Virtual synchronous machine dynamic inertia control method and system based on artificial intelligence technology

Abstract

The invention provides a virtual synchronous machine dynamic inertia control method and system based on an artificial intelligence technology, which are applied to voltage source inverter control, and mainly comprises the following steps: data acquisition and preprocessing are carried out, and dynamic parameters are output through a time sequence neural network and are respectively used for control links of a virtual synchronous machine; a convolutional neural network is adopted to calculate a phase compensation angle for phase correction of pulse width modulation, and then a deep reinforcement learning network is utilized to output duty ratio correction to adjust the pulse width modulation so as to cooperatively control the dynamic inertia of the virtual synchronous machine. According to the invention, dynamically adjusted parameters are obtained through data time sequence processing and neural network application, and adaptive adjustment of core parameters can be realized; the convolutional neural network participates in modulation signal calculation, hardware phase delay is effectively compensated, and the dynamic response precision is improved; and cooperative optimization of inertia control and dynamic response is realized by combining deep reinforcement learning, current overshoot is reduced, harmonic waves are suppressed, and the electric energy quality and the system stability are improved.

Description

Technical Field

[0001] The present invention relates to the field of power supply technology, and particularly relates to a method and system for dynamic inertia control of a virtual synchronous machine based on artificial intelligence technology. Background Art

[0002] With the large-scale access of distributed power sources to the power grid, the virtual synchronous machine (VSG) technology has attracted much attention because it can effectively improve the grid connection stability and power quality of distributed power sources and simulate the inertia and damping characteristics of synchronous generators.

[0003] Among them, the two-level voltage source inverter (2L-VSI) has become a common topology for VSG implementation due to its simple structure and low cost. However, the traditional 2L-VSI type VSG faces many technical problems in actual operation, specifically as follows: Hardware characteristic constraints: The LC filter configured in it has a phase lag, and the switching frequency is limited by hardware conditions, resulting in a delay in the dynamic response of the VSG to grid frequency and voltage fluctuations, and it is difficult to quickly track grid changes; Insufficient parameter adaptability: Using fixed virtual inertia (J) and damping coefficient (D), in the face of various scenarios such as photovoltaic output fluctuations, load mutations, and weak grids, it cannot be dynamically adjusted to meet actual needs, reducing the flexibility and stability of system operation; Limitations of control algorithms: Based on the control strategy of a linear PI regulator, in the face of non-linear disturbances such as harmonic resonance and load mutations, it cannot effectively coordinate the relationship between inertia support and fast response, often resulting in the compromise failure of both, affecting the overall performance of the system.

[0004] In response to the above problems, existing solutions include parameter adaptive adjustment technology, predictive control schemes, and hardware compensation schemes. Although the response speed, stability, etc. of the system have been improved to a certain extent, there are still obvious defects, specifically as follows: Parameter adaptive adjustment scheme: Mostly rely on a rule base or fuzzy logic to adjust J and D, but these methods highly rely on manual experience to build rules, and when facing complex and changeable actual working conditions, the generalization ability is insufficient, and it is difficult to achieve precise self-adaptation.

[0005] Predictive control scheme: Although model predictive control (MPC) can theoretically optimize control, it is necessary to establish an accurate mathematical model, and the 2L-VSI has non-ideal characteristics such as dead zone effect and device aging, resulting in model mismatch and greatly reducing the control effect.

[0006] Hardware compensation scheme: Improve the performance by increasing the filter or raising the switching frequency, but this will significantly increase the system cost and power loss, violating the original intention of the 2L-VSI to pursue economy, and is limited in practical applications.

[0007] In summary, the traditional 2L-VSI type VSG still requires an improved control method and corresponding system to improve the overall dynamic performance and have better multi-scenario adaptability and economy. Summary of the Invention

[0008] In view of the problems existing in the prior art, the present invention provides a system with a simple structure and low cost, and a method for collaborative control of the dynamic inertia of a virtual synchronous machine.

[0009] To achieve the above object, the technical solution adopted by the present invention is as follows: On the one hand, the present invention provides a method for controlling the dynamic inertia of a virtual synchronous machine based on artificial intelligence technology, mainly including the following steps: Collect multi-dimensional power grid information in real time; preprocess the multi-dimensional power grid information and generate a time series data sequence of a preset time window; Input the time series data sequence into a time series neural network and output dynamically adjusted virtual inertia, damping coefficient and virtual impedance; Substitute the virtual inertia and the damping coefficient into the control module of the virtual synchronous machine, and substitute the virtual impedance into the power control voltage and current control link of the virtual synchronous machine; Use a convolutional neural network to extract the phase-frequency characteristics of the filter of the voltage source inverter and generate a phase compensation angle; correct the phase of the pulse width modulation signal used to drive the voltage source inverter through the phase compensation angle; Train a deep reinforcement learning network based on an integrated power quality optimization reward function, use the system state parameters of the voltage source inverter and the grid operation state parameters as the input quantities of the deep reinforcement learning network, and output a duty cycle correction amount of the pulse width modulation signal used to drive the voltage source inverter; input the duty cycle correction amount into the pulse width modulation strategy of the voltage source inverter to achieve collaborative control of the dynamic inertia of the virtual synchronous machine.

[0010] Optionally, the time series neural network is synchronously updated through an online update strategy; The online update strategy includes the following steps: Periodically collect the multi-dimensional power grid information and the historical inertia parameter sequence and form a current operating state data set; Apply a sliding window forgetting factor to the current operating state data set and obtain a training data set; in the training data set, assign an update weight to the latest data and reduce the weight of the historical data in the data set according to the exponential decay method to generate weighted training data; Periodically perform incremental update of the weights of the time series neural network based on the weighted training data.

[0011] Optionally, the temporal neural network is a long short-term memory network and adopts a double-layer long short-term memory structure; In the double-layer long short-term memory structure, the bottom layer has 64 neurons and the top layer has 32 neurons.

[0012] Optionally, the temporal neural network is optimized by transfer learning after being deployed in the corresponding scenario; The transfer learning includes source domain pre-training based on multi-scenario simulation data and target domain fine-tuning using the measured data of the target scenario to update the top-layer network parameters.

[0013] Optionally, the multi-dimensional power grid information includes power grid frequency deviation, active power fluctuation, DC bus voltage change rate, and output current signal; The preprocessing at least includes noise filtering of the collected multi-dimensional power grid information through a filtering algorithm.

[0014] Optionally, the convolutional neural network is a one-dimensional convolutional neural network, adopts an asymmetric convolutional kernel structure, and the kernel weight distribution is obtained through phase difference supervised training.

[0015] Optionally, the deep reinforcement learning network is a deep Q network; The parameters of the power quality comprehensive optimization reward function are at least provided with a frequency deviation reward term, a current overshoot penalty term, and a harmonic distortion penalty term, and the index weights of each term are dynamically set.

[0016] On the other hand, the present invention also provides a virtual synchronous machine dynamic inertia control system based on artificial intelligence technology for implementing the above-mentioned virtual synchronous machine dynamic inertia control method based on artificial intelligence technology; The system includes a data acquisition layer and a control layer; The data acquisition layer is provided with multiple types of sensors for collecting the state information of the current power grid; the data acquisition layer is connected to the control layer through an analog-to-digital conversion unit; The control layer is provided with a heterogeneous multi-processor embedded computing platform for calculating control signals and transmitting the control signals to the voltage source inverter.

[0017] Optionally, the heterogeneous multi-processor embedded computing platform is a local cooperation architecture, including a digital signal processor and a field programmable gate array; The digital signal processor is used for control and algorithm execution; The field programmable gate array is used for hardware acceleration.

[0018] Optionally, it further includes a server; The server is connected to the control layer and is used for deploying and online updating the temporal neural network.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The system of the present invention has a simple structure and low control and construction costs. Dynamically adjusted parameters are obtained through data timing processing and neural network application, which can realize adaptive adjustment of core parameters and solve the problem that traditional fixed parameters are difficult to cope with complex scenarios. Convolutional neural networks are used to participate in modulation signal calculations to effectively compensate for hardware phase delays and improve dynamic response accuracy. Deep reinforcement learning is then combined to achieve coordinated optimization of inertia control and dynamic response, reduce current overshoot and suppress harmonics, improve power quality and system stability, and significantly enhance the adaptability of virtual synchronous machines to dynamic changes in power grids and their comprehensive control performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 is a system diagram in a specific embodiment of the present invention; Figure 2 It is a structural diagram of a time-series neural network in a specific embodiment of the present invention; Figure 3 is a convolutional neural network structure diagram in a specific embodiment of the present invention; Figure 4 It is a comparison diagram of simulation and experiment of the power quality comprehensive optimization reward function in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0024] In the description of the present invention, “plurality” means two or more than two, unless otherwise clearly and specifically defined.

[0025] It should be noted that the methods used in the present invention are all conventional methods unless otherwise specified; the raw materials and devices used are all conventional commercially available products unless otherwise specified, and their sources are not specifically limited.

[0026] On the one hand, this embodiment provides a virtual synchronous machine dynamic inertia control method based on artificial intelligence technology, which mainly includes the following steps: First, as Figure 1 shown in the system, the application scenario of this embodiment is a two-level voltage source inverter (2L-VSI) and a virtual synchronous generator (SVG) of the topology, which is used to achieve dynamic inertia collaborative control.

[0027] Multidimensional grid information is collected in real time through sensors. The information includes grid frequency deviation, active power fluctuation, DC bus voltage change rate, and output current signal. The collected multidimensional grid information is preprocessed through a filtering algorithm for noise filtering, etc., to generate a time series data sequence of a preset time window.

[0028] The time series data sequence is input into a time series neural network optimized by transfer learning. The time series neural network is a two-layer long short-term memory network (LSTM), with 64 neurons in the bottom layer and 32 neurons in the top layer. Its transfer learning process includes: based on multi-scenario simulation data of the IEEE 39-bus system, such as multi-scenario simulation data of photovoltaic fluctuations, load switching, short-circuit faults, etc., perform source domain pre-training to complete the initial LSTM model; in the target scenario, such as when deploying an island microgrid, freeze the weights of the bottom layer network, and only fine-tune the parameters of the top layer fully connected layer using the measured data of the target scenario. For example, using 1% of the measured data of the target scenario to complete rapid fine-tuning, the error convergence speed can be increased by 60%.

[0029] Thus, a time series neural network structure as Figure 2 shown is constructed, specifically: (1) Input layer (Input); Input features: time window data, length = 10ms; grid frequency deviation ; power fluctuation ; DC bus voltage change rate ; historical inertia parameter sequence ， ; dimension: [batch_size, timesteps = 10, features = 4].

[0030] (2) LSTM layer; The first LSTM layer (Layer1): number of neurons: 64; return sequence , that is, transfer the complete timing output to the next layer; activation function: tanh; transfer learning mark: freeze weights, lock after pre-training in the source domain.

[0031] The second LSTM layer (Layer2): number of neurons: 32; return sequence: , that is, only output the result of the last time step; activation function: tanh; transfer learning mark: fine-tuning, update weights when adapting to the target domain.

[0032] (3) Fully connected layer; Feature fusion layer: number of neurons: 32 → 16; activation function: ReLU; input: output of the LSTM layer + real-time grid impedance Z grid , input from the outside; Output layer: number of neurons: 3, corresponding to J 、 D 、 ; activation function: linear, no activation function, directly output parameter values.

[0033] (4) Output layer (Output layer); Output parameters: virtual inertia J , unit: kg•m²; damping coefficient D , unit: N•m•s / rad; virtual impedance , unit: Ω.

[0034] Through the above time series neural network, the dynamically adjusted virtual inertia J , damping coefficient D and virtual impedance can be output.

[0035] Among them, the input of the time series neural network is specifically the grid frequency deviation , power fluctuation , DC bus voltage change rate , and the historical inertia parameter sequence. The 10ms short sequence is suitable for the embedded platform to avoid excessive model complexity. The main considerations are real-time requirements: the system is implemented based on the embedded platform and requires low-latency processing, such as a 10ms window; grid frequency characteristics: 50Hz fundamental wave, and the dynamic fluctuations are usually in the sub-second level. A 1kHz sampling rate is sufficient to capture key harmonics and transient processes. Therefore, a 1kHz sampling rate, that is, 10 points, is a reasonable choice.

[0036] In addition, the spiking neural network maintains adaptability through an online update strategy: periodically collecting multi-dimensional power grid information and historical inertia parameter sequences to form a current operating state data set; applying a sliding window forgetting factor, such as selecting λ = 0.95 to process the data set, assigning an updated weight to the latest data, that is, a higher weight; while historical data reduces the weight according to the exponential decay method to generate weighted training data; incrementally updating the network weights based on the weighted training data every 5 minutes to adapt to long-term changes such as equipment aging and seasonal characteristic drift.

[0037] Substitute the virtual inertia J and damping coefficient D into the control module of the virtual synchronous machine to implement the core logic of active-frequency droop and reactive-voltage regulation; substitute the virtual impedance into the voltage-current control link of the virtual synchronous machine.

[0038] Under the parallel operation conditions of the above dynamic parameter operations, a convolutional neural network is used to extract the phase-frequency characteristics of the LC filter of the voltage source inverter to implement a non-linear delay compensation mechanism. The CNN adopts an asymmetric convolutional kernel structure (the kernel size of the previous layer is 5×1, and the kernel size of the latter layer is 3×1), and the kernel weight distribution is determined by supervised training of the phase difference of the Hilbert transform. After generating the phase compensation angle, the pulse width modulation (PWM) signal is phase-corrected to compensate for the phase lag of the filter. After that, the harmonic content (THD) of the filtered voltage is monitored in real time. When THD > 3%, it switches to the high-precision compensation mode, that is, using a 0.1° resolution.

[0039] Furthermore, in this embodiment, a one-dimensional convolutional neural network (1D-CNN) is used. The LC filter phase lag compensation network architecture formed by training the 1D-CNN is as Figure 3 shown. From left to right in the figure, it is divided into four parts: input layer, convolutional layer, pooling / full connection layer, and output layer. Specifically, the architecture is as follows: 1. Input layer; 1.1 Input data: Waveform: The difference waveform between the inverter output voltage and the filtered voltage , 256 points / cycle; Dimension: 1×256 single-channel time series.

[0040] 2. Convolutional layer; 2.1 The first convolutional layer (Conv1D-1): Number of convolutional kernels (kernel-1): 5; Kernel size: 5 in length; Stride: 2; Activation function: ReLU; Output dimension: 5×126 feature maps; 2.2 Second Convolutional Layer (Conv1D-2): Number of convolutional kernels (kernel-2): 10; Kernel size: 3 in length. Stride: 1; Activation function: ReLU; Output dimension: 10×124.

[0041] 3. Pooling and Fully Connected Layers; 3.1 Max Pooling Layer (MaxPool1D): Pooling size: 2; Stride: 2; Output dimension: 10×62; Flatten layer: Output dimension: 1×620 (10×62); Fully Connected Layers (Dense1,2): Number of neurons: 64→16→1; Activation function: Linear for the last layer, outputting the phase compensation angle θ comp 。

[0042] 4. Output Layer (Output); 4.1 Output: Phase compensation angle θ comp , in degrees, with a precision of 0.1°, which is used for leading compensation.

[0043] The data flow is labeled as follows: a. Input → Conv1D-1: Waveform local feature extraction (such as high-frequency harmonics, phase mutations).

[0044] b. Conv1D-1 → Conv1D-2: Deep feature abstraction (such as phase delay trend).

[0045] c. Conv1D-2 → MaxPool: Dimensionality reduction to retain significant features.

[0046] d. Flatten → Dense: Mapping to the phase compensation amount.

[0047] The weight distribution of the above one-dimensional convolutional neural network is supervised and trained through the Hilbert transform phase difference to calculate the actual phase difference in real time, serving as the true value label for CNN training to ensure that the compensation error ≤ 0.5.

[0048] Based on the Deep Q-Network (DQN) algorithm for multi-objective collaborative optimization, its network structure is a 6-dimensional input layer → hidden layer (64 neurons → 32 neurons, ReLU) → 5-dimensional Q-value output layer. The training strategy of this network is: Prioritized Experience Replay (PER) + Double Q-Network (DoubleDQN) to avoid overestimation of Q-values.

[0049] Furthermore, a corresponding reward function is set in the DQN training stage. Specifically, the reward function of the DQN set in this embodiment is: ; This reward function is for the training of the comprehensive power quality optimization reward function. The complete comprehensive power quality optimization reward function combines a frequency deviation reward term, an overshoot penalty term for current, a harmonic distortion penalty term, and a reward term for the health of the energy storage. The comprehensive power quality optimization reward function The formula is as follows: ; In the formula, is the frequency deviation reward term, α is the frequency stability reward weight coefficient, which determines the contribution of the frequency deviation to the reward. Its role is in a weak power grid, such as when the short-circuit capacity ratio SCR < 2 or the frequency fluctuates violently, increasing α can preferentially encourage the system to maintain frequency stability; β is the frequency deviation sensitivity coefficient, which controls the attenuation rate of the reward with the frequency deviation. When β > 0 and the value is larger, the negative impact of the frequency deviation on the reward is more significant, prompting the system to quickly suppress the frequency deviation; is the overshoot penalty term for current, is the overshoot of the output current in the dq coordinate system, that is, the deviation relative to the rated current, which characterizes the current fluctuation in the dynamic response. γ is the overshoot penalty weight coefficient for current, which adjusts the degree of attention to current stability. Its role is when high-load switching or power mutation occurs, increasing its value can suppress the current peak and avoid the risk of device overcurrent; is the harmonic distortion penalty term, THD is the Total Harmonic Distortion, which measures the harmonic content of the output voltage / current. It can be set to trigger a penalty when THD > 3%. δ is the harmonic distortion penalty weight coefficient, which controls the optimization priority of power quality. For example, in the grid-connected scenario, by increasing δ, it is ensured that THD meets the standard and harmonic pollution of the power grid is avoided; is the reward term for the health of the energy storage, is the health of the energy storage, which can be designed as a piecewise linear function or a Gaussian function to encourage the SOC to be maintained in a reasonable range, such as 30% - 90%, to avoid the impact of deep charge and discharge on the battery life, is the weight coefficient for the health of the energy storage, which adjusts the optimization priority of the energy storage life. Its typical value range is, for example, 0.1 - 0.3.

[0050] Among them, the weights of the various indicators of the power quality comprehensive optimization reward function are dynamically adjusted according to the grid strength and load characteristics. Further, the aforementioned dynamic adjustment is the implementation method of scenario adaptation. Weak grid: increase α, such as from 0.8 → 1.2, strengthen the priority of frequency stability, and suppress low-frequency oscillation; High load: increase γ, such as from 0.5 → 0.8, reduce the stress on the device caused by current overshoot; Harmonic-sensitive scenario: increase δ, such as from 0.3 → 0.6, ensure that the THD meets the standard, referring to THD < 2.5% during PV grid connection. The specific implementation method of automatic dynamic adjustment can refer to Pareto front optimization: obtain the Pareto Optimal solutions under different weight combinations through offline simulation, preset a multi-group (α, β, γ, δ) parameter library, and the DQN dynamically calls according to the real-time grid state (such as SCR value, load rate) to avoid manual parameter adjustment. Furthermore, conduct convergence verification on the power quality comprehensive optimization reward function of this embodiment, such as Figure 4 As shown, draw the function convergence curve under the simulation environment and experimental environment conditions respectively. The experimental steady-state value is 0.48, and the simulation steady-state value is 0.3.

[0051] Therefore, the system state parameters of the voltage source inverter, such as the DC bus voltage, output current, etc., and the grid operation state parameters, such as frequency deviation, power fluctuation, etc., are used as the input of the aforementioned trained deep learning network. By discretizing the action space, for example, dividing the continuous PWM duty cycle correction amount ΔD into a five-level action space of -5%, -2.5%, 0%, +2.5%, +5%, with a step resolution of 0.5%; thereby calculating and generating the PWM duty cycle correction amount, inputting it into the inverter PWM modulation strategy, and realizing the coordinated control of dynamic inertia and response speed.

[0052] On the other hand, this embodiment also provides a Figure 1 virtual synchronous machine dynamic inertia control system based on artificial intelligence technology as shown, which is used to implement the above control method, mainly including a data acquisition layer, a control layer, and a server.

[0053] Among them, the data acquisition layer: deploy multiple types of sensors, such as Hall voltage / current sensors: measure the DC bus voltage Vdc, the inverter output current iabc, the grid voltage vgrid, and a frequency detection module: collect the grid frequency fgrid in real time. Through the above sensors, real-time state information such as grid frequency, voltage, current, and DC bus voltage is collected. After the sensor signals are converted into digital signals by a high-precision analog-to-digital conversion unit (ADC), they are transmitted to the control layer.

[0054] Control layer: An heterogeneous multi-processor embedded computing platform is adopted, and the DSP+FPGA local cooperation architecture adopted in this embodiment. Among them, a digital signal processor (DSP), such as TMS320F28335 is used to execute the core algorithms: running a double-layer LSTM network to generate dynamic virtual inertia, damping coefficient and virtual impedance; optimizing the PWM duty cycle correction amount through the DQN algorithm; implementing the core control logic of the virtual synchronous machine. A field programmable gate array (FPGA), such as Xilinx Zynq is responsible for hardware acceleration: preprocessing the collected data, including normalization and sliding window processing; calculating the phase compensation angle in real time through a fixed 1D-CNN hardware accelerator; generating PWM signals and embedding dead time, and integrating overcurrent / overvoltage protection circuits to achieve nanosecond-level fault response. After receiving the data from the data acquisition layer, the FPGA performs preprocessing and caching, and realizes data transmission with the DSP through direct memory access (DMA).

[0055] The execution layer is the application object of this embodiment, and its main body is a two-level voltage source inverter (2L-VSI). The execution layer also includes a PWM module, an LC filter and a topology. Among them, the topology also includes three-phase bridge arms (IGBT) and DC bus capacitors. The execution layer interacts with the power grid / load, and collects the connection point voltage through sensors v PCC and current i grid .

[0056] Server: Connected to the control layer through a high-speed communication interface, used to deploy the initial LSTM model and support online update: storing multi-scenario simulation data sets for source domain pre-training; receiving the measured data of the target scenario uploaded by the control layer, performing top-level network parameter fine-tuning; periodically sending the updated model weights to the DSP to achieve cross-scenario adaptation and long-term maintenance of the time series neural network.

[0057] During system operation, the data acquisition layer provides real-time input for the control layer. The DSP and FPGA cooperate to complete AI algorithm inference and hardware control. The server supports model training and update, forming a closed loop of "data acquisition - algorithm optimization - hardware execution - model iteration", and realizing the dynamic inertia cooperative control of the virtual synchronous machine in a complex power grid environment.

[0058] Comparative example 1: VSG control of a photovoltaic energy storage system; Application scenario: In a photovoltaic energy storage system, a 2L-VSI is connected to a 380V low-voltage distribution network as a virtual synchronous machine, and it is required to cope with photovoltaic output fluctuations, load mutations and power grid frequency disturbances, and at the same time meet the grid connection harmonic standard of THD < 3%.

[0059] Among them, the system configuration and hardware parameters of the system applying the method of the present invention in this comparative example are as shown in Table 1 below: Table 1 In the same environment, a traditional PI control group is set up, and simulation field tests are carried out through two groups of events respectively. The events are: Event 1, sudden drop in photovoltaic power (cloud shading), the light intensity drops by 70% within 0.5 s, and the photovoltaic output drops from 50 kW to 15 kW; Event 2, grid frequency drop (simulating a fault), the grid frequency drops from 50 Hz to 48.5 Hz and lasts for 200 ms.

[0060] Thus, the experimental comparison results of traditional PI control and the control method of the present invention are obtained, as shown in Table 2 below: Table 2 It can be seen from the comparison that the method and system of the present invention have the following advantages: Multi-time scale coordination: LSTM is responsible for generating inertia parameters at the 10 ms level, DQN optimizes the dynamic response at the 1 ms level, and CNN compensates for hardware delay at the μs level. Economy-performance balance: Based on a low-cost 2L-VSI, it achieves dynamic performance close to that of a 3L-NPC, saving more than 30% in hardware costs. Strong robustness: Combining transfer learning and online fine-tuning to adapt to non-ideal factors such as photovoltaic output fluctuations and device aging.

[0061] Furthermore, the method of the present invention can be further extended to multiple scenarios such as wind power and energy storage. By adjusting the training data set and the weights of the reward function to adapt to different application requirements, it also shows that the method of the present invention has good multi-scenario versatility.

[0062] Comparative Example 2: VSG adaptive regulation under a weak grid; Application scenario: A 2L-VSI virtual synchronous machine is connected to a weak grid with an access impedance ratio SCR < 2, such as a remote island microgrid or a high-proportion new energy access area, which needs to cope with grid impedance mutation, voltage fluctuation and sub-synchronous oscillation risks, while maintaining grid connection stability and power quality.

[0063] Among them, the system configuration and hardware parameters of applying the method of the present invention in this comparative example are as shown in Table 3: Table 3 Another fixed group of traditional VSGs is set up in the same environment and simulation field tests are carried out through two groups of events respectively. The events are: Event 1, grid impedance mutation, such as line switching, the grid impedance steps from 2 mH to 8 mH to simulate line fault isolation; Event 2, background harmonics cause sub-synchronous oscillation, and the amplitude of the 5th harmonic voltage in the grid background rises to 8%, triggering a 90 Hz sub-synchronous oscillation.

[0064] Thus, the fixed value of the traditional VSG is obtained The experimental comparison results controlled by the method of the present invention are shown in Table 4 below: Table 4 It can be seen from the comparison that the method and system of the present invention have the following advantages: Impedance adaptive ability: Dynamically adjusted by LSTM , stable when Z grid changes by ±300%, without manual re-parameterization. Resonant active suppression: Integrating CNN harmonic lead compensation detection and notch filtering, the subsynchronous oscillation decay time < 100 ms. Cross-scenario compatibility: Adversarial training enables the model to adapt to unseen impedance topologies, such as looped microgrids and radial distribution networks.

[0065] Furthermore, the method of the present invention can be further extended to the scenario of multiple VSGs in parallel with a weak grid, and the and J of each unit are coordinated through the Consensus Algorithm to avoid circulating current and oscillation.

[0066] In summary, it can be seen from the comparison that the method of the present invention has a dynamic response speed increase of ≥30% and an inertia support effect error of ≤5% in terms of performance improvement; and under the same THD requirement, the volume of the LC filter is reduced by 20%. Further, in terms of economy, the performance of the multilevel topology is achieved based on the low-cost 2L-VSI hardware, and through the lightweight design of the model, it does not require an additional AI acceleration chip. Furthermore, in terms of generalization ability, transfer learning supports cross-scenarios, such as photovoltaic, wind power, and energy storage, and can be quickly adapted, reducing the training data requirement by 80%.

[0067] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than a limitation on the protection scope of the present invention. Any simple modification or equivalent replacement of the technical solution of the present invention by those of ordinary skill in the art does not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A virtual synchronous machine dynamic inertia control method based on artificial intelligence technology, applied to a voltage source inverter control system, characterized in that: The steps include: Collect multi-dimensional power grid information in real time; pre-process the multi-dimensional power grid information and generate a time series data sequence of a preset time window; Inputting the time series data sequence into a time series neural network, and outputting dynamically adjusted virtual inertia, damping coefficient and virtual impedance; Substituting the virtual inertia and the damping coefficient into a control module of a virtual synchronous machine, and substituting the virtual impedance into a voltage and current control link of the virtual synchronous machine; A convolutional neural network is used to extract the phase frequency characteristics of the filter of the voltage source inverter and generate a phase compensation angle; a pulse width modulation signal for driving the voltage source inverter is phase corrected by the phase compensation angle; A deep reinforcement learning network is trained based on a comprehensive power quality optimization reward function, the system state parameters of the voltage source inverter and the grid operation state parameters are used as input quantities of the deep reinforcement learning network, and a duty cycle correction quantity of a pulse width modulation signal for driving the voltage source inverter is output; the duty cycle correction quantity is input into a pulse width modulation strategy of the voltage source inverter to coordinately control the dynamic inertia of the virtual synchronous machine.

2. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 is characterized in that: The temporal neural network is synchronously updated through an online update strategy; The online update strategy includes the following steps: Periodically collecting the multi-dimensional power grid information and historical inertia parameter sequences, and forming a current operating status data set; Applying a sliding window forgetting factor to the current running state data set to obtain a training data set; In the training data set, the latest data is given an update weight, and the weight of the historical data of the data set is reduced according to the exponential decay method to generate weighted training data; The weights of the temporal neural network are incrementally updated periodically based on the weighted training data.

3. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 is characterized in that: The temporal neural network is a long short-term memory network and adopts a double-layer long short-term memory structure; In the double-layer long short-term memory structure, the bottom layer has 64 neurons and the top layer has 32 neurons.

4. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 or 3, characterized in that: The temporal neural network is optimized through transfer learning after being deployed in the corresponding scenario; The transfer learning includes source domain pre-training based on multi-scenario simulation data and target domain fine-tuning using target scenario measured data to update top-level network parameters.

5. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 is characterized in that: The multi-dimensional grid information includes grid frequency deviation, active power fluctuation, DC bus voltage change rate and output current signal; The preprocessing at least includes performing noise filtering on the collected multi-dimensional power grid information through a filtering algorithm.

6. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 is characterized in that: The convolutional neural network is a one-dimensional convolutional neural network that adopts an asymmetric convolution kernel structure, and the kernel weight distribution is obtained through phase difference supervised training.

7. The method for controlling dynamic inertia of a virtual synchronous machine based on artificial intelligence technology according to claim 1 is characterized in that: The deep reinforcement learning network is a deep Q network; The parameters of the power quality comprehensive optimization reward function include at least a frequency deviation reward item, a current overshoot penalty item and a harmonic distortion penalty item, and the indicator weights of each item are dynamically set.

8. A virtual synchronous machine dynamic inertia control system based on artificial intelligence technology, characterized in that: Used to implement the virtual synchronous machine dynamic inertia control method based on artificial intelligence technology as described in any one of claims 1 to 7; The system includes data acquisition layer and control layer; The data acquisition layer is provided with multiple types of sensors for collecting current power grid status information; the data acquisition layer is connected to the control layer via an analog-to-digital conversion unit; The control layer is provided with a heterogeneous multi-processor embedded computing platform for calculating control signals and transmitting the control signals to the voltage source inverter.

9. The virtual synchronous machine dynamic inertia control system based on artificial intelligence technology according to claim 8 is characterized in that: The heterogeneous multi-processor embedded computing platform is a local collaborative architecture, including a digital signal processor and a field programmable gate array; The digital signal processor is used for control and algorithm execution; The field programmable gate array is used for hardware acceleration.

10. The virtual synchronous machine dynamic inertia control system based on artificial intelligence technology according to claim 8 or 9, characterized in that: Also includes servers; The server is connected to the control layer and is used for deploying and online updating of the temporal neural network.

Images