism; Based on the strategy robustness evaluation vector, a dynamic weight allocation model for the 5G multi-hop transmission protocol is constructed. The priority weights of control instructions are adjusted in real time based on the base station load rate and channel quality to generate an anti-congestion instruction distribution sequence. Performing multi-base station coordinated control conflict detection on the instruction distribution sequence and using an implicit gradient compensation algorithm to eliminate policy execution deviations, wherein the implicit gradient compensation algorithm dynamically corrects the amplitude and phase of the control instructions in the instruction distribution sequence by backpropagation of the conflict residual; The corrected control instructions are timestamped and synchronized through multi-base station collaborative delay balancing technology, and a sliding window-constrained instruction buffer mechanism is used to ensure end-to-end delay is less than 1ms, ultimately outputting a multi-source intelligent control instruction set with low latency and high reliability.

5. A control system for a multi-source intelligent power manager based on 5G communication, characterized in that: The system comprises: An acquisition module is used to acquire multi-source heterogeneous power data of a 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-synchronized time-frequency constraint mechanism, and generate a time-space aligned power state joint tensor; wherein, the multi-source heterogeneous power data of the multi-source intelligent power manager is acquired, and according to the original time domain signals of the photovoltaic array output waveform and the energy storage battery charge and discharge curve in the data, the dynamic threshold gating of the pulse neural network is used for pulse encoding, and the continuous waveform is converted into a pulse trigger sequence to generate a pulse time-space 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 time-space matrix and the ambient temperature field distribution data, a multi-source signal phase difference measurement tensor is constructed, and a cross-modal time calibration is performed 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; The multi-scale pulse tensor is input into the multi-scale pulse fusion tensor decomposition algorithm, and the pulse energy density spectrum is constructed in the time-frequency hybrid domain. The core characteristic sub-tensors of photovoltaic, energy storage, and load are decomposed through the time-frequency constraint mechanism parameterized by the convolution kernel, and the time-frequency aliasing interference is suppressed. The core characteristic sub-tensors are subjected to pulse energy accumulation and spatial topological mapping. A multi-source energy dynamic topological map is constructed based on the pulse-triggered dynamic weight allocation algorithm to generate a spatiotemporally aligned power state joint tensor. The topological map reflects the dynamic coupling relationship between energy nodes in real time through the pulse-triggered edge weight update mechanism. a separation module for inputting the power state joint tensor into a physically constrained adversarial prediction network, reconstructing the multi-source collaborative power supply shape based on a dynamic game framework, separating load mutation and steady-state fluctuation characteristics through a residual-focused frequency domain decomposition algorithm, and outputting an energy dynamic balance vector including a power supply margin prediction; An optimization module is used to perform multi-objective dynamic game optimization on the energy dynamic balance vector, generate a candidate set of power supply strategies using a deep strategy network driven by Monte Carlo tree search, screen the global optimal solution through a Nash equilibrium solver constrained by manifold projection, and output an interference-resistant multi-source collaborative power supply strategy matrix; The output module is used to input the multi-source collaborative power supply strategy matrix into the distributed collaborative control framework driven by edge computing, optimize the strategy execution parameters based on the real-time feedback mechanism verified by digital twins, coordinate 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.

6. The system according to claim 5, characterized in that The separation module is specifically used for: The power state joint tensor is input into a physically constrained adversarial prediction network, a potential representation of the multi-source power supply shape is constructed through a generator network, and a frequency domain sparsity constraint is imposed using a discriminator network 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; Performing frequency-domain sparse residual focusing processing on the adversarial feature vector, extracting the load mutation residual component through a bandpass filter kernel with adaptive bandwidth, and using a steady-state fluctuation suppression algorithm to eliminate background noise, outputting a high-resolution load mutation feature spectrum; Based on the dynamic game framework, a game payoff matrix of load mutation and steady-state fluctuation is constructed. The attention weights of the two types of features are dynamically allocated through an adversarial attention mechanism. The adversarial attention mechanism optimizes the sparsity of the attention mask through the game strategy gradient update mechanism. The load mutation characteristic spectrum is integrated in the time domain and energy is accumulated. Combined with the state of charge constraint of the energy storage battery, a power supply margin prediction model is used to generate an energy dynamic balance vector including the power supply capacity for several seconds in the future. The power supply margin prediction model improves the prediction robustness through adversarial sample enhancement technology.

7. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 4 when run.

8. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 4.

Images