A High- and Low-Voltage Ride-Through Control Method for Energy Storage Converters Based on Predictive Large Models

Abstract

This invention pertains to the field of power supply and distribution, and particularly relates to a high- and low-voltage ride-through control method for energy storage converters based on a large-scale predictive model. This method constructs a dynamic predictive model to identify grid voltage trends in real time, and combines this with a multi-objective optimization algorithm to calculate the support duration for the current fault ride-through. This method effectively expands the range of the converter's fault ride-through curve, allowing it to dynamically extend the support duration based on actual grid needs even after the standard ride-through time has expired. This solves the problem of insufficient reactive power support leading to decreased system stability under grid faults outside the standard curve, enabling the converter to shift from passive compliant grid disconnection to proactive extended support. This invention utilizes a large-scale predictive model to anticipate voltage recovery trends in advance, dynamically adjusting the ride-through strategy while ensuring equipment safety. This avoids the thermal stress risks associated with blindly extending the support duration and enhances grid resilience.

Description

Technical Field

[0001] This invention belongs to the field of power supply and distribution, and particularly relates to a high and low voltage ride-through control method for energy storage converters based on a large predictive model. Background Technology

[0002] According to the requirements of GBT34120-2023 Technical Specification for Energy Storage Converters in Electrochemical Energy Storage Systems, A1 and A2 type energy storage converters must meet the corresponding high and low voltage ride-through requirements. At the same time, the converters must be able to meet at least two consecutive low-to-high voltage ride-throughs without disconnecting from the grid within the curve range, and be able to inject the corresponding reactive current into the grid in accordance with the standard requirements during the ride-through process.

[0003] The high and low voltage ride-through requirements described in GB / T 34120-2023, "Technical Specification for Energy Storage Converters in Electrochemical Energy Storage Systems," represent the minimum requirements for converters. However, converters themselves possess a longer ride-through capability. Currently, converters perform fault ride-through according to the time specified in the standard, actively disconnecting from the grid before the grid fault is resolved, even after reaching the curve limit time. While this logic satisfies the fault ride-through requirements in the standard, it fails to provide reactive current support for fault times outside the standard ride-through curve. This results in insufficient reactive power support for voltage faults outside the standard curve, potentially leading to the expansion of grid faults. Especially during short-circuit faults or sudden load changes in the grid, voltage fluctuations are frequent and large in amplitude, making it difficult for traditional control strategies to respond quickly and maintain system stability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a high- and low-voltage ride-through control method for energy storage converters based on a large predictive model. By using the large predictive model to anticipate voltage recovery trends in advance, the ride-through strategy is dynamically adjusted while ensuring equipment safety. This avoids the thermal stress risks caused by blindly extending the support and improves the resilience of the power grid.

[0005] To address the aforementioned technical problem, the present invention employs the following technical solution: When a voltage drop occurs in the power grid and exceeds the time specified by the standard ride-through curve, the converter collects parameters such as DC-side bus voltage, inductor current, and IGBT junction temperature to assess its own thermal stress state in real time. Combining this with the current reactive power demand and output capacity, a preset power grid fault prediction model is used to calculate the predicted fault duration. If it is determined that continuous operation conditions are met, the converter automatically enters extended ride-through mode, maintaining grid connection and injecting the maximum permissible reactive current into the power grid until the voltage recovers or the equipment safety boundary is reached.

[0006] By using a pre-defined power grid fault prediction model, the duration of the fault is predicted, providing a confidence interval for the current fault persistence trend, and the reactive power support strength is dynamically adjusted. When the predicted voltage is about to recover to the normal range, the output current is gradually reduced to avoid secondary impacts; if the fault persists and thermal stress accumulation approaches the safety threshold, a protection mechanism is activated to disconnect the power grid in an orderly manner, ensuring equipment safety. During the extended ride-through process, the converter continuously monitors the dynamic recovery of the power grid voltage and its own operating status, achieving closed-loop optimization of the support strategy. By integrating the fault recovery trend output by the large prediction model with the real-time thermal stress assessment results, the reactive current setpoint is dynamically adjusted to achieve a balance between system stability and equipment safety.

[0007] This method dynamically calculates the safely extended ride-through time by real-time monitoring of the magnitude and duration of grid fault voltage, combined with the converter's own thermal margin and reactive power output capability. Under the premise of not exceeding the equipment's tolerance limits, it intelligently determines whether to maintain grid-connected operation and continuously provides reactive power support, thereby improving the system's voltage recovery capability. This technology effectively fills the control gap outside the coverage of the standard curve, enhancing the converter's adaptability and support capability against prolonged voltage dips.

[0008] The beneficial effects of this invention are as follows: This invention identifies the grid voltage trend in real time by constructing a dynamic prediction model, and calculates the support duration and support current magnitude for the current fault ride-through using a multi-objective optimization algorithm. This method effectively expands the range of the converter's fault ride-through curve, enabling it to dynamically extend the support duration based on actual grid needs even after the standard ride-through time has ended. This solves the problem of insufficient reactive power support leading to decreased system stability under grid faults outside the standard curve, realizing the converter's transformation from "passive compliant grid disconnection" to "active delayed support." This method significantly improves the dynamic support capability for the grid while ensuring converter safety. Through this intelligent calculation method, the converter can remain connected to the grid even after the standard disconnection time, extending the support time by several seconds to tens of seconds, significantly improving the system voltage recovery window. The application of this method can effectively reduce the risk of cascading grid disconnection caused by voltage faults and improve the resilience of the grid under complex disturbances. Attached Figure Description

[0009] Figure 1 This is an overall block diagram of an energy storage converter equipped with the method described in Example 1;

[0010] Figure 2 This is a block diagram illustrating the principle of the method described in Example 1;

[0011] Figure 3 This is a flowchart of the method described in Example 1;

[0012] Figure 4 This is a schematic diagram illustrating the training and prediction of a random forest prediction model.

[0013] Figure 5 A schematic diagram for calculating the safety margin of an energy storage converter;

[0014] Figure 6 This is the sequential control loop during fault crossing. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0016] Example 1

[0017] This embodiment discloses a high- and low-voltage ride-through control method for energy storage converters based on a large predictive model. The controlled object of this method is a typical LC-type filtered grid-connected converter. Figure 1 The diagram shows the overall block diagram of an energy storage converter equipped with the method described in Example 1. It includes a main power circuit, a DSP control loop, and a predictive calculation MCU. Part A, marked with a solid line, represents the method described in this example. Upcc is the grid connection point, Rs and Linv are the equivalent resistance of the converter and their filter inductance, Rg and Lg are the equivalent resistance and inductance of the grid impedance, and Uinv and Iinv are the AC voltage and current of the converter, respectively. When no fault occurs, the converter performs system phase-locking by acquiring the converter voltage Uinv, simultaneously determining grid faults based on the Uinv value, calculating the instantaneous power of the converter by acquiring Iinv, and performing dual-loop control. This is a conventional energy storage converter control method.

[0018] The predictive fault ride-through control method described in this embodiment does not start when no grid fault occurs. The system adopts a conventional power loop nested current loop control strategy. When a grid fault occurs, the control loop switches to this method (part A marked with solid line). The two are selected by software logic.

[0019] like Figure 2 As shown, this method is implemented through an external storage device, a predictive computing MCU, and a DSP control loop. The external storage device can be a common FLASH or EEPROM chip, the predictive computing MCU should be a processor with performance of STM32F407 or higher, and the DSP should be a processor with performance of TMS320F20335 or higher.

[0020] External high-capacity storage devices are primarily used for storing historical data and preset parameters, providing initial parameters for large-scale model training. The predictive MCU, acting as the carrier of the predictive algorithm, continuously extracts historical data and preset parameters from the external memory for large-scale model training and updates the predictive model parameters in real time. Simultaneously, the trained large-scale model is combined with current power grid state data for dynamic inference to determine the fault duration and voltage recovery trend. The DSP control loop is used for comprehensive logic judgment of power grid fault ride-through and closed-loop control of fault ride-through current.

[0021] like Figure 3 As shown, the present invention includes the following steps:

[0022] Step 1: Obtain the grid voltage and determine whether the grid voltage meets the high and low voltage ride-through requirements. If it does, proceed to Step 2; otherwise, continue with Step 1.

[0023] Step 2: Collect current power grid characteristic quantities, input them into the pre-trained power grid fault prediction model, and obtain the predicted fault duration;

[0024] Step 3: Collect historical operating data and hardware parameters of the energy storage converter, calculate the instantaneous junction temperature margin and comprehensive cost function, input the current operating data of the converter, the instantaneous junction temperature margin and comprehensive cost function into the pre-trained deep learning model, and obtain the converter safety margin for the next N sampling times.

[0025] Step 4: Calculate the minimum crossing support time required by the system;

[0026] Step 5: Issue a fault ride-through start signal and start the ride-through timer. Monitor the grid connection point voltage, current and system operating status in real time, and support the grid voltage through reactive current regulation.

[0027] Step 6: Real-time detection and judgment of whether the grid connection point voltage has returned to the normal operating range. If yes, issue a fault crossing stop signal, exit the fault crossing and return to the normal operating state. Otherwise, judge whether the time point when the normal state has not been restored has reached the minimum crossing support time. If yes, return to step 5; otherwise, proceed to step 7.

[0028] Step 7: Compare the predicted fault duration with the set threshold to determine the fault type. If it is a short-term or long-term fault, compare the converter safety margin with the set threshold. If the converter safety margin is greater than the set threshold, the converter safety margin is sufficient and the maximum reactive current is output to support the grid voltage. Otherwise, the reactive current command is derated. When it is a permanent fault, a fault ride-through stop signal is issued, the fault ride-through control is exited and the operation ends.

[0029] In this embodiment, the power grid fault prediction model is a random forest prediction model, such as... Figure 4 As shown, it includes two stages: training inference and model prediction. The specific construction process is as follows:

[0030] First, data related to power grid operation status and faults are extracted from the local power grid historical database to construct an initial sample set for training the fault prediction model. The initial sample set mainly includes the following four types of feature information: First, power grid fault type characteristics, including typical distribution network fault categories such as voltage sags, single-phase grounding faults, and phase-to-phase short circuits; second, local meteorological environment information, including temperature, humidity, and weather conditions (sunny, cloudy, rainy, snowy, etc.); third, geographical features of the fault location, including terrain type (mountains, hills, plains, etc.) and natural disaster influencing factors (mudslides, typhoons, icing, etc.); and fourth, load operation information at the fault location, including load rate, overload status, and load type (such as inductive loads like electric motors).

[0031] Based on the initial sample data mentioned above, the multi-source features are normalized and vectorized according to the corresponding formula. Construct multi-source fault feature vectors and form a training sample set. Wherein: For multi-source fault feature vectors, This represents the actual duration of the fault.

[0032] Among them, the multi-source fault feature vector is used to characterize the changes in the power grid operation status and external environment before and after the fault occurs, and the actual fault duration is used as the prediction target and label of the model to achieve fault duration prediction and fault development trend analysis.

[0033] During the training of a random forest, the Bootstrap resampling method is used to randomly sample multiple subsets from the training dataset D. Each subset is used to train a decision tree. If the random forest contains M decision trees, then M training subsets are generated through resampling. .

[0034] For each subset We construct a regression decision tree. During the splitting process at each node, we randomly select L features from all features for the optimal split, thereby reducing the risk of model overfitting.

[0035] Let the node partitioning function be ,in: As input features, Parameters are assigned to the decision tree.

[0036] According to the formula The mean squared error (MSE) is obtained, and the optimal partition is achieved by minimizing the MSE of samples within each node. Where: This represents the actual duration of the fault. Predict the duration of the failure for the model.

[0037] After training all decision trees, combine the M decision trees to form a random forest prediction model. For each input feature vector X, each decision tree outputs a prediction result. .

[0038] According to the formula The average of all decision tree outputs is obtained and used as the final prediction. Where: M represents the number of decision trees used to predict the duration of a failure.

[0039] When the system detects a voltage drop in the power grid, the predictive MCU collects current power grid characteristics and environmental information in real time, constructs a multi-source fault feature vector, and inputs it into a trained random forest model for inference calculation to obtain the predicted fault duration. .

[0040] Figure 5 The flowchart below shows the calculation process for the safety margin of an energy storage converter. The safety margin is a measure of the overload capacity of the IGBT devices in the energy storage converter, representing the critical point at which the IGBT may fail. It is characterized by a comprehensive cost function to achieve predictive thermal protection rather than reactive current limiting, thereby extending the device's lifespan and reducing system operation and maintenance costs. The specific construction process is as follows:

[0041] First, collect historical operating data of the energy storage converter during continuous operation, including: maximum output power. Overload duration Maximum junction temperature of the device Simultaneously, converter hardware parameters are collected, including: saturation voltage drop constant. On-state resistance Activate energy Shutting off energy Diode reverse recovery energy Rated current Rated voltage Rated junction temperature Foster fourth-order thermal network parameters ( (i=1-4), filter inductor L, DC support capacitor C, heat sink thermal resistance wait.

[0042] Power loss is calculated in real time during each sampling period:

[0043] according to Calculate conduction loss.

[0044] according to Calculate switching losses.

[0045] according to Calculate the total loss.

[0046] in, For collector real-time current, This is the average collector current. The above calculations represent the effective value of the collector current for a single IGBT device. The overall loss requires calculating the relevant parameters for each IGBT device separately and then summing them up based on the number of devices or the number of parallel connections. For switching frequency, This is the DC bus voltage.

[0047] The instantaneous junction temperature is calculated using a fourth-order Foster thermal network model. Through state-space discretization and recursion, and with the initial state using the steady-state value of the previous cycle for hot start-up, the following equation can be obtained:

[0048] ,

[0049] in, This is the state vector of a fourth-order Foster thermal network, where each element corresponds to the temperature state of a heat capacity node in a first-order RC network. Here is the state transition matrix. For the input matrix, For the output matrix, This refers to the instantaneous total power loss of the IGBT. This represents the instantaneous junction temperature of the IGBT. This is the instantaneous value of the ambient temperature. This represents the increase in radiator temperature relative to the environment.

[0050] Then according to Calculate the instantaneous junction temperature margin.

[0051] After obtaining the above indicators, a comprehensive cost function is constructed based on the actual situation to characterize the overall risk by designing weights:

[0052] ,

[0053] in, These are weighting coefficients, which should be set based on practical experience. To allow for a cumulative overload time limit, For the instantaneous output power of the energy storage converter, This represents the maximum output power from historical data.

[0054] A training dataset is constructed, with samples organized in a sliding window manner: the input of each sample is a multi-dimensional feature vector of the past 60 sampling points, covering DC voltage, positive and negative sequence dq current components, junction temperature, power loss, active power, switching frequency, overload flag, instantaneous junction temperature margin, and comprehensive cost function value. The sample is gradually transitioned from historical operating data at the beginning and filled with the currently collected real-time data at the end, forming a continuous input sequence containing long-term historical context and short-term dynamic information; the corresponding label is the safety margin sequence of the next 30 sampling points. With the help of data augmentation algorithms, a total of no less than 10,000 samples are finally generated, and divided into training, validation, and test sets in an 8:1:1 ratio.

[0055] Supervised learning is used to optimize the parameters of the LSTM network. The network structure adopts a two-layer LSTM architecture, followed by an attention mechanism to improve the modeling ability of key time points in the sequence. Finally, it is mapped to the prediction margin vector through a linear layer. To balance prediction accuracy and engineering safety, the loss function is designed as a multi-objective composite form.

[0056] ,

[0057] in To predict the safety margin, M is the actual safety margin. Let J be the predicted cost function, and J be the true cost function. This is a safety threshold; The mean squared error of the margin sequence is the principal loss term. To predict the auxiliary consistency loss between the cost function and the true cost function, As an out-of-bounds penalty term, it generates an additional gradient when the prediction cost function exceeds the safety threshold, pushing the model to tend towards conservative estimation; This is a hyperparameter.

[0058] The AdamW optimizer is used to improve training performance. After typical training, the root mean square error (RMSE) of the validation set margin sequence is controlled within a reasonable range. The predicted safety margin is extracted from the output sequence, compared with the safety threshold, and passed to the subsequent control decision layer to dynamically adjust the reactive power support strategy and current limiting threshold of the converter, maximizing the voltage support capability while ensuring the safety of power electronic devices.

[0059] After completing the prediction of grid fault duration and the assessment of converter safety margin, the prediction results are used as inputs to make ride-through control decisions for the energy storage converter. The specific process is as follows:

[0060] The control system acquires the grid connection point voltage in real time. The system then determines whether the grid voltage exceeds the set operating range. When the grid voltage is within the normal range, the system maintains normal operation and continuously monitors the voltage; when the grid voltage is detected to be outside the normal range, it determines that a voltage anomaly has occurred in the grid and enters the ride-through control process.

[0061] When the grid voltage is abnormal, the control system obtains the predicted fault duration output by the random forest prediction model. And the predicted safety margin output by the safety margin prediction module.

[0062] The minimum crossing support time required by the calculation system is calculated according to GBT34120-2023.

[0063] When an abnormal grid voltage is detected, the system issues a fault ride-through initiation signal and enters ride-through control mode. In this mode, the grid connection point voltage, current, and system operating status are monitored in real time. Reactive current regulation is used to support the grid voltage.

[0064] Calculate the reference value of reactive current based on the voltage deviation at the grid connection point: ,

[0065] in, This is a reference value for reactive current. Rated voltage; This refers to the voltage at the grid connection point. This is the voltage support coefficient.

[0066] When the system detects an abnormal grid voltage and enters fault ride-through control mode, the control system will continuously monitor changes in the grid connection point voltage. If the grid connection point voltage recovers to the normal operating range within the minimum ride-through support time, a fault ride-through stop signal will be immediately issued, exiting fault ride-through control and returning to normal operation. If the actual duration of the system fault (i.e., the duration for which the grid connection point voltage deviates from the normal operating range) exceeds the minimum ride-through support time, the fault duration will be predicted. Fault type determination is performed based on the preset fault duration threshold: if the predicted fault duration value... If the value is not greater than the short-term fault threshold, it is determined to be a short-term fault; if the fault duration is less than the predicted value... If the fault duration is greater than the short-term fault threshold but not greater than the long-term fault threshold, it is determined to be a long-term fault; if the fault duration is predicted... If the failure exceeds the long-term failure threshold, it is considered a permanent failure.

[0067] When a short-term or long-term fault is identified, the converter's safety margin is first assessed. The real-time safety margin is compared to a set safety threshold: if the real-time safety margin is less than the set safety threshold, reactive current derating control is applied; if the real-time safety margin is greater than the set safety threshold, maximum reactive current is output to support the grid voltage. When a permanent fault is identified, the control system issues a fault ride-through stop signal, exits fault ride-through control, and terminates operation.

[0068] Upon receiving a fault ride-through command, the DSP loop switches the control algorithm to a predictive fault ride-through algorithm. The current control loop, for example... Figure 6 As shown, the system decomposes the collected voltage and current into positive and negative sequences, constructing a positive-sequence DQ control loop and a negative-sequence DQ control loop. This dual-sequence DQ current control continues until a fault ride-through termination command is received. The current reference value is updated periodically to respond to the latest output of the predictive algorithm, thus achieving real-time adaptation to grid fault dynamics. When the positive-sequence voltage amplitude continuously and stably returns to the normal range, and the negative-sequence component remains below the engineering threshold for a certain time delay, the ride-through process ends. The controller gradually restores the current reference value to the setpoint under normal power control. All data collected during the ride-through process is recorded and appended to the historical dataset to support further optimization of the predictive model.

[0069] The above description is merely the basic principle and preferred embodiment of the present invention. Improvements and substitutions made by those skilled in the art based on the present invention are within the scope of protection of the present invention.

Claims

1. A method for high and low voltage ride-through control of an energy storage converter based on a large prediction model, characterized by: include: Step 1: Obtain the grid voltage and determine whether the grid voltage meets the high and low voltage ride-through requirements. If it does, proceed to Step 2; otherwise, continue with Step 1. Step 2: Collect current power grid characteristic quantities, input them into the pre-trained power grid fault prediction model, and obtain the predicted fault duration; The power grid fault prediction model is a random forest prediction model. The Bootstrap resampling method is used to randomly sample multiple subsets from the training dataset D. Each subset is used to train a decision tree. After all decision trees are trained, they are combined to form a random forest prediction model. For the input feature vector, each decision tree outputs a prediction result. The average of the prediction results of all decision trees is calculated and used as the final prediction value of the random forest prediction model, that is, the predicted fault duration. Step 3: Collect historical operating data and hardware parameters of the energy storage converter, calculate the instantaneous junction temperature margin and comprehensive cost function, input the current operating data of the converter, instantaneous junction temperature margin and comprehensive cost function into the pre-trained deep learning model to obtain the converter safety margin for the next N sampling points; The formula for calculating the comprehensive cost function is: ， wherein is a comprehensive cost function, which is a measure of the overload capability of the energy storage converter IGBT device, is a transient junction temperature margin, , , is a weight coefficient, is the instantaneous output power of the energy storage converter, is the maximum output power in the historical data, is the overload duration, is the allowable overload cumulative time limit; Step 4: Calculate the minimum crossing support time required by the system; Step 5: Issue a fault ride-through start signal and start the ride-through timer. Monitor the grid connection point voltage, current and system operating status in real time, and support the grid voltage through reactive current regulation. Step 6: Real-time detection and judgment of whether the grid connection point voltage has recovered to the normal operating range. If yes, issue a fault ride stop signal, exit the fault ride and return to normal operation. Otherwise, determine whether the minimum ride support time has been reached. If yes, return to step 5; otherwise, proceed to step 7. Step 7: Compare the predicted fault duration with the set threshold to determine the fault type. If it is a short-term or long-term fault, compare the real-time converter safety margin with the set threshold. If the real-time converter safety margin is greater than the set threshold, the converter safety margin is sufficient and the maximum reactive current is output to support the grid voltage. Otherwise, the reactive current command is derated.

2. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 1, characterized in that: Training dataset ,in The fault multi-source feature vector includes power grid fault type features, local meteorological environment information, geographical features of the fault location, and load operation information of the fault location. To predict the actual duration of the fault, the input features of the random forest prediction model also include local meteorological environment information and geographical features of the fault location.

3. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 1, characterized in that: When training a decision tree, during the splitting process at each node, L features are randomly selected from all features for optimal splitting. Let the node splitting function be... ,in For input features, To divide the parameters of the decision tree, according to the formula The mean squared error (MSE) is obtained, and the optimal partition is achieved by minimizing the mean squared error of samples within each node. This represents the actual duration of the fault. The model predicts the duration of the fault, where n is the optimal step size.

4. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 1, characterized in that: The formula for calculating the instantaneous junction temperature margin is: ， in For instantaneous junction temperature margin, This is the highest junction temperature of the power devices in the energy storage converter. This represents the instantaneous junction temperature of the power devices in the energy storage converter. This is the rated junction temperature.

5. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 4, characterized in that: The instantaneous junction temperature was calculated using a fourth-order Foster thermal network model. The calculation formula is: ， in Here is the state vector of a fourth-order Foster thermal network. Here is the state transition matrix. For the input matrix, For the output matrix, This refers to the instantaneous total power loss of the IGBT. This represents the instantaneous junction temperature of the IGBT. This is the instantaneous value of the ambient temperature. This represents the increase in radiator temperature relative to the environment.

6. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 1, characterized in that: The deep learning model is an LSTM network. Supervised learning is used to optimize the parameters of the LSTM network. The network structure adopts a two-layer LSTM architecture, followed by an attention mechanism to improve the ability to model key time points in the sequence. Finally, it is mapped to the prediction margin vector through a linear layer.

7. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 6, characterized in that: The loss function of the LSTM network is: ， in The mean square error of the margin sequence. To predict the safety margin, M is the actual safety margin. For the prediction cost function With the true cost function Auxiliary consistency loss, For boundary crossing penalties, As a safety threshold, This is a hyperparameter.

8. The high and low voltage ride-through control method for energy storage converters based on a large predictive model according to claim 1, characterized in that: In step 7, when the fault is permanent, a fault crossing stop signal is issued, the fault crossing control is exited, and the operation ends.

Images