A neuromorphic control method and apparatus for a power electronic converter

Abstract

The application relates to the technical field of power electronic control, and provides a neuromorphic control method and device for a power electronic converter, which comprises the following steps: acquiring local electrical quantities of a node where the power electronic converter is located; inputting the local electrical quantities into a pulse neural network to deduce remote state information of a power grid; and generating a PWM signal for controlling the power electronic converter based on the output of the pulse neural network. The application deduces remote state information from local electrical quantities by using a pulse neural network, eliminates the dependence on a traditional communication layer, effectively avoids communication delay and network attack risks, and significantly improves the response speed, safety and self-adaptive capability of the power electronic converter.

Description

Technical Field

[0001] This invention relates to the field of power electronic control technology, specifically to a neuromorphic control method and device for a power electronic converter. Background Technology

[0002] With the rapid development of power electronics technology and smart grids, the penetration rate of distributed energy sources in power systems is constantly increasing, posing new challenges to the operation and control of power systems. In traditional power electronic converter control architectures, coordinated control between distributed power sources typically relies on communication networks. This communication-layer-based control method requires the transmission of state information such as voltage, current, and power through communication links to achieve global optimized scheduling and stable control.

[0003] However, existing control schemes based on the communication layer have significant limitations. First, the communication process inevitably introduces transmission delays. When disturbances or faults occur in the power grid, these delays can lead to lag in the response of the control system, thus affecting system stability. Second, communication networks may experience packet loss and bit errors during data transmission, resulting in missing or erroneous control information and reducing control reliability. More seriously, reliance on external communication networks makes the power system vulnerable to cyberattacks. Attackers could potentially tamper with control commands by intruding into communication links, severely threatening the safe operation of the power grid. Therefore, how to achieve fast, safe, and adaptive control of power electronic converters while ensuring control performance, without relying on traditional communication layers, has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the issues of latency, low reliability, and insufficient security caused by the reliance on communication networks in existing power electronic converter control technologies, this application proposes a neuromorphic control method and device for power electronic converters. This method utilizes a spiking neural network to infer remote state information from local electrical quantities, thereby achieving fast adaptive control without a communication layer.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A neuromorphic control method for a power electronic converter includes: acquiring local electrical quantities of the node where the power electronic converter is located; inputting the local electrical quantities into a pulse neural network to infer remote state information of the power grid; and generating a PWM signal for controlling the power electronic converter based on the output of the pulse neural network.

[0007] The above scheme utilizes the powerful nonlinear fitting capability and event-driven characteristics of spiking neural networks to infer the remote state information of the power grid based solely on local measurement data, thereby generating control signals. This completely eliminates the dependence on the traditional communication layer, effectively avoids the risks caused by communication delays, packet loss, and network attacks, and significantly improves the response speed and system security of power electronic converters.

[0008] As one implementation, the spiking neural network employs a leakage integral excitation neuron model.

[0009] The above scheme, by employing a leakage integral-triggered neuron model, leverages its biological interpretability and event-driven characteristics to process dynamic signals from the power grid more efficiently and reduce computational energy consumption.

[0010] In one implementation, the spiking neural network is configured to update network weights based on unsupervised Hebbian learning rules according to the power grid operating status.

[0011] The above scheme introduces unsupervised Heb learning rules, enabling the control system to automatically adjust network parameters according to the real-time operating status of the power grid, thereby achieving online adaptation to power grid parameter mismatch and topology changes and enhancing the robustness of the system.

[0012] As one implementation, the step of generating a PWM signal for controlling the power electronic converter based on the output of the spiking neural network includes: acquiring the membrane potential of the spiking neural network; and converting the membrane potential into a PWM pulse sequence through a dynamic threshold decoding mechanism.

[0013] The above scheme establishes a direct mapping relationship between the neural network output and the PWM modulation signal through a dynamic threshold decoding mechanism, which simplifies the control logic and realizes the deep integration of neuromorphic computing and power electronic modulation.

[0014] In one implementation, the dynamic threshold changes periodically over time, and its variation pattern matches the PWM carrier signal.

[0015] The above scheme ensures the accuracy of control signal generation by setting a dynamic threshold that matches the carrier signal, so that the output pulse of the pulse neural network can accurately correspond to the PWM modulation waveform.

[0016] As one implementation, the step of obtaining the local electrical quantities of the node where the power electronic converter is located adopts an event-driven sampling mechanism, which triggers sampling only when a dynamic change event in the power grid is detected.

[0017] The above scheme uses an event-driven sampling mechanism to collect and process data only when the power grid undergoes dynamic changes, which greatly reduces the collection of invalid data and redundant calculations, and further reduces the power consumption and computational burden of the control system.

[0018] As one implementation, the local electrical quantity is subjected to noise suppression processing before being input into the pulse neural network.

[0019] The above scheme effectively filters out measurement noise interference by suppressing the input signal, thereby improving the inference accuracy and stability of the system in low signal-to-noise ratio environments.

[0020] As one implementation, the method further includes an energy efficiency optimization step: dynamically adjusting the calculation frequency of the spiking neural network according to the grid load level.

[0021] The above solution achieves on-demand allocation of computing resources by dynamically adjusting the computing frequency according to the load level, further optimizing the overall energy efficiency of the system.

[0022] Furthermore, the present invention also provides a neuromorphic control device for a power electronic converter, comprising: a data acquisition module for acquiring local electrical quantities of the node where the power electronic converter is located; a pulse neural network processing module for inferring remote state information of the power grid based on the local electrical quantities; and a PWM generation module for generating a PWM signal for controlling the power electronic converter based on the output of the pulse neural network processing module.

[0023] The aforementioned device implements the control method through a modular design. The modules work together to complete the local inference and control functions without a communication layer.

[0024] In one embodiment, the spiking neural network processing module includes a spiking neural network for outputting membrane potential; the PWM generation module includes a dynamic threshold decoding unit configured to perform the steps of the method described above.

[0025] Beneficial effects:

[0026] The present invention provides a neuromorphic control method and apparatus for a power electronic converter, which has the following beneficial effects:

[0027] 1. Improved system security and reliability: This invention infers remote state information from local electrical quantities through a spiking neural network, completely eliminating the dependence on the traditional communication layer, cutting off the path of external network attacks, and avoiding control failures caused by communication delays and packet loss, thus significantly improving the security and operational reliability of the power electronic system.

[0028] 2. Improved response speed and control efficiency: By adopting event-driven sampling and spiking neural network processing mechanisms, calculations are triggered only when the power grid undergoes dynamic changes, which significantly reduces data redundancy and computational latency, and achieves millisecond-level rapid response to power grid disturbances.

[0029] 3. Enhanced system adaptability: By introducing unsupervised Heb learning rules, the system can update network weights online, automatically adapt to changes in power grid topology and parameter mismatch, and optimize and adjust control strategies without manual intervention.

[0030] 4. Innovative integration of control mechanisms: The dynamic threshold decoding mechanism proposed in this invention directly maps the pulse output of neuromorphic computing to PWM control signals, realizing the unification of artificial intelligence algorithms and power electronic physical modulation processes, and providing a new technical path for edge computing of smart grids. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the power supply and remote information mode provided in an embodiment of the present invention;

[0032] Figure 2 This is a diagram of the neural network leakage integral and excitation model provided in the embodiments of the present invention;

[0033] Figure 3 This is a simulation test diagram provided in an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the coordinated control results provided in an embodiment of the present invention;

[0035] Figure 5 This is a fault ride-through compliance voltage and current waveform diagram provided in an embodiment of the present invention;

[0036] Figure 6 This is a parameter adaptive voltage waveform diagram provided in an embodiment of the present invention;

[0037] Figure 7 This is a schematic diagram of spikes in the hidden layer of a neural network provided in an embodiment of the present invention;

[0038] Figure 8 This is a schematic diagram of the experimental apparatus provided in an embodiment of the present invention;

[0039] Figure 9 This is a waveform diagram of the load change and voltage sag experimental results provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0042] Example 1:

[0043] This embodiment provides a neuromorphic control method for power electronic converters, applied to smart grid edge computing scenarios. For example... Figure 1 As shown, the core of this method lies in utilizing the inference capabilities of spiking neural networks (SNNs) to achieve remote sensing and control of the power grid status locally, thereby eliminating the reliance on communication networks in traditional control architectures.

[0044] Specifically, the method includes the following steps:

[0045] Step S100: Obtain the local electrical quantities of the node where the power electronic converter is located.

[0046] In this embodiment, local electrical quantities refer to physical quantities that can be directly measured at the grid connection point of the power electronic converter, including but not limited to node voltage, branch current, frequency, or combinations thereof. Figure 1 As shown in (b), voltage and current signals in analog form are acquired in real time using voltage and current sensors deployed locally on the converter. Unlike existing technologies, this step relies solely on local measurement data and does not require obtaining status information from remote nodes via communication links, thus eliminating the risk of control failure due to communication delays, packet loss, or network attacks at the data source.

[0047] Step S200: Input the local electrical quantity into the pulse neural network to infer the remote state information of the power grid.

[0048] This is the core step of this embodiment. In traditional solutions, to obtain the voltage or power distribution of other nodes in the power grid, data must be transmitted through a wide-area communication network. This embodiment utilizes the coupling characteristics of the power grid's physical topology, where the dynamic changes in local electrical quantities implicitly contain information about the power flow distribution across the entire grid. This implicit correlation is mined using a trained spiking neural network. Specifically, the local voltage and current signals collected in step S100 are preprocessed (e.g., normalized, encoded) and then input into the spiking neural network. This network simulates the information processing mechanism of a biological nervous system, performing nonlinear mapping and feature extraction on the input signal through the integration of membrane potentials and pulse firing activities of internal neurons. Finally, the remote state information of key nodes in the power grid (such as remote bus voltage, line power flow, etc.) is reconstructed at the output layer. It should be understood that this "inference" is not a simple numerical calculation, but a data-driven intelligent sensing process. Its response speed is limited only by the inference calculation time of the neural network, typically in the microsecond range, much faster than communication transmission.

[0049] Step S300: Based on the output of the pulse neural network, a PWM signal for controlling the power electronic converter is generated.

[0050] After obtaining the inference results from the remote state information, this step converts them into specific control commands. The output signal of the spiking neural network (such as a pulse sequence or membrane potential) is sent to the PWM generation unit. This unit maps the neural network output to a pulse width modulation (PWM) signal with a corresponding duty cycle, based on a preset control strategy (such as voltage regulation or power distribution). This PWM signal directly drives the switching transistors (such as IGBTs or MOSFETs) in the power electronic converter, thereby achieving rapid regulation of the injected grid current or port voltage. This forms a closed loop of "local sensing - intelligent inference - local control," completely eliminating the "communication transmission" link found in traditional architectures.

[0051] This embodiment also provides a neuromorphic control device for a power electronic converter, used to perform the above-described method. For example... Figure 8 As shown, the device includes a data acquisition module, a pulse neural network processing module, and a PWM generation module.

[0052] The data acquisition module is used to acquire the local electrical quantities of the node where the power electronic converter is located. Specifically, this module can be composed of voltage transformers (PT), current transformers (CT), and corresponding signal conditioning circuits (such as filters and analog-to-digital converters ADC), which are responsible for converting high voltage and high current into low voltage digital signals that can be processed by the controller.

[0053] The spiking neural network processing module is used to infer remote state information of the power grid based on the local electrical quantities. This module is the "brain" of the device, and its hardware carrier can be a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or an embedded microcontroller. Deploying a spiking neural network model on this hardware leverages the advantages of hardware parallel computing to efficiently perform neuron integration, activation, and synaptic transmission operations, enabling real-time inference of the remote state.

[0054] The PWM generation module generates a PWM signal for controlling the power electronic converter based on the output of the pulse neural network processing module. This module typically includes a PWM controller and drive circuitry, converting the digital control quantity output by the neural network into a high-frequency switching signal to drive the converter's main circuitry.

[0055] Through the above scheme, this embodiment constructs a decentralized, communication-independent power electronic control architecture. By eliminating the communication link, the system not only avoids the limitation of control bandwidth caused by communication latency, but also completely eliminates network attack paths targeting the communication link, significantly improving the response speed and operational security of smart grid edge devices.

[0056] Example 2:

[0057] Based on the above embodiment 1, this embodiment provides a detailed description of the internal structure of the spiking neural network and its adaptive learning mechanism.

[0058] Specifically, the spiking neural network employs a leaky integral excitation (LIF) neuron model. For example... Figure 2 As shown, the LIF neuron model simulates the dynamic behavior of membrane potential in biological neurons. Its core mechanism lies in "integration" and "firing": the neuron integrates the input pulse signal over time, causing the membrane potential to gradually increase; when the membrane potential reaches a preset threshold voltage, the neuron immediately fires an output pulse, and the membrane potential is then reset to the resting potential, entering a brief refractory period. Compared to the traditional integral firing (IF) model, the LIF model introduces a "leakage" term, meaning the membrane potential decays exponentially when there is no input. This characteristic makes it particularly suitable for power grid signal processing: transient disturbances in the power grid are often short-lived and sudden, while steady-state signals or high-frequency noise are naturally filtered out by the "leakage" mechanism. Therefore, LIF neurons can effectively capture key dynamic events in the power grid (such as voltage sags and frequency spikes) while ignoring steady-state background noise, achieving "event-driven" sparse computation. This means the network is only activated when a valid event is detected, significantly reducing the computational load and energy consumption of the controller.

[0059] Furthermore, the spiking neural network is configured to update network weights based on unsupervised Hebbian learning rules according to the grid operating status. Traditional neural network training typically relies on offline supervised learning with large amounts of labeled data. However, at the edge of a smart grid, obtaining labeled data covering all fault types and topology changes is extremely difficult and costly. This embodiment employs unsupervised Hebbian learning rules, the core idea of ​​which is that "neurons that are excited together are connected together." In a grid scenario, this means that if a certain change pattern of local electrical quantities (such as voltage phase angle jumps) frequently occurs simultaneously with certain remote state information (such as power oscillations at neighboring nodes), the corresponding synaptic weights between neurons will be enhanced.

[0060] Specifically, the network adjusts weights in real time based on the power grid's operating status (such as frequency deviation and voltage fluctuation amplitude). For example, when a topology change occurs in the power grid (such as a line disconnection) leading to parameter mismatch, the network can automatically adjust the connection strength between neurons through an online learning mechanism, thereby re-establishing an accurate mapping between local measurements and remote states without manual intervention. Figure 6 As shown, the voltage waveforms with and without a parameter adaptive mechanism are compared. It can be seen that without adaptive learning ( Figure 6 In (a), parameter mismatch leads to continuous voltage oscillations or even divergence; however, after introducing the Hebbian learning rule in this embodiment ( Figure 6 In (b), the network weights quickly converged to the new optimal value, and the voltage waveform stabilized within a short time. This effect demonstrates that the online learning mechanism endows the control system with strong robustness and environmental adaptability, enabling it to cope with uncertainties in power grid operation.

[0061] Example 3:

[0062] Based on the above embodiments, this embodiment elaborates in detail how to convert the output of a pulse neural network into a PWM signal that can directly drive a power electronic converter.

[0063] Specifically, the step of generating a PWM signal for controlling the power electronic converter based on the output of the pulse neural network in step S300 includes the following sub-steps:

[0064] Step S301: Obtain the membrane potential of the spiking neural network. In the LIF neuron model, the membrane potential is a cumulative representation of the neuron's internal state. When a neuron receives an input pulse, the membrane potential undergoes an integral change. In this embodiment, the membrane potential of the output layer neurons is no longer merely an intermediate calculation variable, but is directly used as a modulation signal source. Unlike traditional control methods that require the neural network output to undergo digital-to-analog conversion before being input into an independent PWM generator, this embodiment treats the membrane potential itself as the waveform to be modulated, thereby achieving the fusion of calculation and modulation.

[0065] Step S302 involves converting the membrane potential into a PWM pulse sequence using a dynamic threshold decoding mechanism. This step is one of the core innovations of this invention. Traditional spiking neural network decoding typically employs "counting decoding" or "time decoding," which is difficult to directly adapt to the PWM modulation requirements of the power electronics field. This embodiment introduces a specific dynamic threshold mechanism, allowing the neuron's activation process to directly correspond to the PWM pulse generation process.

[0066] Furthermore, the dynamic threshold changes periodically with time, and its variation pattern matches that of the PWM carrier signal. Specifically, the dynamic threshold is described by the following formula:

[0067]

[0068] in, The pulse firing threshold at the current moment, The system sampling period or PWM carrier period. For the current absolute time variable, This is the current time slice number (a positive integer).

[0069] To understand this mechanism more clearly, it is necessary to combine it with... Figure 1 (c) The modulation principle on the right is explained. In traditional sinusoidal pulse width modulation (SPWM), the modulating wave (sine wave) is compared with the carrier wave (usually a triangular wave or sawtooth wave), and a high level is output when the modulating wave is greater than the carrier wave. The dynamic threshold formula in this embodiment actually constructs a sawtooth wave shape that changes linearly with time, and its physical meaning is equivalent to the PWM carrier signal. The formula... Indicates the first The center time offset of each time slice is calculated using a linear mapping to normalize the relative positions within the time window to a threshold space.

[0070] When the membrane potential of the spiking neural network (As a modulated signal) exceeding the aforementioned dynamic threshold When the signal is used as a carrier signal, the neuron fires a pulse and outputs a high level; conversely, it does not fire a pulse and outputs a low level. This process simulates the "excitation" characteristic of biological neurons and simultaneously completes the "comparison" action in PWM modulation at the physical level.

[0071] This embodiment achieves the unification of the neural network computation process and the power electronic modulation process through the above design. The integration process of the membrane potential corresponds to the formation of the modulated wave, the periodic change of the dynamic threshold corresponds to the generation of the carrier wave, and the pulse emission time precisely corresponds to the turn-on or turn-off time of the PWM switch. This "computation as modulation" mechanism eliminates the time lag between algorithm calculation and PWM generation in traditional digital control, greatly improving the real-time performance and response bandwidth of the control system. At the same time, since the dynamic threshold is directly driven by the time variable, no additional carrier generator hardware is required, simplifying the controller architecture and reducing system complexity.

[0072] Example 4:

[0073] Based on the above embodiments, this embodiment further optimizes the sampling mechanism, signal preprocessing and energy efficiency management of the control method to improve the robustness and operating efficiency of the system in complex power grid environments.

[0074] Specifically, the step of acquiring the local electrical quantities of the node where the power electronic converter is located adopts an event-driven sampling mechanism, triggering sampling only when a dynamic change event in the power grid is detected. Traditional control strategies typically employ periodic sampling at a fixed frequency (e.g., sampling thousands of times per second). Regardless of whether the power grid state changes, the controller continuously performs high-intensity data acquisition and computation, resulting in significant waste of computing resources and energy consumption. This embodiment introduces an event-driven mechanism, changing the sampling mode from "time-driven" to "event-driven." Figure 1 (c) As shown on the left, the system presets dynamic change event thresholds and monitors the rate or magnitude of change of local electrical quantities (such as voltage and current) in real time. Sampling is triggered only when the change in the electrical quantity exceeds the preset threshold, indicating a "dynamic change event" (such as a fault occurrence or load change). Figure 7 As shown, the spike activity of neurons in the hidden layer is illustrated. It can be seen that the spikes are mainly concentrated at moments of signal abrupt change, while neurons remain silent during steady-state periods. This mechanism significantly reduces the acquisition and transmission of redundant data, allowing the spiking neural network to be activated only when necessary. This significantly reduces the computational load and power consumption of the controller, while also enabling a faster response to power grid disturbances.

[0075] Furthermore, before inputting the local electrical quantity into the pulse neural network, noise suppression processing is performed on the local electrical quantity. In actual industrial environments, voltage and current signals are often superimposed with high-frequency switching noise and harmonic interference. If the noisy signal is directly input into the event-driven sampling module, noise spikes are easily misjudged as dynamically changing events, leading to false triggering and invalid calculations. This embodiment sets up a preprocessing stage before sampling, using a low-pass filter or moving average filtering algorithm to smooth the original electrical quantity. Specifically, the filtering algorithm adaptively adjusts the filtering strength according to the signal characteristics, filtering out high-frequency noise while retaining the dynamic change characteristics of the signal. This step ensures the accuracy of subsequent event detection, avoids false triggering due to noise, and improves the system's operational stability in low signal-to-noise ratio environments.

[0076] Furthermore, the method includes an energy efficiency optimization step: dynamically adjusting the computation frequency of the spiking neural network based on the grid load level. The load level of a smart grid typically exhibits significant peak-valley characteristics. During peak load periods, the system requires high-frequency computation to ensure control accuracy; however, during off-peak periods, maintaining high-frequency computation leads to unnecessary energy consumption. This embodiment establishes a mapping relationship between load level and computation frequency. When a low grid load level is detected, the clock frequency of the spiking neural network is automatically reduced or the number of activated neurons is decreased, thereby reducing chip power consumption. When the load level increases, the computation frequency is correspondingly increased to ensure control performance. This strategy of allocating computing resources on demand enables the control system to achieve optimal energy efficiency while ensuring dynamic response performance, making it particularly suitable for power-sensitive edge computing devices.

[0077] Example 5:

[0078] To verify the actual performance of the neuromorphic control method and device for the power electronic converter provided by this invention, this embodiment built an improved IEEE 14-bus system simulation model based on the MATLAB / Simulink platform and conducted tests using a hardware experimental bench. Figure 3 As shown, the test system comprises four distributed inverters (Inv1-Inv4), connected to different nodes in the system (e.g., nodes #2, #3, #6, and #8), to simulate a complex power grid environment with multiple inverters operating in parallel. The tests focused on inference accuracy, fault ride-through capability, and system response speed.

[0079] First, the accuracy of the inferences made by this invention in multi-machine coordinated control is verified. For example... Figure 4 As shown, this demonstrates the inference results and coordinated control effects of the spiking neural network processing modules deployed on each node in the IEEE 14-node system on remote state information. Figure 4(a) shows the regression accuracy of the modulated signal output by the neural network and the target signal. It can be seen that the two are highly consistent, which proves that the spiking neural network can accurately reconstruct the modulation information required for global control based solely on local electrical quantities. Figure 4 (c) and Figure 4 (d) The reactive power sharing results and average voltage regulation results of each inverter are shown respectively. Without relying on the communication layer, each inverter can automatically adjust its output reactive power based on the inferred remote information, achieving accurate power sharing and stable voltage regulation, thus verifying the effectiveness of the method of the present invention in communication-free coordinated control.

[0080] Secondly, verify the system's ability to overcome fault conditions and its rapid response characteristics. For example... Figure 5 As shown, the system is configured to experience a voltage sag fault during operation. Figure 5 (a) and Figure 5 (b) Demonstrates the voltage instability and overcurrent problems that occur in the system without the control strategy of this invention, with severe waveform oscillations and even divergence. However, after applying the control method of this invention, as shown... Figure 5 (c) and Figure 5 As shown in (d), when a fault occurs, the system can quickly detect voltage dips and current surges, and rapidly generate corrected PWM signals through online inference and dynamic threshold decoding mechanisms using a spiking neural network. The waveforms show that the voltage and current stabilize very quickly after the fault occurs, without sustained oscillations or overcurrent tripping, demonstrating the excellent fault ride-through (FRT) compliance of this invention. This effect is attributed to the event-driven sampling mechanism's keen capture of dynamic events and the millisecond-level inference speed of the spiking neural network.

[0081] Furthermore, combined Figure 9 The hardware experimental results are used to analyze the load mutation scenario. The experimental setup is as follows: Figure 8 As shown, it includes two three-phase DC-AC converters (SNN 1 and SNN 2), which are connected to resistive loads respectively. Figure 9 (a) and Figure 9 (e) shows the output voltage and current waveforms of the converter at the moment of load change. It can be seen that at the moment the load is connected or disconnected, the voltage waveform has only a small fluctuation and then quickly returns to steady state, and the current waveform also transitions smoothly. Figure 9 (d) and Figure 9 (h) Recorded the peak activity of hidden layer neurons during load changes. It can be clearly observed that the peaks are mainly concentrated during the dynamic process of load changes, while neuronal activity is sparse in steady state. This result intuitively demonstrates the effectiveness of the event-driven sampling mechanism: the system triggers high-frequency sampling and computation only when dynamic changes in the power grid are detected, while maintaining a low-power silent state in steady state, thus achieving optimized allocation of computing resources and energy efficiency.

[0082] Finally, combining Figure 7 The sparse computational characteristics of neural networks are explained. Figure 7 (a) shows the temporal relationship between the input signal va and the hidden layer spike. Figure 7 (b) illustrates the generation of spikes within the moving window. It can be seen that the generation of spikes is highly synchronized with the dynamic changes of the signal, and there are almost no spikes during the signal's stable phase. This sparsity not only reduces the computational load of the hardware implementation but also makes the control system more robust to high-frequency noise, as steady-state noise typically cannot trigger the integral of the neuron's membrane potential to reach the threshold. In summary, through simulation and experimental verification, the control method provided by this invention successfully achieves accurate inference of remote state information without the support of a communication layer, and exhibits comprehensive advantages of fast response, high stability, and low power consumption under extreme conditions such as fault ride-through and load surges.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, such as using other types of spiking neuron models to replace the LIF model, or using other forms of dynamic threshold functions to replace the sawtooth wave threshold, should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A neuromorphic control method for a power electronic converter, characterized in that, The method includes: Obtain the local electrical quantities of the node where the power electronic converter is located; The local electrical quantities are input into a pulse neural network to infer the remote state information of the power grid; Based on the output of the pulse neural network, a PWM signal for controlling the power electronic converter is generated.

2. The method according to claim 1, characterized in that, The spiking neural network uses a leakage integral-triggered neuron model.

3. The method according to claim 2, characterized in that, The spiking neural network is configured to update network weights based on unsupervised Hebbian learning rules according to the power grid operating status.

4. The method according to claim 1, characterized in that, The step of generating a PWM signal for controlling the power electronic converter based on the output of the pulse neural network includes: Obtain the membrane potential of the spiking neural network; The membrane potential is converted into a PWM pulse sequence through a dynamic threshold decoding mechanism.

5. The method according to claim 4, characterized in that, The dynamic threshold changes periodically over time, and its variation pattern matches that of the PWM carrier signal.

6. The method according to claim 1, characterized in that, The step of obtaining the local electrical quantities of the node where the power electronic converter is located adopts an event-driven sampling mechanism, which triggers sampling only when a dynamic change event in the power grid is detected.

7. The method according to claim 1, characterized in that, Before inputting the local electrical quantity into the pulse neural network, the local electrical quantity is subjected to noise suppression processing.

8. The method according to claim 1, characterized in that, The method further includes an energy efficiency optimization step: dynamically adjusting the calculation frequency of the spiking neural network according to the grid load level.

9. A neuromorphic control device for a power electronic converter, characterized in that, include: The data acquisition module is used to acquire the local electrical quantities of the node where the power electronic converter is located; A pulse neural network processing module is used to infer remote state information of the power grid based on the local electrical quantities; The PWM generation module is used to generate a PWM signal for controlling the power electronic converter based on the output of the pulse neural network processing module.

10. The apparatus according to claim 9, characterized in that, The spiking neural network processing module includes a spiking neural network, which is used to output membrane potential; The PWM generation module includes a dynamic threshold decoding unit, which is configured to perform the steps of the method as described in claim 4 or 5.

Images