A machine learning-based solid-state transformer grid-connected island smooth switching control method

Abstract

The application belongs to the technical field of power electronic conversion and microgrid control. A kind of grid-connected island smooth switching control method of solid state transformer based on machine learning, it is characterized by comprising the following steps: S1: multi-dimensional electrical quantity acquisition and preprocessing: real-time acquisition of multi-dimensional electrical quantity of each link of solid state transformer and power grid, load side by voltage, current sensor;S2: accurate identification of mode based on CNN-LSTM-Attention network: input the time series feature matrix constructed in step S1 into the pre-trained CNN-LSTM-Attention multi-modal time series network;S3: weight self-adaptive unified control based on SAC deep reinforcement learning;S4: three-level topology coordinated control based on TCN time series prediction;S5: island to grid smooth switching of embedded pre-synchronization. Realize zero impact, high reliability, full working condition adaptive operation of solid state transformer bidirectional switching of double mode, meet the real-time requirement of power electronic control.

Description

Technical Field

[0001] This invention belongs to the field of power electronic conversion and microgrid control technology, specifically relating to the operation mode control technology of solid-state transformers (also known as power electronic transformers), and particularly to a control method that integrates deep learning and reinforcement learning algorithms to achieve seamless and smooth switching between grid-connected operation and islanded operation modes of solid-state transformers. Background Technology

[0002] In the process of transforming new power systems towards distributed, flexible, and intelligent operation, solid-state transformers (SSTs), with their flexible voltage transformation, reliable electrical isolation, precise power flow control, and integrated power quality management capabilities, have become core and critical equipment for distributed renewable energy consumption, integrated photovoltaic-storage-charging power station construction, and AC / DC hybrid microgrid operation. To ensure the continuous and stable operation of the microgrid system, SSTs must be capable of both grid-connected and islanded operation modes, and the switching between these modes must be seamless and smooth. This is crucial to prevent issues such as large voltage / frequency fluctuations and current overshoot during switching, which could lead to load shutdowns, damage to power electronic devices, or even the collapse of the entire microgrid system.

[0003] Currently, the mainstream solid-state transformer mode switching control strategy generally adopts a hard switching method of PQ constant power control and VF constant voltage and frequency control. Its core drawback is that it relies on mode switching to switch between two independent control loops. There are natural control breakpoints during the switching process, which inevitably generate voltage surges and frequency fluctuations, making it difficult to meet the requirements of high precision and high reliability operation.

[0004] To improve the handover effect, some existing technologies have proposed a unified control architecture for weighted transition, but there are still many problems that need to be solved: First, pattern recognition relies on threshold judgment of single parameters such as voltage amplitude and frequency at the PCC point, which has weak anti-interference ability and is easily affected by factors such as grid harmonics, load changes, and voltage sags and dips, resulting in serious misjudgments and failure to operate, and it cannot predict grid faults in advance, thus missing the optimal handover time. Second, the weighted transition curve relies on manually preset linear or exponential rules, which cannot adapt to the dynamic changes of different load types (resistive, inductive, nonlinear, unbalanced loads) and different grid operating conditions, making it difficult to achieve handover under all operating conditions. The first problem is that the switching effect is not optimal globally. The second problem is that most solutions only focus on the control optimization of the output stage inverter of the solid-state transformer, completely ignoring the coordinated control of its three-level topology (input stage rectifier unit, isolation stage bidirectional DAB conversion unit, and output stage inverter unit). The DC bus voltage fluctuation caused by power change during the switching process will seriously deteriorate the switching effect and even trigger the system protection shutdown. The third problem is that the control robustness is insufficient under complex operating conditions. In nonlinear and unbalanced load scenarios, traditional PI / PR controllers are difficult to guarantee the output power quality during the switching process. Moreover, the pre-synchronization link of islanding to grid connection needs to rely on an additional control loop, which is prone to secondary switching impact.

[0005] In recent years, the application of machine learning technology in the field of power electronic control has gradually increased. Some studies have attempted to use it for inverter pattern recognition, but they have only reached the level of single grid fault classification and have not achieved deep integration of machine learning with the full-link control of solid-state transformer mode switching. At the same time, existing machine learning solutions have not fully considered the microsecond-level real-time requirements and high reliability constraints of power electronic control. Most of them have only completed simulation verification and lack the feasibility of engineering deployment, thus failing to meet the needs of practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a machine learning-based control method for smooth switching between grid-connected and islanded solid-state transformers, which enables zero-impact, high-reliability, and adaptive operation of solid-state transformers in both modes and conditions, thus meeting the real-time requirements of power electronic control.

[0007] To achieve the above-mentioned objectives, the present invention employs the following technical solution: a machine learning-based control method for smooth switching between grid-connected and islanded modes of a solid-state transformer. The solid-state transformer adopts a three-stage topology, consisting of an input-stage PWM rectifier unit, an isolation-stage bidirectional DAB DC-DC converter unit, and an output-stage LC filter and inverter unit connected sequentially. The control method achieves seamless and smooth switching between grid-connected and islanded modes through multi-dimensional sensing, precise pattern recognition, adaptive weight control, end-to-end cooperative disturbance suppression, and embedded pre-synchronization. The method is characterized by the following steps: S1: Multi-dimensional electrical quantity acquisition and preprocessing: The solid-state transformer and the power grid and load side collect multi-dimensional electrical quantities in real time through voltage and current sensors. These include the instantaneous values ​​of three-phase voltage and three-phase current at the PCC point (point of common coupling) on ​​the power grid side, the DC bus voltage and current at the input stage, the DC bus voltage and current at the high and low voltage sides of the isolation stage, the three-phase output voltage and three-phase output current at the output stage, and the active and reactive power at the load side. The collected electrical quantities are filtered by a sliding window to eliminate high-frequency noise interference, and then standardized to unify the dimensions. Finally, a time-series feature matrix with a length of 20 control cycles (the control cycle is set to 20μs) is constructed to provide high-quality input data for subsequent pattern recognition and disturbance prediction. S2: Accurate pattern recognition based on CNN-LSTM-Attention network: The temporal feature matrix constructed in step S1 is input into the pre-trained CNN-LSTM-Attention (convolutional neural network-long short-term memory network-attention mechanism) multimodal temporal network to achieve accurate identification of power grid status and operation mode, and output three types of operation mode instructions: grid-connected mode, islanded mode, and switching transition mode. The specific structure and deployment of the CNN-LSTM-Attention multimodal temporal network are as follows: The input layer receives 12-dimensional temporal features, which cover the voltage amplitude, frequency, phase, total harmonic distortion (THD), online identification value of grid impedance, DC bus voltage fluctuation, load active / reactive power change rate, output voltage imbalance, and load nonlinearity at the PCC point. The feature extraction layer consists of two 1D CNN convolutional layers with a kernel size of 3 and ReLU activation function, used to extract local waveform and spectral features of the temporal data. The temporal dependency layer consists of two LSTM long short-term memory networks with 64 hidden neurons, used to mine the long-short-term dependencies of the temporal data and effectively distinguish the temporal differences between grid faults and load mutations. The attention weighting layer uses the Bahdanau attention mechanism, which can automatically assign higher weights to key fault features and suppress the influence of non-key interference features. The output layer uses the Softmax activation function to output the probability values ​​of three operating modes, and takes the mode corresponding to the highest probability value as the final operating mode instruction. The network adopts an "offline pre-training + online lightweight inference" deployment approach: In the offline stage, a dataset (200,000 samples) covering 12 operating conditions such as symmetrical / asymmetrical power grid faults, voltage swell / droop, load mutation, and power grid harmonics is constructed in the MATLAB / Simulink simulation environment. The Adam optimizer and cross-entropy loss function are used to complete the model training. After training, the model is deployed on the FPGA control chip through INT8 quantization and model pruning optimization. The single-step inference latency is controlled within 8μs. At the same time, a hysteresis threshold and a 200μs anti-false judgment delay are set to avoid frequent mode switching. S3: Weight adaptive unified control based on SAC (Soft Actor-Critic) deep reinforcement learning: A unified control architecture is constructed by connecting the PQ power control loop and the VF voltage-frequency control loop in parallel. The outputs of the two control loops are weighted and summed using weighted coefficients, and then used as the reference value for the inner current loop. In grid-connected mode, the weight of the PQ control loop is α=1, the weight of the VF control loop is β=0, and α+β=1, focusing on active and reactive power tracking. In islanded mode, α=0 and β=1, focusing on maintaining constant output voltage and frequency. In the transition mode, a pre-trained SAC soft actor-commentator deep reinforcement learning agent continuously and smoothly adjusts the values ​​of α and β to achieve uninterrupted unified control. The specific design of the SAC deep reinforcement learning agent is as follows: The state space includes the current operating mode command, PQ / VF weight coefficients α / β, output voltage amplitude error, output frequency error, output current overshoot, DC bus voltage fluctuation, load nonlinearity, and synchronization phase error during islanding to grid connection; the action space is the value of the continuously adjusted PQ control loop weight α, with an action range of [0,1], a step size of 0.01, and β=1-α synchronous updates; the reward function focuses on minimizing switching shock and transition time, and the specific formula is: In the formula, Reward is the reward function. The relative error between the output voltage amplitude and the rated value. The relative error between the output frequency and the rated value. This refers to the relative overshoot of the output current compared to its rated value. For switching transition time; , , , These are weighting coefficients, with values ​​of 0.4, 0.3, 0.2, and 0.1 respectively. Training and deployment of the agent: In the offline stage, pre-training was completed in the Simulink / PLECS (Power Electronics Circuit Simulation) co-simulation environment, covering all operating conditions. A dual critic network and target network update mechanism were adopted, with an experience replay pool capacity of 100,000 and 500 training rounds. During online operation, an action limit constraint of ≤0.1 for the weight change rate within a single cycle was set. At the same time, an emergency fallback mechanism was added - when the output voltage fluctuation exceeds ±5% of the rated value, it automatically switches to the traditional PI control with preset linear weight transition to ensure the reliability of system operation. S4: Three-level topology cooperative control based on TCN (Temporal Convolutional Network) temporal prediction: During mode switching, the pre-trained TCN temporal convolutional network predicts DC bus disturbances and performs coordinated control on the input stage, isolation stage, and output stage simultaneously, providing stable DC support for output stage switching and effectively solving the problem of the deterioration of switching effect caused by the disturbances in the preceding stage. The specific design of the TCN temporal convolutional network is as follows: the input consists of the load power, output stage current, high and low voltage DC bus voltages, and DAB phase shift angle timing data of the past 20 control cycles; the output consists of the load power mutation value, high voltage DC bus voltage fluctuation prediction value, and low voltage DC bus voltage fluctuation prediction value for the next 5 control cycles. The model adopts 4 causal convolutional layers with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, and the activation function is ReLU. The mean squared error (MSE) is used as the loss function. After offline training, the network is quantized and deployed, with a single-step inference latency of ≤5μs. The specific implementation of the three-level topology collaborative control is as follows: The grid-side disturbance and DC bus voltage fluctuation values ​​predicted by the TCN are used as feedforward compensation values ​​and superimposed on the voltage loop output of the input stage PWM rectifier unit to adjust the rectifier modulation wave in advance and suppress the high-voltage DC bus fluctuations caused by grid-side disturbances and power surges; The load power surge values ​​predicted by the TCN are used as feedforward compensation values ​​to adjust the phase shift angle of the DAB unit in advance, quickly match the input-output power difference, and stabilize the low-voltage DC bus voltage; After the mode switching command is issued, the input stage and isolation stage enter the disturbance suppression mode one control cycle in advance, and are synchronized with the mode switching action of the output stage to achieve full-link collaboration; S5: Embedded pre-synchronized smooth handover from islanded to grid-connected systems: When the CNN-LSTM-Attention (Convolutional Neural Network-Long Short-Term Memory Network-Attention Mechanism) network recognizes that the power grid has returned to normal, a pre-synchronization tracking loop is embedded in the VF control loop of the unified control architecture, eliminating the need to switch to an additional pre-synchronization control loop and avoiding secondary switching impact. The SAC agent progressively adjusts the voltage amplitude, frequency, and phase reference values ​​of the VF control loop to ensure that the synchronization error between the output stage voltage and the grid-side PCC point voltage meets the preset thresholds (amplitude error ≤1%, frequency error ≤0.1Hz, phase error ≤1°). After synchronization is completed, the SAC agent automatically and smoothly increases the weight α of the PQ control loop and decreases the weight β of the VF control loop, smoothly completing the zero-impact transition from islanded mode to grid-connected mode.

[0008] Compared with the prior art, the present invention has the following significant advantages: 1. High pattern recognition accuracy and strong anti-interference capability: By using the CNN-LSTM-Attention multimodal temporal network to integrate multi-dimensional temporal features, the system can accurately identify the power grid status and operating mode. The pattern recognition accuracy reaches 99.9%, and the false judgment / failure rate is reduced by two orders of magnitude compared with traditional methods. It can predict power grid faults 1-2 control cycles in advance, obtain the optimal timing for mode switching, and effectively resist interference from power grid harmonics, load changes, etc.

[0009] 2. Seamless and shock-free switching process: Based on SAC deep reinforcement learning, the PQ / VF weights are adaptively and smoothly adjusted under all operating conditions, eliminating the need for manually preset transition curves and effectively eliminating control breakpoints. During the bidirectional switching between grid-connected and islanded systems, the output voltage fluctuation is ≤±1%, the frequency fluctuation is ≤±0.1Hz, and the switching transition time is ≤2ms, which is far superior to the national standard requirements, achieving truly seamless switching.

[0010] 3. Strong end-to-end coordination and high stability: The TCN timing prediction model realizes end-to-end coordinated control of the three-level topology of solid-state transformers, suppressing DC bus voltage fluctuations within ±0.5% during mode switching, effectively solving the problem of the deterioration of the output stage switching effect by the front-stage disturbance, and greatly improving the system's operational stability.

[0011] 4. Excellent adaptability to operating conditions and power quality: Embedded pre-synchronization is achieved within a unified control architecture to avoid secondary impacts during islanded to grid-connected transitions; Under complex operating conditions such as nonlinear and unbalanced loads, the total harmonic distortion (THD) of the output voltage is ≤2% and the voltage imbalance is ≤1% during the switching process, fully meeting the national standards for power quality.

[0012] 5. High engineering feasibility: All machine learning models adopt the deployment method of "offline pre-training + online lightweight inference", and the single-step inference latency is controlled within 10μs, which fully meets the microsecond-level real-time requirements of power electronic control. At the same time, a complete emergency backup mechanism is set up to ensure the high reliability of the system. It can be directly deployed on FPGA control chips and has good prospects for engineering applications. Attached Figure Description

[0013] Figure 1 This is a three-level topology diagram of the solid-state transformer described in this invention.

[0014] Figure 2 This is a block diagram of the overall architecture of the control strategy described in this invention.

[0015] Figure 3 This is a schematic diagram of the CNN-LSTM-Attention pattern recognition model described in this invention.

[0016] Figure 4 This is a schematic diagram of the SAC deep reinforcement learning weight adaptive unified control described in this invention.

[0017] Figure 5 This is a block diagram of the TCN timing prediction and three-level topology cooperative control described in this invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to specific embodiments, so as to more clearly demonstrate the implementation process and technical advantages of the present invention.

[0019] Example Parameter Settings: The rated parameters of the solid-state transformer described in this embodiment are as follows: rated capacity 100kVA, rated line voltage on the grid side 380V / 50Hz, rated voltage of the input stage high-voltage DC bus 800V, rated voltage of the isolation stage low-voltage DC bus 400V, rated line voltage on the output side 380V / 50Hz, switching frequency 10kHz, control cycle 20μs. An FPGA is used as the core control chip to implement all control algorithms and machine learning models. The control cycle is set to 20μs, much smaller than the power device switching cycle of 100μs (corresponding to a 10kHz switching frequency), aiming to achieve a higher precision control algorithm execution frequency, thereby responding more quickly to system state changes and updating control commands.

[0020] A machine learning-based control method for smooth grid-connected islanding switching of a solid-state transformer, wherein the solid-state transformer adopts a three-stage topology, consisting of an input-stage PWM rectifier unit and an isolation-stage bidirectional DAB, specifically including the following steps: Step 1, Multi-dimensional Electrical Quantity Acquisition and Preprocessing: The following electrical quantities are collected in real time using high-precision voltage and current sensors: three-phase voltage at the PCC point on the grid side (u a u b u c ), three-phase current (i a i b i c ), Input stage DC bus voltage (U dc1 ), current (I) dc1 ), isolation stage high-voltage side DC bus voltage (U dc1 ), low-voltage side DC bus voltage (U dc2 ), corresponding current (I) dc1 I dc2 ), output stage three-phase output voltage (u oa u ob u oc ), three-phase output current (i oa i ob i oc ), active power on the load side (P) L ), reactive power (Q) L ).

[0021] The collected electrical quantities are subjected to sliding window filtering (window length is 5 control cycles) to eliminate high-frequency noise interference. Then, per-unit processing is performed to convert all electrical quantities into dimensionless parameters, and a 12-dimensional time-series feature matrix with a length of 20 control cycles is constructed. The features specifically include: PCC point voltage amplitude, frequency, phase, THD, online identification value of grid impedance, high-voltage DC bus voltage fluctuation, load active power change rate, load reactive power change rate, output voltage imbalance, load nonlinearity, low-voltage DC bus voltage fluctuation, and output current peak value.

[0022] Step 2, Pattern recognition based on CNN-LSTM-Attention network: The preprocessed 12-dimensional temporal feature matrix (size (20,12)) is input into the pre-trained CNN-LSTM-Attention network, the specific structure of which is as follows: 1. Input layer: Receives a temporal feature matrix of size (20, 12); 2. First CNN convolutional layer: 32 convolutional kernels, size 3, stride 1, padding method is Same, ReLU activation function; 3. Second CNN convolutional layer: 64 convolutional kernels, size 3, stride 1, padding method Same, ReLU activation function; 4. First LSTM layer: 64 hidden neurons, returns a complete time sequence; 5. Second LSTM layer: 64 hidden neurons, returns the final temporal output; 6. Bahdanau Attention Layer: Weights the feature vector output by the LSTM to focus on key features; 7. Fully connected layer: 32 neurons, ReLU activation function; 8. Output layer: 3 neurons, with Softmax activation function, corresponding to the probability values ​​of grid-connected mode, island mode, and transition mode respectively.

[0023] The mode corresponding to the maximum output probability is taken as the final operating mode command. At the same time, a hysteresis threshold and a delay to prevent false judgment are set: when the grid voltage amplitude drops to below 85% of the rated value and the duration exceeds 200μs, it is judged as a grid fault and an islanding mode command is output; when the grid voltage recovers to the range of 90%-110% of the rated value and the duration exceeds 500μs, it is judged as the grid has returned to normal and a grid-connected mode command is output.

[0024] The specific parameters for offline training of this network are as follows: a dataset of 200,000 samples covering 12 types of working conditions is constructed, the Adam optimizer is used, the initial learning rate is 0.001, the batch size is 64, the training epochs are 100, and the cross-entropy loss function is used. After training, NT8 quantization, structured pruning, and a highly optimized FPGA pipeline design are used to compress the number of model parameters to 30% of the original. After deployment on the FPGA, the measured single-step inference latency is 7.2μs, which fully meets the 20μs control cycle requirement.

[0025] Step 3, Adaptive unified control of weights based on SAC deep reinforcement learning: A unified control architecture is constructed by connecting the PQ power control loop and the VF voltage-frequency control loop in parallel. The PQ control loop uses a PI controller to achieve accurate tracking of active power P and reactive power Q in grid-connected mode. The VF control loop also uses a PI controller to achieve constant control of the output voltage amplitude and frequency in islanded mode. The outputs of the two outer loops are weighted and summed by weighting coefficients α and β, and used as reference values ​​for the d-axis and q-axis of the inner current loop, always satisfying α+β=1.

[0026] The current operating mode command, weight coefficient α, output voltage amplitude error, output frequency error, output current overshoot, DC bus voltage fluctuation, load nonlinearity, and phase synchronization error (when switching from islanding to grid connection) are input into the pre-trained SAC agent. The agent outputs the weight coefficient α of the current control cycle, and β is synchronously updated to 1-α to achieve continuous and smooth adjustment of the weight.

[0027] Offline training parameters for the SAC agent: In the Simulink / PLECS co-simulation environment, it covers a full range of operating conditions including resistive, inductive, nonlinear, and unbalanced loads, as well as different grid fault types. It adopts a dual-commentator network structure, with an experience replay pool capacity of 100,000, a soft update coefficient of 0.005, a discount factor of 0.99, and 500 training rounds. During online operation, a limit constraint of ≤0.1 for the single-cycle weight change rate is set, and an emergency fallback mechanism is set: when the output voltage fluctuation exceeds ±5% of the rated value, the SAC agent output is immediately frozen, and it switches to traditional PI control with a preset 10ms linear weight transition to ensure that the system does not run out of control.

[0028] Step 4, Three-level topology cooperative control based on TCN timing prediction: The load active power (P) over the past 20 control cycles L ), output stage d / q axis current (i od i oq ), High voltage DC bus voltage (U dc1 ), low-voltage DC bus voltage (U dc2 The DAB phase shift angle (θ) time series data is input into the pre-trained TCN time series prediction model, which outputs the load power mutation value (ΔPL) and the high-voltage DC bus voltage fluctuation prediction value (ΔU) for the next 5 control cycles. dc1 ), the predicted value of low-voltage DC bus voltage fluctuation (ΔUdc2).

[0029] The specific parameters of the TCN model are as follows: 4 causal convolutional layers with dilation coefficients of 1, 2, 4 and 8 respectively, kernel size of 3, 64 output channels per layer, ReLU activation function, dropout rate of 0.2, and mean squared error (MSE) as the loss function; offline training is also deployed through INT8 quantization and highly pipelined optimization, and the measured single-step inference latency is 4.6μs.

[0030] During mode switching, the specific execution of coordinated control includes: predicting ΔU... dc1 As a feedforward compensation, it is superimposed on the voltage loop PI controller output of the input-stage PWM rectifier unit to pre-adjust the modulation waveform of the rectifier and suppress high-voltage DC bus fluctuations caused by grid-side disturbances; the predicted ΔP L As a feedforward compensation, the phase shift angle θ of the DAB unit is adjusted in advance to quickly match the input-output power difference and stabilize the low-voltage DC bus voltage U. dc2 After the mode switching command is issued, the input stage and the isolation stage enter the disturbance suppression mode one control cycle in advance, and are synchronized with the weight adjustment action of the output stage to achieve full-link coordinated control of the input stage, isolation stage and output stage.

[0031] Step 5, Embedded pre-synchronized island-to-grid handover: Once the CNN-LSTM-Attention network recognizes that the power grid has returned to normal, a pre-synchronization tracking loop is embedded in the VF control loop of the unified control architecture, eliminating the need to switch to an additional pre-synchronization control loop and simplifying the control logic. Under the guidance of the reward function, the SAC agent progressively adjusts the voltage amplitude, frequency, and phase reference value of the VF control loop, tracking the changes in the grid voltage at the PCC point in real time, until the amplitude error between the output stage voltage and the grid voltage is ≤1%, the frequency error is ≤0.1Hz, and the phase error is ≤1°, at which point the pre-synchronization is considered complete.

[0032] After pre-synchronization is completed, the SAC agent smoothly increases the weight coefficient α linearly from 0 to 1 and decreases β linearly from 1 to 0 within 10ms, completing the zero-impact smooth transition from island mode to grid-connected mode. The entire switching process is free of current overshoot and voltage fluctuation.

[0033] Example verification results: This embodiment verifies the control effect of the present invention through MATLAB / Simulink simulation and FPGA hardware-in-the-loop experiment. The specific results are as follows: 1. Grid-connected to islanded switching: When a three-phase short-circuit fault occurs in the power grid, the control strategy of this invention can complete the switching within 1.6ms. During the switching process, the output voltage fluctuation is ≤0.8%, the frequency fluctuation is ≤0.08Hz, there is no current overshoot, and the load operation is stable. 2. Island to grid connection switching: After the grid returns to normal, the pre-synchronization completion time is ≤8ms, the output voltage fluctuation is ≤0.6% and the frequency fluctuation is ≤0.05Hz during the switching process, with no secondary impact and smooth transition to grid connection mode; 3. DC bus stability: During mode switching, the voltage fluctuation of the high-voltage DC bus is ≤0.4%, and the voltage fluctuation of the low-voltage DC bus is ≤0.3%, with no obvious oscillation, providing stable DC support for output stage switching; 4. Adaptability to complex operating conditions: When carrying a three-phase uncontrolled rectifier nonlinear load (nonlinearity 60%), under the synergistic effect of the control strategy, the output voltage THD during the switching process is ≤1.8%; when carrying a 25% unbalanced load, the output voltage imbalance is ≤0.8%, and the power quality fully meets the requirements of GB / T 14549-1993 "Power Quality - Harmonics in Public Power Grids" and GB / T 15543-2008 "Power Quality - Three-Phase Voltage Imbalance".

Claims

1. A machine learning-based control method for smooth grid-connected islanding switching of solid-state transformers, characterized in that... Includes the following steps: S1: Multi-dimensional electrical quantity acquisition and preprocessing: The solid-state transformer and the power grid and load side are collected in real time by voltage and current sensors, including the instantaneous values ​​of three-phase voltage and three-phase current at the PCC point on the power grid side, the DC bus voltage and current of the input stage, the DC bus voltage and current of the isolation stage high and low voltage sides, the three-phase output voltage and three-phase output current of the output stage, and the active power and reactive power on the load side. The collected electrical quantities are filtered by sliding window to eliminate high-frequency noise interference, and then standardized to unify the dimensions. Finally, a time-series feature matrix with a length of 20 control cycles is constructed, with the control cycle set to 20μs, to provide high-quality input data for subsequent pattern recognition and disturbance prediction. S2: Accurate pattern recognition based on CNN-LSTM-Attention network: The temporal feature matrix constructed in step S1 is input into the pre-trained CNN-LSTM-Attention multimodal temporal network to achieve accurate identification of power grid status and operation mode, and output three types of operation mode instructions: grid-connected mode, islanded mode, and switching transition mode. S3: Weight Adaptive Unified Control Based on SAC Deep Reinforcement Learning: A unified control architecture is constructed by connecting the PQ power control loop and the VF voltage-frequency control loop in parallel. The outputs of the two control loops are weighted and summed using weighted coefficients, and then used as the reference value for the inner current loop. In grid-connected mode, the weight of the PQ control loop is α=1, the weight of the VF control loop is β=0, and α+β=1, focusing on active and reactive power tracking. In islanded mode, α=0 and β=1, focusing on maintaining constant output voltage and frequency. In the transition mode, a pre-trained SAC soft actor-commentator deep reinforcement learning agent continuously and smoothly adjusts the values ​​of α and β to achieve uninterrupted unified control. S4: Three-level topology cooperative control based on TCN timing prediction: During mode switching, the pre-trained TCN temporal convolutional network predicts DC bus disturbances and simultaneously performs coordinated control on the input stage, isolation stage, and output stage, providing stable DC support for output stage switching. S5: Embedded pre-synchronized smooth handover from islanded to grid-connected systems: When the CNN-LSTM-Attention network recognizes that the power grid has returned to normal, a pre-synchronization tracking loop is embedded in the VF control loop of the unified control architecture, eliminating the need to switch to an additional pre-synchronization control loop and avoiding the impact of secondary switching. The voltage amplitude, frequency, and phase reference value of the VF control loop are progressively adjusted by the SAC agent so that the synchronization error between the output stage voltage and the PCC point voltage on the power grid side meets the preset threshold. After synchronization is completed, the SAC agent automatically and smoothly increases the weight α of the PQ control loop and decreases the weight β of the VF control loop, thus smoothly completing the zero-impact transition from islanded mode to grid-connected mode.

2. The machine learning-based solid-state transformer grid-connected islanding smooth switching control method according to claim 1, characterized in that... In step S2, the specific structure and deployment of the CNN-LSTM-Attention multimodal temporal network are as follows: The input layer receives 12-dimensional temporal features, which cover the voltage amplitude, frequency, phase, total harmonic distortion rate, online identification value of grid impedance, DC bus voltage fluctuation, load active / reactive power change rate, output voltage imbalance, and load nonlinearity at the PCC point; the feature extraction layer consists of two one-dimensional CNN convolutional layers with a kernel size of 3 and ReLU activation function, used to extract local waveform features and spectral features of the temporal data; The temporal dependency layer consists of two layers of LSTM (Long Short-Term Memory) network with 64 hidden neurons. It is used to mine the long-short-term dependencies of temporal data and effectively distinguish the temporal differences between power grid faults and load mutations. The attention weighting layer adopts the Bahdanau attention mechanism, which can automatically assign higher weights to key fault features and suppress the influence of non-key interference features. The output layer adopts the Softmax activation function, outputs the probability values ​​of three types of operating modes, and takes the mode corresponding to the highest probability value as the final operating mode instruction. The network adopts an "offline pre-training + online lightweight inference" deployment approach: In the offline stage, a dataset covering 12 operating conditions such as symmetrical / asymmetrical power grid faults, voltage swell / droop, load mutation, and power grid harmonics is constructed in the MATLAB / Simulink simulation environment, with a total of 200,000 samples. The Adam optimizer and cross-entropy loss function are used to complete the model training. After training, the model is deployed on the FPGA control chip through INT8 quantization and model pruning optimization, with the single-step inference latency controlled within 8μs. At the same time, a hysteresis threshold and a 200μs anti-false judgment delay are set to avoid frequent mode switching.

3. The machine learning-based solid-state transformer grid-connected islanding smooth switching control method according to claim 1, characterized in that... In step S3, the specific design of the SAC deep reinforcement learning agent is as follows: the state space includes the current operating mode instruction, PQ / VF weight coefficients α / β, output voltage amplitude error, output frequency error, output current overshoot, DC bus voltage fluctuation, load nonlinearity, and synchronization phase error during islanding to grid connection; the action space is the value of continuously adjusting the weight α of the PQ control loop, with an action range of [0,1] and a step size of 0.01, and β=1-α is updated synchronously; the reward function focuses on minimizing the switching impact and transition time, and the specific formula is: In the formula, Reward is the reward function. This refers to the relative error between the output voltage amplitude and the rated value. The relative error between the output frequency and the rated value. This refers to the relative overshoot of the output current compared to its rated value. For switching transition time; , , , These are weighting coefficients, with values ​​of 0.4, 0.3, 0.2, and 0.1 respectively. Training and deployment of the agent: In the offline stage, pre-training was completed in the Simulink / PLECS co-simulation environment, covering all working conditions. A dual critic network and target network update mechanism was adopted, with an experience replay pool capacity of 100,000 and 500 training rounds. During online operation, an action limit constraint of ≤0.1 for the weight change rate within a single cycle was set. At the same time, an emergency fallback mechanism was added - when the output voltage fluctuation exceeds ±5% of the rated value, it automatically switches to the traditional PI control with preset linear weight transition to ensure the reliability of system operation.

4. The machine learning-based solid-state transformer grid-connected islanding smooth switching control method according to claim 1, characterized in that... In step S4, the specific design of the TCN temporal convolutional network is as follows: the input consists of the load power, output stage current, high and low voltage DC bus voltages, and DAB phase shift angle timing data of the past 20 control cycles; the output consists of the load power mutation value, high voltage DC bus voltage fluctuation prediction value, and low voltage DC bus voltage fluctuation prediction value for the next 5 control cycles; the model adopts 4 causal convolutional layers with dilation coefficients of 1, 2, 4, and 8, a convolutional kernel size of 3, the activation function is ReLU, the mean squared error is used as the loss function, and quantization deployment is performed after offline training, with a single-step inference latency ≤5μs; The specific implementation of the three-level topology collaborative control is as follows: The grid-side disturbance and DC bus voltage fluctuation values ​​predicted by the TCN are used as feedforward compensation values ​​and superimposed on the voltage loop output of the input stage PWM rectifier unit to adjust the rectifier modulation wave in advance and suppress the high-voltage DC bus fluctuations caused by grid-side disturbances and power surges; The load power surge values ​​predicted by the TCN are used as feedforward compensation values ​​to adjust the phase shift angle of the DAB unit in advance, quickly match the input-output power difference, and stabilize the low-voltage DC bus voltage; After the mode switching command is issued, the input stage and isolation stage enter the disturbance suppression mode one control cycle in advance, and are synchronized with the mode switching action of the output stage to achieve full-link collaboration.

5. The machine learning-based solid-state transformer grid-connected islanding smooth switching control method according to claim 1, characterized in that... In step S5, the synchronization error between the output stage voltage and the grid-side PCC point voltage is made to meet the preset thresholds: amplitude error ≤1%, frequency error ≤0.1Hz, and phase error ≤1°.

Images