A power quality management method and system based on the collaboration of self-power supply and power electronic module

Abstract

The present invention discloses a power quality management method and system based on the collaboration of self-power supply and power electronic modules. The method comprises: dividing the operating scenarios into high-voltage scenarios, low-voltage scenarios, three-phase unbalance scenarios and harmonic pollution scenarios according to the power quality problems in the substation area, and adopting different response strategies respectively; if it is a high-voltage scenario, adopting a reinforcement learning algorithm to optimize the reactive output strategy of SVG; if it is a low-voltage scenario, adopting a proximal strategy to optimize the energy storage scheduling optimization strategy of PPO; if it is a three-phase unbalanced scenario, adopting a current transfer optimization strategy of a convolutional neural network (CNN); if it is a harmonic pollution scenario, adopting a harmonic compensation current generation strategy of a generative adversarial network (GAN); the present invention solves the problems of increased line loss, severe voltage fluctuation, three-phase imbalance and severe harmonic pollution through a self-power supply module and a power electronic module.

Description

Technical Field

[0001] The present invention relates to a method and system for managing power quality, and in particular to a method and system for managing power quality based on collaboration between self-power supply and power electronic modules. Background Art

[0002] Distributed renewable energy has been widely adopted due to its environmental and energy-saving advantages, particularly in the photovoltaic and wind power sectors. According to statistics, the penetration rate of distributed renewable energy in the power grid is expected to reach over 20% by 2025.

[0003] However, the integration of large numbers of distributed photovoltaic systems into the grid has also led to significant power quality issues, particularly in substations (areas powered by distribution transformers). During days of abundant sunlight, full photovoltaic power generation leads to power backflow, making it difficult to eliminate high voltage in the substation even with the transformer set to its lowest setting. At night, when photovoltaic power generation is halted and power load surges, low voltage is common at the end of the substation. This "high during the day, low at night" voltage instability and fluctuation has become a major cause of deteriorating regional power quality. The intermittent output of photovoltaic systems increases line losses in the substation, especially when photovoltaic output fluctuates significantly. The integration of distributed photovoltaic systems often leads to three-phase current imbalance in the substation, further exacerbating line losses and equipment wear. The widespread use of power electronic equipment such as photovoltaic inverters has led to increasingly serious harmonic pollution problems in the substation, impacting the safe operation of electrical equipment.

[0004] While traditional SVGs can address some power quality issues, they suffer significant losses during operation (full-load losses ≥ 3%), further increasing line losses in substations. Furthermore, SVGs have a single control strategy and insufficient dynamic sensing and response capabilities, making them difficult to adapt to the complex and changing substation operating environment. Retrofitting older substations is costly, requiring a long payback period, and existing equipment lacks compatibility and scalability, making it difficult to meet the needs of future power system development. Existing power quality management devices typically rely on external power grids, increasing operating costs and grid dependence, which is inconsistent with the development trend of green energy. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a power quality management method to reduce line losses, stabilize voltage fluctuations, achieve three-phase balance and filter harmonics, and on the other hand, to provide a power quality management system.

[0006] Technical solution: The power quality control method of the present invention includes:

[0007] Based on the power quality issues in the substation, the operation scenarios are divided into high voltage scenario, low voltage scenario, three-phase imbalance scenario and harmonic pollution scenario;

[0008] In high-voltage scenarios, a reinforcement learning algorithm is used to optimize the SVG's reactive power output strategy.

[0009] In low-voltage scenarios, the proximal strategy is used to optimize the PPO energy storage scheduling optimization strategy;

[0010] In the case of three-phase unbalanced scenarios, a convolutional neural network (CNN) is used to extract features from the three-phase current time series and generate a current transfer optimization strategy.

[0011] In the case of harmonic pollution, a harmonic compensation current generation strategy based on generative adversarial network (GAN) is adopted.

[0012] Preferably, the conditions for determining the running scenario are as follows:

[0013] High voltage scenario: V > 1.1V nominal ;

[0014] Low voltage scenario: V < 0.9V nominal ;

[0015] Three-phase unbalance scenario: three-phase current imbalance ΔI unbalance >15%;

[0016] Harmonic pollution scenario: Total harmonic distortion rate THDi>5%.

[0017] Preferably, the reactive power output strategy objective function is as follows:

[0018] ;

[0019] in, Indicates voltage deviation, Indicates line loss power, Indicates the reactive power output of SVG. represents the weight coefficient;

[0020] The constraints are as follows:

[0021] Voltage upper limit: V≤1.1V nominal ;

[0022] SVG capacity limitations: ;

[0023] Self-powered module output: ;

[0024] The penalty function is as follows:

[0025] ;

[0026] Among them, K1 and K2 represent penalty coefficients.

[0027] Preferably, the energy storage scheduling optimization strategy reward function is as follows:

[0028] ;

[0029] ;

[0030] in, represents the voltage deviation penalty, Indicates line loss power penalty, Indicates energy storage SOC, Indicates the penalty for constraint violation;

[0031] The constraints are as follows:

[0032] Voltage lower limit: V ≥ 0.9V nominal ;

[0033] Energy storage charging and discharging power limit: ;

[0034] Energy storage SOC safety range: SOC ≥ 20%;

[0035] SVG capacity limitations: .

[0036] Preferably, the energy storage scheduling optimization strategy includes: optimizing the PPO energy storage scheduling optimization strategy through a proximal strategy to interact with the environment and collect trajectory data; estimating the advantage value using GAE; limiting the strategy update amplitude through a clipping mechanism to maximize the objective function; and minimizing the mean square error loss of the value network.

[0037] Preferably, the objective function of the current transfer optimization strategy is as follows:

[0038] ;

[0039] ;

[0040] in, Indicates the three-phase imbalance weight, represents line loss weight;

[0041] The constraints are as follows:

[0042] Current transfer limit: ;

[0043] Three-phase unbalance safety threshold: ΔI unbalance ≤10%;

[0044] IGBT switching frequency limit: ;

[0045] The penalty function is as follows:

[0046] .

[0047] Preferably, the current transfer optimization strategy includes: optimizing the PPO energy storage scheduling optimization strategy through a proximal strategy to interact with the environment and collect trajectory data; estimating the advantage value using GAE; limiting the strategy update amplitude through a clipping mechanism to maximize the objective function; and minimizing the mean square error loss of the value network.

[0048] Preferably, the objective function of the harmonic compensation current generation strategy is as follows:

[0049] ;

[0050] in, represents the harmonic distortion weight, represents line loss weight;

[0051] The constraints are as follows:

[0052] Harmonic distortion safety threshold: THDi≤5%;

[0053] Compensation current bandwidth limitation: ;

[0054] Inverter capacity limitation: ;

[0055] The penalty function is as follows:

[0056] .

[0057] Preferably, the harmonic compensation current generation strategy includes: collecting harmonic spectrum data from the actual system or simulation and inputting it into the power system simulation tool; using the mean square error (MSE) to preliminarily train the generator to generate an ideal compensation current; alternately training the generator and the discriminator to gradually optimize the generation strategy; and dynamically adjusting the generator parameters through real-time data.

[0058] The power quality management system of the present invention includes:

[0059] A self-powered module, which is used to provide power to the system through photovoltaic power generation and power the power electronic module;

[0060] The power electronics module is used to provide different response strategies based on different operating scenarios. The strategies include the reactive output strategy of SVG optimized by reinforcement learning algorithm, the energy storage scheduling optimization strategy of proximal policy optimization (PPO), the current transfer optimization strategy, and the harmonic compensation current generation strategy of generative adversarial network (GAN).

[0061] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The power electronic module is powered by the self-powered module, reducing dependence on the external power grid, while optimizing the reactive compensation strategy, reducing line loss, and the power generation of the self-powered module is not included in the gateway table, further reducing the line loss rate; 2. By adjusting the phase and amplitude of the AC output voltage of the bridge converter circuit, reactive power is dynamically compensated, the system voltage is stabilized near the set reference value, and voltage fluctuations are controlled; 3. By driving the IGBT, the unbalanced phase current is transferred from the large phase to the small phase, thereby achieving three-phase balance on the grid side; 4. The IGBT power converter generates a compensation current that is equal to the harmonics in the system and opposite in phase, thereby achieving harmonic filtering. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0063] like Figure 1 As shown in the figure, based on the power quality issues in the substation, the operation scenarios are divided into high voltage scenario, low voltage scenario, three-phase unbalance scenario and harmonic pollution scenario;

[0064] Scene determination conditions:

[0065] High voltage scenario: V > 1.1V nominal , the photovoltaic power generation power is greater than the load demand of the substation; during the period of sufficient sunlight during the day (such as 10:00-15:00);

[0066] Low voltage scenario: V < 0.9V nominal , the load surges at night and the photovoltaic power generation stops; the peak load period at night (such as 18:00-22:00);

[0067] Three-phase unbalance scenario: three-phase current imbalance ΔI unbalance >15%, the load distribution in the substation is uneven and the load fluctuation is large;

[0068] Harmonic pollution scenarios: When the total harmonic distortion rate (THDi) is greater than 5%, power electronic equipment is densely populated, and photovoltaic output fluctuates greatly.

[0069] In high-voltage scenarios, a reinforcement learning algorithm optimizes the SVG's reactive power output strategy, dynamically adjusting reactive power compensation based on real-time feedback of voltage changes. Through continuous learning, the system can find the optimal reactive power compensation strategy in complex and changing operating environments.

[0070] Algorithm flow:

[0071] State space: Substation voltage, PV output, and load demand; Action space: SVG reactive output; Reward function: The smaller the voltage deviation, the higher the reward; the lower the line loss, the higher the reward;

[0072] The objective function is as follows:

[0073] ;

[0074] Among them, it means Voltage deviation , Indicates line loss power, Indicates the reactive power output of SVG. represents the weight coefficient;

[0075] The constraints are as follows:

[0076] Voltage upper limit: V≤1.1V nominal ;

[0077] SVG capacity limitations: ;

[0078] Self-powered module output: ;

[0079] The penalty function is as follows:

[0080] ;

[0081] Among them, K1 and K2 represent penalty coefficients.

[0082] Solution: Deep Q Network DQN:

[0083] (1) Neural network structure:

[0084] Input layer: 4 nodes (corresponding to the state space dimension);

[0085] Hidden layer: 2 fully connected layers (24 nodes, ReLU activation);

[0086] Output layer: 3 nodes (corresponding to the Q values ​​of 3 discrete actions);

[0087] (2) Experience replay:

[0088] Buffer size: 2000 experiences (state, action, reward, next state, termination flag);

[0089] Sampling strategy: Randomly sample batches (batch_size=32) for training to break data correlation;

[0090] (3) Greedy strategy:

[0091] Initial exploration rate ϵ = 1.0;

[0092] Minimum exploration rate: ϵmin=0.01;

[0093] Decay rate: ϵdecay=0.995ϵdecay=0.995, gradually reducing the exploration rate to balance exploration and utilization.

[0094] (4) Target network:

[0095] Update frequency: Copy the main network parameters to the target network every 100 steps to stabilize the training process.

[0096] In low-voltage scenarios, a proximal strategy is used to optimize the PPO energy storage scheduling optimization strategy. A surge in nighttime load causes the substation voltage to fall below the rated value, requiring voltage boost through energy storage discharge and SVG capacitive reactive power compensation while minimizing line losses and energy storage lifespan.

[0097] State space: station voltage , load demand , energy storage SOCϵ[0.2, 1.0] (20%~100%), SVG reactive output (kVar);

[0098] Action space: continuous action, energy storage charging and discharging power (Negative value is charging, positive value is discharging);

[0099] Reward function:

[0100] ;

[0101] ;

[0102] in, Indicates voltage deviation penalty (target voltage 0.95V nominal ), Indicates line loss power penalty, Indicates energy storage SOC, change penalty (encourages stable charging and discharging), Indicates constraint violation penalties (e.g., voltage exceeding limit, SOC exceeding limit);

[0103] The constraints are as follows:

[0104] Voltage lower limit: V ≥ 0.9V nominal ;

[0105] Energy storage charging and discharging power limit: ;

[0106] Energy storage SOC safety range: SOC ≥ 20%;

[0107] SVG capacity limitations: .

[0108] PPO optimizes the energy storage scheduling strategy while ensuring training stability through importance sampling and strategy clipping;

[0109] Neural network structure: Policy network (Actor), outputs the mean and standard deviation (Gaussian distribution) of energy storage charge and discharge power;

[0110] The input layer has 4 nodes (corresponding to the state space); the hidden layer has 2 fully connected layers (64 nodes, ReLU activation);

[0111] Output layer 1 node (mean ) + 1 node (standard deviation );

[0112] The value network (Critic) evaluates the value of the state:

[0113] The input layer has 4 nodes; the hidden layer has 2 fully connected layers (64 nodes, ReLU activation);

[0114] Output layer 1 node (state value V(s)).

[0115] Algorithm steps:

[0116] (1) Collect trajectory data by interacting with the environment through the current strategy;

[0117] (2) Use GAE (Generalized Advantage Estimation) to estimate the advantage value ;

[0118] (3) Limit the strategy update range through the clipping mechanism to maximize the objective function:

[0119] ;

[0120] (4) Update the value function to minimize the mean square error loss of the value network;

[0121] (5) Hyperparameter setting: discount factor =0.99, GAE parameters =0.95, clipping threshold ϵ=0.2, learning rate , batch size 64, and number of training rounds 1000.

[0122] Training steps:

[0123] (1) Initialization environment: set initial voltage, load demand, and energy storage SOC;

[0124] (2) Collect trajectory data (state, action, reward, next state) generated by the current strategy;

[0125] (3) Use GAE to estimate the advantage value of each state ;

[0126] (4) Update the Actor and Critic networks through the PPO pruning mechanism;

[0127] (5) Repeat steps 2-4 until convergence.

[0128] Key implementation details:

[0129] Action mapping, which outputs actions Linear mapping to actual charge and discharge power:

[0130] ;

[0131] Add a penalty term for SOC exceeding the limit to the reward function:

[0132] .

[0133] Reward function adjustment: Increase the reward for voltage recovery speed to avoid excessive discharge of energy storage; exploration strategy: Add noise in the early stage of training to improve strategy diversity; parallel training: Use multiple environments to collect data in parallel to accelerate the training process.

[0134] Take a low-voltage area in Jiangsu as an example:

[0135] Initial state: voltage 0.88 pu ., load 300 kW, energy storage SOC=50%;

[0136] Results after training:

[0137] The voltage returns to 0.95 pu ., line loss reduced by 15%; energy storage SOC dynamically adjusted, without triggering the lower limit (SOC>25%);

[0138] Compared with traditional strategies:

[0139] The PPO strategy shortens voltage recovery time by 30% and reduces energy storage life loss by 20%.

[0140] Through the PPO algorithm, the system can intelligently dispatch the energy storage charging and discharging power and SVG reactive output, optimizing line losses and energy storage life while increasing voltage. Next, it can be combined with photovoltaic output forecasts to achieve multi-timescale optimization, further improving economy and reliability.

[0141] In the case of a three-phase unbalanced scenario, features are extracted from the three-phase current time series using a convolutional neural network (CNN) to generate a current transfer optimization strategy.

[0142] In the three-phase unbalanced scenario, uneven load distribution leads to significant differences in current between phases. The high-phase current needs to be transferred to the low-phase through the IGBT drive circuit to achieve three-phase balance while minimizing line losses.

[0143] State Space:

[0144] Instantaneous value of three-phase current: (time series data, sampling frequency 1 kHz);

[0145] Current imbalance: ;

[0146] Line loss power: ;

[0147] Action space: continuous action, phase current transfer (Positive value indicates transition from high phase to low phase, negative value indicates transition in the opposite direction);

[0148] Objective function:

[0149] ;

[0150] ;

[0151] in, Indicates the three-phase imbalance weight, Indicates the line loss weight.

[0152] The constraints are as follows:

[0153] Current transfer limit: ;

[0154] Three-phase unbalance safety threshold: ΔI unbalance ≤10%;

[0155] IGBT switching frequency limit: ;

[0156] The penalty function is as follows:

[0157] ;

[0158] K1=10,K2=50.

[0159] Convolutional Neural Network (CNN) design: CNN is used to extract features from the three-phase current time series and generate the optimal current transfer strategy.

[0160] (1) Input data preprocessing: Input shape is T × 3 (time step T = 100, three-phase current); the current data is normalized by phase (mean 0, variance 1);

[0161] (2) Network structure: input layer (100, 3), convolution layer Conv1D (32, kernel_size=5, activation='relu'): extract local time features, MaxPooling1D (2) downsampling to reduce the amount of calculation; fully connected layer Dense (64, activation='relu'): feature fusion, Dense (1, activation='tanh'): output current transfer amount (range [-1, 1], needs to be linearly mapped to the actual value);

[0162] (3) Loss function:

[0163] ;

[0164] MSE represents minimizing the mean square error between the predicted current transfer amount and the target value, and Penalty represents the constraint violation penalty.

[0165] Training and Optimization:

[0166] (1) Data preparation: Use power system simulation software (such as MATLAB / Simulink or OpenDSS) to simulate different load distribution scenarios, generate input three-phase current time series (T = 100), and output the optimal current transfer amount (calculated using traditional optimization algorithms or expert strategies); data enhancement, adding noise (±5% current fluctuations) and random load mutations to improve model robustness;

[0167] (2) Hyperparameter tuning: learning rate 0.001 (Adam optimizer), batch size 32, number of training rounds 200, convolution kernel size 5 (balancing local features and computational efficiency);

[0168] (3) Performance verification indicators:

[0169] Unbalance reduction rate: ;

[0170] Line loss reduction rate: .

[0171] Take a three-phase unbalanced transformer area as an example:

[0172] Initial state: three-phase current , imbalance , line loss ;

[0173] After CNN optimization: current transfer (Transfer from phase A to phase C), balanced current , imbalance , line loss (A 36% decrease).

[0174] High-precision feature extraction: CNN captures the local timing characteristics of three-phase current through convolution kernels, making it more adaptable to dynamic load changes than traditional FFT analysis. Real-time response capability: a single inference time of <1ms, meeting the 10kHz switching frequency requirement. The model can adapt to new load distribution patterns through online learning without redesigning the control logic. Multi-objective collaboration: line loss and imbalance are integrated into the loss function to achieve dual optimization of economy and power quality.

[0175] S5. If it is a harmonic pollution scenario, the harmonic compensation current generation strategy of the generative adversarial network (GAN) is adopted.

[0176] In harmonic pollution scenarios, the dense concentration of power electronic equipment (such as photovoltaic inverters and frequency converters) leads to a total harmonic distortion rate (THDi>5%), which requires the generation of a compensation current with the same magnitude and opposite phase as the harmonics ( ), to achieve harmonic filtering and minimize line losses.

[0177] State Space: Harmonic Current Spectrum (3rd, 5th, 7th and other characteristic harmonic amplitudes), satisfying opposite phases and matching amplitudes.

[0178] Objective function:

[0179] ;

[0180] in, represents the harmonic distortion weight, represents line loss weight;

[0181] .

[0182] The constraints are as follows:

[0183] Harmonic distortion safety threshold: THDi≤5%;

[0184] Compensation current bandwidth limitation: ;

[0185] Inverter capacity limitation: ;

[0186] The penalty function is as follows:

[0187] ;

[0188] in, , represents the penalty coefficient.

[0189] Generative Adversarial Network (GAN) Design:

[0190] GAN generates a compensated current spectrum through a generator, and a discriminator distinguishes between real harmonic data and generated data, gradually optimizing the generation strategy.

[0191] (1) Generator: Input random noise vector and real-time harmonic spectrum , output compensation current spectrum (N is the harmonic number).

[0192] (2) Network structure: fully connected layer (128 nodes, ReLU), fully connected layer (64 nodes, ReLU), output layer (NN nodes, tanh activation);

[0193] Output mapping, linearly mapping the tanh output (range [-1,1]) to the actual compensation current amplitude:

[0194] ;

[0195] Indicates the maximum allowable compensation current of each harmonic;

[0196] Fully connected layer (64 nodes, LeakyReLU), fully connected layer (32 nodes, LeakyReLU), output layer (1 node, sigmoid activation).

[0197] (3) Loss function:

[0198] Generator loss: (The first one deceives the discriminator, the second one penalizes constraint violations);

[0199] Discriminator loss: .

[0200] (4) Training process:

[0201] Data acquisition, collecting harmonic spectrum data from actual system or simulation;

[0202] Generator pre-training, using mean square error (MSE) to preliminarily train the generator to generate a rough compensation current;

[0203] Adversarial training, alternately training the generator and discriminator to gradually optimize the generation strategy;

[0204] Online fine-tuning, dynamically adjusts generator parameters through real-time data after deployment.

[0205] (5) Training and optimization: Use power system simulation tools (such as MATLAB / Simulink or PLECS) to simulate different harmonic scenarios and generate: input harmonic spectrum IharmonicIharmonic, output ideal compensation current IcomptargetIcomptarget (calculated through FFT analysis); data enhancement, add random harmonic perturbations (±10% amplitude variation) and noise (Gaussian white noise) to improve model robustness;

[0206] Hyperparameter settings: noise dimension 100, number of training rounds 500, batch size 32, learning rate generator (0.0001), discriminator (0.0003), penalty weight λ =10.

[0207] Performance indicators:

[0208] THDi reduction rate: ;

[0209] Line loss reduction rate: ;

[0210] Take the harmonic pollution area of ​​a certain industrial park as an example:

[0211] Initial state: ;

[0212] THDi=8.5%, line loss .

[0213] After GAN compensation: generating compensation current ;

[0214] After compensation, THDi=3.2%, line loss (A 33% decrease).

[0215] GAN uses adversarial training to generate compensation currents that closely match the actual harmonics, outperforming traditional FFT+PI control. Its online fine-tuning mechanism enables the model to adapt to changes in the harmonic spectrum (such as sudden load changes) in real time. It also compensates for characteristic harmonics such as the 3rd, 5th, and 7th harmonics, avoiding the lag inherent in successive filtering.

Claims

1. A power quality management method based on the collaboration of self-power supply and power electronic module, characterized in that: The following steps are involved: The self-powered module uses photovoltaic power generation to provide operating power to the power electronics modules of the entire system, which include the SVG, energy storage device, current transfer device and harmonic compensation device. The self-powered module powers the power electronics modules in the SVG and energy storage device, and also powers the IGBT drive circuit of the current transfer and harmonic compensation device. S1. Based on the power quality issues in the substation, the operation scenarios are divided into high voltage scenario, low voltage scenario, three-phase imbalance scenario and harmonic pollution scenario; S2. If the voltage is high, use a reinforcement learning algorithm to optimize the SVG's reactive power output strategy. S3. If it is a low voltage scenario, the proximal strategy is used to optimize the PPO energy storage scheduling optimization strategy; S4. If it is a three-phase unbalanced scenario, extract features from the three-phase current time series through the convolutional neural network (CNN) to generate a current transfer optimization strategy; S5. If it is a harmonic pollution scenario, the harmonic compensation current generation strategy of the generative adversarial network (GAN) is adopted.

2. The power quality control method according to claim 1, characterized in that: The conditions for determining the operating scenario described in S1 are as follows: High voltage scenario: V>1.1Vnominal; Low voltage scenario: V<0.9Vnominal; Three-phase unbalance scenario: three-phase current imbalance ΔI unbalance >15%; Harmonic pollution scenario: Total harmonic distortion rate THDi>5%.

3. The power quality control method according to claim 1, characterized in that: The objective function of the strategy described in S2 is as follows: min(w1·ΔV+w2P loss +w3Q SVG ); Where ΔV represents the voltage deviation, P loss Indicates line loss power, Q SVG Indicates the reactive power output of SVG. w1, w2, and w3 represent weight coefficients; The constraints are as follows: Voltage upper limit: V≤1.1Vnominal; SVG capacity limit: Q SVG max ;​ Self-powered module output: P self ≤1kW; The penalty function is as follows: Penalty=k1·max(0,V-1.1Vnominal)+k2·max(0,Q SVG -Q max ); Among them, k1 and k2 represent penalty coefficients.

4. The power quality control method according to claim 1, characterized in that: The reward function of the strategy described in S3 is as follows: R=w1·(Vnominal-0.95) 2 +w2·P loss +w3·ΔSOC+w4·Penalty; w1=-1.0, w2=-0.5, w3=-0.1, w4=-10.0; Wherein, w1 represents the voltage deviation penalty coefficient, w2 represents the line loss power penalty coefficient, w3 represents the energy storage SOC change ΔSOC penalty coefficient, w4 represents the constraint violation penalty coefficient, Penalty represents the penalty function, and ΔSOC represents the energy storage SOC change; The constraints are as follows: Voltage lower limit: V ≥ 0.9Vnominal; Energy storage charging and discharging power limit: |P ESS |≤P ESS,max ; Energy storage SOC safety range: SOC ≥ 20%: SVG capacity limit: Q SVG max .​ 5. The power quality control method according to claim 1, characterized in that: The S3 includes: optimizing the PPO energy storage scheduling optimization strategy through the proximal strategy and interacting with the environment to collect trajectory data; estimating the advantage value using GAE; limiting the strategy update amplitude through the clipping mechanism to maximize the objective function; and minimizing the mean square error loss of the value network.

6. The power quality control method according to claim 1, characterized in that: The objective function of the strategy described in S4 is as follows: min(w1·ΔI unbalance +w2P loss ); w1=1.0, w2=0.5; Among them, w1 represents the three-phase imbalance weight, w2 represents the line loss weight; The constraints are as follows: Current transfer limit: |ΔI transfer |≤I max ; Three-phase unbalance safety threshold: ΔI unbalance ≤10%; IGBT switching frequency limit: f sw ≤10kHz; The penalty function is as follows: Penalty=k1·max(0,ΔI unbalance -10%)+k2·max(0,|ΔI transfer |-I max ); Among them, k1 and k2 represent penalty coefficients.

7. The power quality control method according to claim 1, characterized in that: The S4 includes: normalizing the current data by phase; simulating different load distribution scenarios through power system simulation software and inputting three-phase current time series; outputting the optimal current transfer amount through traditional optimization algorithms or expert strategies; adding noise and random load mutations; and integrating the loss function of MSE and constraint penalty, where MSE represents minimizing the mean square error between the predicted current transfer amount and the target value.

8. The power quality control method according to claim 1, characterized in that: The objective function of the strategy described in S5 is as follows: min(w1 THDi+w2 P loss , Among them, w1 represents the harmonic distortion weight, w2 represents the line loss weight; The constraints are as follows: Harmonic distortion safety threshold: THD i ≤5%; Compensation current bandwidth limit: f comp ≤2kHz; Inverter capacity limit: |I comp,h |≤I inv,max ; The penalty function is as follows: Penalty=k1·max(0,THDi-5%)+k2·max(0,|I comp,h |-I inv,max ); Among them, k1 and k2 represent penalty coefficients.

9. The power quality control method according to claim 1, characterized in that: The S5 includes: collecting harmonic spectrum data from the actual system or simulation and inputting it into the power system simulation tool; using the mean square error (MSE) to preliminarily train the generator to generate an ideal compensation current; alternately training the generator and the discriminator to gradually optimize the generation strategy; and dynamically adjusting the generator parameters through real-time data.

10. A power quality management system based on the collaboration of self-power supply and power electronic module, characterized in that: include: Self-powered module, used to power the power electronic modules of the entire system through photovoltaic power generation, and also to power the power electronic modules in the SVG and energy storage equipment, and the IGBT drive circuits of the current transfer and harmonic compensation equipment; The power electronics module includes an SVG, energy storage equipment, current transfer equipment, and harmonic compensation equipment. It is used to provide different response strategies according to different operating scenarios. The strategies include a reactive power output strategy optimized by a reinforcement learning algorithm for SVG, an energy storage scheduling optimization strategy optimized by a proximal policy optimization (PPO), a current transfer optimization strategy, and a harmonic compensation current generation strategy using a generative adversarial network (GAN).

Images