Artificial intelligence driving cascade direct-hanging SVG carrier phase shift optimization method and system

Abstract

The invention provides an artificial intelligence driving cascade direct-hanging SVG carrier phase shift optimization method and system, and the method comprises the steps: constructing a sensor input layer, a slow decision layer and a fast execution layer; multidimensional power grid state information collected by the sensor input layer is transmitted to the slow decision-making layer and the fast execution layer; the slow decision-making layer outputs modulation control parameters based on the reinforcement learning neural network; and the fast execution layer outputs a modulation signal based on the modulation control parameter and the multi-dimensional power grid state information, and is used for driving the cascaded H-bridge module. The system is simple in structure and low in control and construction cost, a fast and slow double-layer coordination framework is constructed, and a reinforcement learning neural network is combined in a method level, so that the nonlinear fitting capability is realized, dynamic parameter adjustment and collaborative optimization can be carried out for multiple dimensions and multiple targets, the stability of a cascade structure is enhanced, and the stability of the cascade structure is improved. And the comprehensive performance of the SVG is improved.

Description

Technical Field

[0001] The present invention relates to the field of power supply technology, and in particular to an artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method and system. Background Art

[0002] As the demand for dynamic regulation of power quality in medium and high voltage power grids increases, the defects of traditional control strategies and modulation methods in terms of dynamic adaptability, multi-parameter collaborative optimization and real-time control efficiency are becoming increasingly prominent, as follows: 1. The static and adaptability of traditional control strategies are insufficient Dynamic response hysteresis: Traditional SVG control that relies on PID or fixed rules requires manual preset parameters. When the grid load suddenly changes or fails, the adjustment is delayed and the response time exceeds 10ms, resulting in reduced reactive compensation accuracy. Limited harmonic suppression capability: In nonlinear load scenarios, it is difficult to track harmonic spectrum changes in real time, and the output current THD exceeds 5%, affecting power quality; Poor robustness: When faced with changes in grid impedance or voltage fluctuations, fixed control rules are prone to instability and require frequent manual intervention.

[0003] 2. Challenges of fixation and coupling of CPS-SPWM modulation parameters Complex parameter coupling: There are nonlinear interactions between parameters such as phase shift angle and modulation ratio. Traditional empirical adjustment methods are difficult to balance conflicts among multiple objectives such as harmonic distribution and switching frequency. Dynamic modulation failure: When the SVG output power changes rapidly, fixed parameters are prone to over-modulation or under-modulation, causing DC side capacitor voltage fluctuations, with the amplitude reaching ±15% of the rated value in some scenarios; Insufficient harmonic optimization: The fixed phase shift strategy cannot adaptively suppress specific harmonics, increasing the cost of filter design.

[0004] 3. The contradiction between computational complexity and real-time performance of multi-parameter collaborative optimization High-dimensional search space: Cascading SVGs requires coordination of dozens of parameters, such as H-bridge modulation wave phase, current loop gain, etc. Traditional exhaustive or gradient descent methods take more than 100ms for a single optimization, which cannot meet real-time requirements. Local optimal trap: Traditional optimization methods such as genetic algorithms are prone to fall into local optimality, resulting in reduced system efficiency; Lack of online updates: Relying on offline optimization, it is impossible to dynamically adjust parameters based on real-time operating data, such as temperature and device aging.

[0005] In addition to the above problems with traditional control methods, the cascade direct-hanging structure also has unique technical difficulties, specifically: capacitor voltage balancing problem: the traditional sorting voltage balancing method relies on high-frequency switching action, and the switching loss accounts for more than 30% of the total system loss. The risk of voltage balancing failure is high in dynamic scenarios; poor module fault tolerance: when a single H-bridge fails, the control strategy cannot be reconstructed through parameter adaptiveness, resulting in reduced system availability; common-mode voltage and electromagnetic interference: the fixed modulation strategy aggravates the common-mode voltage, causing electromagnetic compatibility problems, and additional hardware suppression measures are required. At the hardware level, there is also the problem of insufficient trade-off between system-level energy efficiency and life, such as the loss-harmonic paradox: reducing switching losses requires reducing switching frequency, but it will aggravate harmonics, and traditional methods cannot dynamically balance; thermal stress accumulation: the fixed parameter strategy causes excessive fluctuations in IGBT junction temperature under heavy load, shortening the life of the device; capacitor aging is ignored: the voltage balancing strategy is not adjusted according to the health status of the capacitor, which accelerates the attenuation of the capacitor value.

[0006] Therefore, in the face of the above problems, the existing technology has introduced artificial intelligence technology, which has developed rapidly in recent years. However, it also has corresponding application difficulties, such as strong data dependence: neural networks require a large amount of high-quality data, actual power grid fault data is scarce, and the model generalization ability is insufficient; another example is low online learning efficiency: traditional reinforcement learning algorithms converge slowly, and even require thousands of iterations, which is difficult to meet the millisecond-level decision-making requirements of SVG. Furthermore, the introduction of artificial intelligence technology also lacks security and interpretability: black box models may output over-limit control instructions, such as over-modulation problems, and it is difficult to pass the power grid compliance verification. Furthermore, although the simple mounting of artificial intelligence technology at the hardware level can solve some problems, it still faces inherent factors and requirements of the scene, especially: (1) Strong coupling of multiple parameters: CPS-SPWM modulation ratio, carrier frequency, phase difference and other parameters have nonlinear interactions, which increases the difficulty of optimization. (2) Control delay constraints: From state acquisition to parameter update, it must be completed within 50μs, which requires extremely high real-time computing capabilities. (3) Memory limitations: The neural network model needs to be compressed into the DSP on-chip RAM, which restricts the model complexity and accuracy.

[0007] In summary, the traditional methods have significant defects in dynamic adaptability, multi-parameter coupling optimization, real-time computing efficiency and system-level multi-objective trade-offs. It is urgent to achieve dynamic optimization of cascaded SVG modulation parameters and improve system performance through improved control methods and systems. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a system with simple structure and low cost. The method can be combined with a neural network of reinforcement learning to achieve dynamic optimization of cascaded SVG modulation parameters and improve system performance.

[0009] To achieve the above purpose, the technical solution adopted by the present invention is as follows: On the one hand, the present invention provides an artificial intelligence driven cascade direct-mounted SVG carrier phase shift optimization method, which mainly includes the following steps: Construct the sensor input layer, slow decision layer, and fast execution layer; The sensor input layer collects multi-dimensional grid state information and transmits the multi-dimensional grid state information to the slow decision layer and the fast execution layer; Constructing a strategy mathematical model for the slow decision layer, and training and deploying a reinforcement learning neural network based on the strategy network in the strategy mathematical model, wherein the reinforcement learning neural network outputs a modulation control parameter based on the multi-dimensional power grid state information; A signal generation mathematical model is constructed for the fast execution layer, and the signal generation mathematical model outputs a modulation signal based on the modulation control parameter and the multi-dimensional power grid state information, so as to drive a cascaded H-bridge module.

[0010] Optionally, the following steps are also included: The reinforcement learning neural network is updated based on the knowledge distillation technology, and the reinforcement learning neural network is lightweighted by dynamic weights; The dynamic weight adjustment mechanism is based on the training gradient amplitude of the reinforcement learning neural network, and the model compression adopts an iterative pruning algorithm.

[0011] Optionally, the training of the reinforcement learning neural network includes the following steps: Use full-precision training on the server side to generate a full-precision training model and optimize the multi-objective loss function; The full-precision training model simulates the inference error of the embedded device through pseudo-quantization nodes, and maintains full-precision training for the quantization-sensitive layer in the model to obtain a reinforcement learning neural network after quantization-aware training.

[0012] Optionally, the deployment of the reinforcement learning neural network includes the following steps: Before deployment, the reinforcement learning neural network is quantized to 8-bit fixed-point operations, and the critical path retains floating-point calculations; After deployment, the reinforcement learning neural network accelerates matrix operations through a coprocessor.

[0013] Optionally, the multi-dimensional grid state information includes total harmonic distortion, DC bus voltage change rate, thermal stress coefficient and grid frequency offset; The modulation control parameters include modulation ratio and carrier frequency.

[0014] Optionally, the calculation of the modulation signal includes the following steps: Based on the multi-dimensional grid state information and the modulation control parameter, solving the carrier phase value by a direct digital frequency synthesis algorithm; Mapping the carrier phase value to a triangular wave lookup table to obtain a carrier waveform, and outputting a carrier signal; The phase difference is calculated by a THD real-time algorithm, and a multi-channel phase difference is generated by a nonlinear algorithm; each channel of the multi-channel phase difference corresponds to one channel of the cascaded H-bridge module; Based on the carrier signal and the multi-path phase difference, a plurality of pairs of complementary PWM signals are calculated by a modulation signal generation algorithm; The multiple pairs of complementary PWM signals are the modulation signals.

[0015] Optionally, the THD real-time algorithm includes a sliding window Fourier transform algorithm, and the sliding window Fourier transform algorithm is used to calculate the total harmonic distortion rate of the current power grid system in real time.

[0016] On the other hand, the present invention also provides an artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization system, which is used to implement the above method; The system includes a sensor input layer, a slow decision layer, and a fast execution layer; The sensor input layer is provided with multiple types of sensors for collecting current power grid status information; the sensor input layer is connected to the slow decision layer through an analog-to-digital conversion unit; The slow decision layer is provided with a digital signal processor, and the slow decision layer is deployed with a reinforcement learning neural network; the slow decision layer is also connected to the fast execution layer; The fast execution layer is provided with a field programmable gate array, and the fast execution layer is connected to the cascade H-bridge module of the current power grid; The digital signal processor and the field programmable gate array form a collaborative architecture and are connected via a data bus.

[0017] Optionally, it further includes a server, which is connected to the slow decision layer and is used to deploy and update the reinforcement learning neural network online; The slow decision layer is also configured with a coprocessor for accelerating calculations.

[0018] Optionally, the fast execution layer is further configured with a multi-phase DDS generator, a real-time THD calculation unit, a phase difference adjustment module and a PWM generation module; The multi-phase DDS generator is used to independently generate carrier signals of different phases; The real-time THD calculation unit is used to calculate the phase difference; The phase difference adjustment module is used to generate multiple phase differences, and each of the multiple phase differences corresponds to one of the cascaded H-bridge modules; The PWM generation module is used to generate a PWM signal; A triangle wave lookup table is stored in the block random access memory of the field programmable gate array.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The system structure of the present invention is simple and the control and construction costs are low. By constructing a fast and slow double-layer coordinated framework and combining it with a reinforcement learning neural network at the method level, nonlinear fitting capability is achieved. Dynamic parameter adjustment and collaborative optimization can be performed for multiple dimensions and multiple targets, thereby improving the stability of the cascade structure and enhancing the comprehensive performance of SVG. 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 architecture diagram in a specific embodiment of the present invention; Figure 2 It is a schematic diagram of the process of pre-training and deployment phase in a specific embodiment of the present invention; Figure 3 This is a comparison diagram of THD recovery curves under load mutation between the solution of the present invention and the traditional solution in comparative example 1; Figure 4 This is a comparison diagram of switching losses under load mutation between the solution of the present invention and the traditional solution in comparative example 1; Figure 5 This is a comparison chart of the dynamic response time under load mutation of the solution of the present invention and the traditional solution in comparative example 1. 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 is worth noting 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, and their sources are not specifically limited unless otherwise specified.

[0026] On the one hand, this embodiment provides an artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method, which constructs a sensor input layer, a slow decision layer and a fast execution layer.

[0027] Among them, Figure 1 As shown, the sensor input layer is constructed based on the existing detection sensors, and the voltage / current sensor, multi-channel temperature sensor, grid frequency detection sensor and DC bus voltage sensor are connected to other layers. Specifically, the sensor input layer collects the above-mentioned multi-dimensional grid state information; correspondingly, the multi-dimensional grid state information includes total harmonic distortion rate, DC bus voltage change rate, thermal stress coefficient and grid frequency offset. The multi-dimensional grid state information is transmitted to the slow decision layer and the fast execution layer.

[0028] The slow decision layer is deployed with a reinforcement learning neural network, which outputs modulation control parameters based on multi-dimensional grid state information.

[0029] Among them, in order to formally describe the control decision-making process of the cascade SVG system at this layer, it is necessary to build a theoretical framework, which is used to convert the operating state, control strategy and optimization goal of the physical system into a computable mathematical expression. It has never been possible to input and output specifications for the reinforcement learning neural network model and set constraints that meet the requirements. Therefore, a strategy mathematical model is built for the slow decision layer, including: Establish state observation phasors based on multi-dimensional power grid state information , the formula is: ; Where THD is the total harmonic distortion (THD), is the DC bus voltage change rate, is the thermal stress coefficient, is the grid frequency offset, is a 4-dimensional set of real numbers.

[0030] The strategy network mapping is established based on the state observation phasor, which is expressed as: ; In the formula, is a parameterized policy network, is the network weight, is the action space constraint, m is the modulation ratio, is the carrier signal frequency, are the carrier phase offsets between cascaded H-bridge modules. Furthermore, the cascaded SVG is usually composed of multiple H-bridge modules connected in series. For example, in this embodiment, a 7-way design is taken as an example, which includes a 7-level structure, corresponding to 7 H-bridges. The core of CPS-SPWM is to cancel out the harmonics of the total output voltage by orderly offsetting the carrier phase of each H-bridge. For N cascaded modules, the traditional method requires setting N-1 independent phase differences (phase offsets between adjacent modules), but in this embodiment, the independent variables are simplified to 2 through a symmetrical grouping strategy. Specifically, the symmetrical grouping strategy divides the 7 H-bridges into two groups, such as the first 3 groups and the last 3 groups. The middle group is used as a benchmark, and the modules in each group share the same base phase difference. , reducing independent parameters through symmetry.

[0031] State Observation Vector in Mathematical Modeling It is the model input of the reinforcement learning neural network, which clarifies the physical quantities that need to be collected. It is the output range of the model, which constrains the feasible solution space of the parameters.

[0032] The policy network mapping is used as a standard decision logic structure, and the reinforcement learning neural network is used as the execution structure. The mathematical form of the policy network, such as the number of neural network layers and activation functions, determines the architecture of the model. For example, a 3-layer fully connected network is used, with 4-dimensional input → 32-dimensional → 16-dimensional → 4-dimensional output, combined with the ReLU activation function to achieve nonlinear mapping from state to parameters. Constrained optimization objectives: Multiple objectives in mathematical modeling, such as THD minimization and loss balance, are achieved through model training. For example, the loss function in online transfer learning directly corresponds to the optimization objective in mathematical modeling, ensuring that the model output parameters meet the dual requirements of harmonic suppression and efficiency improvement.

[0033] Therefore, if Figure 2 As shown in Figure 1, the training and deployment of a reinforcement learning neural network includes the following steps: First, in the pre-training stage, the server uses full-precision training to generate a full-precision training model and optimizes the multi-objective loss function; that is, the initial model training is completed under FP32 precision, and the multi-objective loss function is optimized at the same time. Specifically, the model training adopts a mixed precision training strategy, and uses a GPU / CPU cluster to perform FP32 full-precision training on the model during the server-side pre-training stage, optimizing the multi-objective loss function including THD and loss; Before deployment, the reinforcement learning neural network is quantized to 8-bit fixed-point operations, and floating-point calculations are retained in the critical path. Optionally, the full-precision training model simulates the inference error of the embedded device through pseudo-quantization nodes, and maintains full-precision training for the quantization-sensitive layers in the model to obtain the reinforcement learning neural network after quantization-aware training. Specifically, pseudo-quantization nodes are then inserted for quantization-aware training to simulate INT8 inference errors, while maintaining FP32 precision training for quantization-sensitive layers such as the policy network output layer, and then the FP32 weights are converted to INT8 and verified to ensure that the THD error is <2%.

[0034] After deployment, the reinforcement learning neural network accelerates matrix operations through the coprocessor, specifically using the CLA coprocessor to accelerate INT8 reasoning, improving computing efficiency through fixed-point format data and dynamic quantization scaling, and maintaining FP32 fixed-point operations for key modules such as thermal stress prediction at the next level. The quantization parameters are stored in BRAM and support dynamic precision switching. Finally, different quantization models such as high-frequency dynamics and thermal safety are dynamically loaded through code according to working conditions such as the grid frequency change rate and thermal stress gradient, realizing deep collaborative optimization of control strategies and hardware architectures, so that the reinforcement learning neural network model meets the RAM ≤ 256KB storage limit and ≤ 35μs inference delay in the slow decision-making layer hardware, while ensuring THD error < 2%, strategy similarity of 91.3% and other performance.

[0035] In order to improve the accuracy and adaptability of the model to multiple working conditions, this embodiment also introduces an online transfer learning mechanism. The server dynamically adjusts the student network through the knowledge distillation framework, as follows: Based on the knowledge distillation technology, the reinforcement learning neural network is updated and the model is lightweighted through dynamic weights. A lightweight knowledge distillation framework is constructed during training. The loss function consists of a Q-value matching term (mean square error) to ensure the consistency of the value function and a strategy distribution alignment term to retain the diversity of the teacher's strategy decision. The formula is as follows: ; In the formula, are the Q-value functions of the teacher and student networks respectively; is the strategy distribution of teachers and students, that is, in the state s Next select action The probability of is the dynamic weight coefficient; is the Jensen-Shannon divergence, which is used to measure the difference in strategy distribution; State-Action Distribution D Find the expectation, that is, the statistical average based on historical data or real-time sampling data.

[0036] Dynamic Weight The formula is: ; In the formula, For the Sigmoid function, the input value is compressed to [0,1] to ensure Dynamically change within the effective range; τ is a sliding time window, such as a typical value of 100ms, used to calculate the past τ The average gradient amplitude over time avoids the influence of instantaneous noise and smoothes the weight adjustment process; Network parameters for students θ The gradient amplitude reflects the current learning difficulty. The larger the gradient, the more difficult the learning.

[0037] The dynamic weight adjustment mechanism is based on the training gradient amplitude of the reinforcement learning neural network. In the dynamic weight, when the gradient amplitude of the student network increases (learning difficulty), the Q value item weight is automatically increased. , strengthen supervised learning. In the mathematical property verification, Lyapunov stability analysis proves that dynamic weight adjustment can ensure: ; In the formula, is a non-negative constant.

[0038] During the knowledge distillation training process, the dynamic weights are adjusted in real time according to the learning status of the student network. Through Lyapunov stability analysis, it can be theoretically proved that under this dynamic adjustment mechanism, the training process of the student network is convergent and will not oscillate or diverge, thus ensuring the effectiveness and stability of the model training.

[0039] Therefore, in the optimization of the multi-objective loss function, THD and loss are minimized simultaneously through gradient descent, achieving a balance between "harmonic suppression" and "energy efficiency optimization".

[0040] The weight is adjusted according to the gradient amplitude of the student network, and the weight of the Q-value item is increased to strengthen supervised learning when learning is difficult. Among them, the model compression adopts an iterative pruning algorithm, that is, iterative amplitude pruning is used to make the student model parameter amount ≤ 30% of the teacher model to meet the lightweight constraint. Among them, in this embodiment, in order to lightweight design, it is also necessary to establish corresponding hardware resource constraints to ensure that the student network can run efficiently on embedded hardware and balance model performance with hardware resource limitations, such as computing speed, storage capacity, etc. Its core includes model complexity constraints and computing delay constraints, and the formula is as follows: Model complexity constraints: ; Calculate the latency constraint: ; In the formula, is the parameter set of the student network, is the parameter set of the teacher network; It is the single inference time of the model, which is the computation time from input state to output action.

[0041] In model compression, we construct a sparse constraint and use iterative magnitude pruning to satisfy constraint: ; In the formula, is a dynamic threshold, which increases in each round; is the mask matrix, are the elements of the mask matrix.

[0042] The above method steps are aimed at the reinforcement learning neural network in the slow decision layer, and together constitute a complete link of "model training → deployment adaptation → real-time optimization". The core is to achieve efficient operation of the reinforcement learning strategy on a resource-constrained embedded platform through a combination of static compression before deployment and dynamic tuning of online migration after deployment, ultimately achieving performance improvements such as THD reduction and switching loss reduction.

[0043] As can be seen from the above content, the modulation control parameters output by the slow decision layer of this embodiment are provided with parameters such as modulation ratio and carrier frequency, and are input into the fast execution layer together with the aforementioned multi-dimensional grid status information.

[0044] Since the slow decision layer and the fast execution layer are a collaborative architecture, a corresponding signal generation mathematical model is constructed in the fast execution layer, which is a high-frequency dynamic optimization model that focuses on quickly adjusting the aforementioned phase difference parameters of CPS-SPWM modulation to achieve real-time harmonic suppression.

[0045] Furthermore, the modulation signal generation function as follows: ; In the formula, input: modulation parameters ; Output: 7 pairs of complementary PWM signals, a binary matrix of dimension 7x2. The purpose of the function is to convert continuous modulation parameters into discrete PWM pulse sequences and achieve harmonic optimization through carrier phase shifting technology.

[0046] Therefore, the calculation of the modulation signal includes the following steps: (1) Carrier phase generation; Based on the modulation control parameters and multi-dimensional grid status information output by the upper layer, the carrier phase is generated through a multi-phase direct digital frequency synthesis (DDS) algorithm: a 17-bit phase accumulator is adapted, a 200MHz clock is used, and 2000 phase accumulations are completed within a 10μs fast adjustment layer update cycle to ensure phase adjustment resolution.

[0047] The high 7 bits are extracted from the 17-bit accumulator as the base phase, and combined with the offset to generate 7 sets of independent carrier phase offsets, each corresponding to an H-bridge module, to achieve precise phase shifting between adjacent carriers.

[0048] (2) Carrier signal generation; The carrier phase value of each H-bridge is mapped to a 12-bit resolution triangle wave lookup table, which is pre-stored in the BRAM of the fast execution layer. The carrier waveform is generated by looking up the table through the phase offset index; Given by the output of the upper strategy network, the phase accumulation step size is based on Dynamic adjustment.

[0049] Seven carrier signals are output, and the phase difference of each signal is bound one by one to the hardware channel of the corresponding H-bridge module.

[0050] (3) Real-time THD calculation and phase difference feedback; Synchronously sample the voltage / current signal, use the fast Fourier transform (FFT) algorithm to perform harmonic analysis on the waveform data in the sliding window, and calculate the amplitude of each harmonic. The formula is: ; In the formula, is the real-time THD value, For the time window [ t-τ,t ] in the waveform data.

[0051] The THD calculation results are fed back to the upper layer as the input of the reinforcement learning neural network. The phase difference parameters are generated by the policy network and dynamically adjusted through multi-objective optimization.

[0052] (4) PWM signal generation; Based on the carrier signal and the modulation ratio of the output, a unipolar PWM signal is generated by comparing the modulation wave with the carrier. When the modulation wave amplitude is greater than the carrier amplitude, a high level (1) is output, otherwise a low level (0) is output. A pair of complementary PWM signals is generated for each H-bridge, i.e., the upper tube / lower tube drive. Hardware dead time is inserted to avoid direct conduction of the bridge arm. Finally, 7 pairs of complementary PWM signals are output to drive the cascaded H-bridge module.

[0053] On the other hand, the present embodiment further provides an artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization system, which is used to implement the aforementioned artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method.

[0054] The system of this embodiment includes a sensor input layer, a slow decision layer, and a fast execution layer; The sensor input layer is equipped with multiple types of sensors to collect the current power grid status information; the sensor input layer is connected to the slow decision layer through an analog-to-digital conversion unit; The slow decision layer is equipped with a digital signal processor (DSP), and the slow decision layer is deployed with a reinforcement learning neural network; the slow decision layer is also connected to the fast execution layer; optionally, the DSP of this embodiment uses TMS320F28337, with a built-in 16-bit ADC, a sampling rate of 1MSPS, and synchronous acquisition of voltage / current / temperature / grid frequency / DC bus voltage; this layer also integrates a 50MHz CLA coprocessor, completes 16×16 fixed-point multiplication and accumulation in a single cycle, and sets RAM≤256KB on the chip to store lightweight neural network models. Furthermore, the DSP also has a built-in ADC to synchronously acquire multi-sensor signals, with a sampling rate of 1MHz and synchronized with the lower clock domain.

[0055] The fast execution layer is equipped with a field programmable gate array (FPGA), and the fast execution layer is connected to the cascade H-bridge module of the current power grid; optionally, the FPGA uses XC7K70T, works in the 200MHz clock domain, occupies 58% of the LUTs resources, occupies 72% of the BRAM resources, and the BRAM stores the 12-bit resolution triangle wave lookup table and quantization parameters, and uses DSP48E1 Slice to implement floating-point operations; further, the FPGA implements phase generation through Verilog, that is, a 17-bit phase accumulator, inputs a 16-bit phase difference, accumulates after sign extension, and outputs 7 phase offsets; BRAM stores the triangle wave lookup table to generate 7 carrier signals, and outputs 7 pairs of complementary PWM signals after modulation wave-carrier comparison, each pair drives an upper and lower bridge arm of an H-bridge module, and inserts a 2μs dead time. The FPGA also generates the carrier phase through multi-phase DDS technology, and uses DSP48E1 Slice to implement 128-point FFT to calculate the 50μs sliding window THD, and the result is transmitted back to the DSP via GPIO.

[0056] The FPGA of the fast execution layer is also equipped with a multi-phase DDS generator, a real-time THD calculation unit, a phase difference adjustment module, and a PWM generation module. The multi-phase DDS generator is used to independently generate carrier signals of different phases; the real-time THD calculation unit is used to calculate the phase difference; the phase difference adjustment module is used to generate multiple phase differences, and each phase difference corresponds to one phase of the cascaded H-bridge module; the PWM generation module is used to generate a PWM signal.

[0057] Optionally, the FPGA also integrates a hardware protection module, which directly shuts down the PWM output within a preset time when overcurrent or overtemperature is detected.

[0058] DSP+FPGA form a collaborative architecture and are connected through the EMIF bus to achieve 100MB / s bandwidth for interactive modulation parameters and status data, supplemented by 10MHz SPI communication to transmit configuration instructions.

[0059] The system also includes a server locally or in the cloud, which is connected to the slow decision layer and is used to deploy and / or online update the reinforcement learning neural network.

[0060] Comparative Example 1; In a specific embodiment, a system using the method of the present invention and a corresponding system using the traditional method are simulated for the same power grid environment, thereby forming a comparison between the solution of the present invention and the traditional solution. The simulation simulates a sudden load change, the time is uniformly set at 0.5ms, and a 50%→100% step change is adopted. Then, the THD recovery performance is tested through the simulation results, such as Figure 3 As shown in Table 1 below.

[0061] Table 1 From the graphs and tables, we can see that in the traditional solution, the THD peak reaches 5.0% at 0.7ms after the mutation, and it takes 2.1ms to recover to 3%. In the solution of the present invention, the THD peak reaches 3.5% at 0.8ms after the mutation, and it takes 0.8ms to recover to 2.8%. Therefore, it can be seen that the solution of the present invention has extremely strong harmonic suppression ability and dynamic response speed, and is significantly higher than the traditional solution, achieving a 30% reduction in THD peak and a 61.9% reduction in recovery time.

[0062] Furthermore, the switching loss and response time are tested in the simulation to form Figure 4 , Figure 5 and Table 2 below.

[0063] Table 2 From the graphs and tables, we can see that in the comparison of switching losses, the dynamic loss of the solution of the present invention is reduced by 18.3% and the peak loss is reduced by 53% compared with the traditional solution, which shows that the solution of the present invention has the effect of dynamic adjustment of carrier frequency and the effect of suppressing loss during load mutation. The solution of the present invention also reduces the dynamic response time from 120μs of the traditional solution to 28μs through the FPGA hardware acceleration effect, which is a reduction of 76.7%.

[0064] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. An artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method, characterized in that: The steps include: Construct the sensor input layer, slow decision layer, and fast execution layer; The sensor input layer collects multi-dimensional grid state information and transmits the multi-dimensional grid state information to the slow decision layer and the fast execution layer; Constructing a strategy mathematical model for the slow decision layer, and training and deploying a reinforcement learning neural network based on the strategy network in the strategy mathematical model, wherein the reinforcement learning neural network outputs a modulation control parameter based on the multi-dimensional power grid state information; A signal generation mathematical model is constructed for the fast execution layer, and the signal generation mathematical model outputs a modulation signal based on the modulation control parameter and the multi-dimensional power grid state information, so as to drive a cascaded H-bridge module.

2. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 1 is characterized by: The following steps are also included: The reinforcement learning neural network is updated based on the knowledge distillation technology, and the reinforcement learning neural network is lightweighted by dynamic weights; The dynamic weight adjustment mechanism is based on the training gradient amplitude of the reinforcement learning neural network, and the model compression adopts an iterative pruning algorithm.

3. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 1 is characterized by: The training of the reinforcement learning neural network includes the following steps: Full-precision training is used on the server side to generate a full-precision training model and optimize the multi-objective loss function; The full-precision training model simulates the inference error of the embedded device through pseudo-quantization nodes, and maintains full-precision training for the quantization-sensitive layer in the model to obtain a reinforcement learning neural network after quantization-aware training.

4. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 3 is characterized by: The deployment of the reinforcement learning neural network includes the following steps: Before deployment, the reinforcement learning neural network is quantized to 8-bit fixed-point operations, and the critical path retains floating-point calculations; After deployment, the reinforcement learning neural network accelerates matrix operations through a coprocessor.

5. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 1 is characterized by: The multi-dimensional grid state information includes total harmonic distortion, DC bus voltage change rate, thermal stress coefficient and grid frequency offset; The modulation control parameters include modulation ratio and carrier frequency.

6. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 5 is characterized by: The calculation of the modulation signal comprises the following steps: Based on the multi-dimensional grid state information and the modulation control parameter, solving the carrier phase value by a direct digital frequency synthesis algorithm; Mapping the carrier phase value to a triangular wave lookup table to obtain a carrier waveform, and outputting a carrier signal; The phase difference is calculated by a THD real-time algorithm, and a multi-channel phase difference is generated by a nonlinear algorithm; each channel of the multi-channel phase difference corresponds to one channel of the cascaded H-bridge module; Based on the carrier signal and the multi-path phase difference, a plurality of pairs of complementary PWM signals are calculated by a modulation signal generation algorithm; The multiple pairs of complementary PWM signals are the modulation signals.

7. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization method according to claim 6 is characterized by: The THD real-time algorithm includes a sliding window Fourier transform algorithm, and the sliding window Fourier transform algorithm is used to calculate the total harmonic distortion rate of the current power grid system in real time.

8. An artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization system, characterized by: Used to implement the artificial intelligence driven cascade direct-mounted SVG carrier phase shift optimization method described in any one of claims 1 to 7; The system includes a sensor input layer, a slow decision layer, and a fast execution layer; The sensor input layer is provided with multiple types of sensors for collecting current power grid status information; the sensor input layer is connected to the slow decision layer through an analog-to-digital conversion unit; The slow decision layer is provided with a digital signal processor, and the slow decision layer is deployed with a reinforcement learning neural network; the slow decision layer is also connected to the fast execution layer; The fast execution layer is provided with a field programmable gate array, and the fast execution layer is connected to the cascade H-bridge module of the current power grid; The digital signal processor and the field programmable gate array form a collaborative architecture and are connected via a data bus.

9. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization system according to claim 8 is characterized by: Also included is a server, the server is connected to the slow decision layer, and is used to deploy and update the reinforcement learning neural network online; The slow decision layer is also configured with a coprocessor for accelerating calculations.

10. The artificial intelligence driven cascaded direct-mounted SVG carrier phase shift optimization system according to claim 8, characterized in that: The fast execution layer is also configured with a multi-phase DDS generator, a real-time THD calculation unit, a phase difference adjustment module and a PWM generation module; The multi-phase DDS generator is used to independently generate carrier signals of different phases; The real-time THD calculation unit is used to calculate the phase difference; The phase difference adjustment module is used to generate multiple phase differences, and each of the multiple phase differences corresponds to one of the cascaded H-bridge modules; The PWM generation module is used to generate a PWM signal; A triangle wave lookup table is stored in the block random access memory of the field programmable gate array.

Images