A Power Feedforward Decoupling Control Method for Electrolytic Hydrogen Production PWM Rectifier under Renewable Energy Fluctuations

By introducing a predictive feedforward control channel and dynamic confidence factor weighting into the PWM rectifier, the problems of response lag and strong coupling of the control channel in the PWM rectifier under renewable energy fluctuations are solved, realizing a fast response to power disturbances and improving system stability.

CN123122272APending Publication Date: 2026-11-17HEBEI JIANTOU NEW ENERGY CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611263785.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing PWM rectifier control strategies suffer from lag in response and strong coupling of control channels when faced with rapid and large power fluctuations from renewable energy sources. This leads to DC bus voltage fluctuations and AC side current distortions, making it difficult to achieve effective decoupling and self-correction.

Method used

A predictive feedforward control channel based on a long short-term memory network is introduced. By combining feedback regulation and predictive correction components, future power changes are predicted using historical data from deep learning. Control commands are then weighted using dynamic confidence factors to achieve active suppression of power disturbances.

Benefits of technology

It improves the dynamic response speed of the system, avoids the response lag caused by bandwidth limitations in traditional control, and achieves rapid suppression of power disturbances and ensures system stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122272A_ABST
    Figure CN123122272A_ABST
Patent Text Reader

Abstract

This application provides a power feedforward decoupling control method for a PWM rectifier in hydrogen electrolysis under renewable energy fluctuations. The method includes: acquiring operating data of the PWM rectifier; determining future power changes based on the operating data using a pre-built prediction model and calculating a feedforward compensation component; calculating a feedback adjustment component based on the difference between the setpoint and measured value of the DC bus voltage; estimating the predicted trajectory of the DC bus voltage based on future power changes; generating a prediction correction component based on the deviation between the predicted and measured trajectories, and determining a dynamic confidence factor based on the historical statistical characteristics of the deviation; and dynamically weighting the feedback adjustment component, the feedforward compensation component, and the prediction correction component to synthesize the final d-axis current reference value. This application can suppress DC bus voltage fluctuations, achieve dynamic decoupling between the AC and DC sides, and improve the system stability under power fluctuations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power electronic conversion technology, and in particular to a power feedforward decoupling control method for an electrolytic hydrogen production PWM rectifier under renewable energy fluctuations. Background Technology

[0002] In systems that utilize renewable energy sources such as wind and solar power for hydrogen electrolysis, the PWM rectifier is a crucial interface connecting the fluctuating power supply to the DC electrolyzer. Its control performance determines the stability of the DC bus voltage and the quality of the AC power, ensuring the efficient and safe operation of the electrolyzer and maintaining grid friendliness.

[0003] To address power fluctuations in renewable energy sources, current mainstream PWM rectifier control strategies primarily employ a dual closed-loop control architecture based on a voltage outer loop and a current inner loop in the dq coordinate system. To improve dynamic response, some studies introduce a direct feedforward compensation channel based on DC-side load current or power measurements into the voltage outer loop, aiming for rapid adjustment after disturbances occur.

[0004] However, the aforementioned existing technologies still face technical problems such as response lag and strong coupling of control channels when dealing with rapid and large power fluctuations. Specifically, the following technical problems exist: In traditional dual-loop control, the bandwidth of the outer voltage loop is usually limited (e.g., below 20Hz) to ensure system stability, resulting in an inherent delay in its response to input power disturbances and an inability to suppress DC bus voltage fluctuations caused by power changes in a timely manner. These DC bus voltage fluctuations are transmitted to the inner current loop through the feedback channel, causing distortion and phase shift of the AC side input current, forming strong coupled oscillations on both the AC and DC sides. Although a direct feedforward channel is introduced, this feedforward is essentially open-loop, its performance depends on the accuracy of the model, and it cannot sense and correct its own compensation errors. In actual operating conditions with varying fluctuation modes, inaccurate feedforward compensation may even inject disturbances in the opposite direction, further deteriorating system stability. Therefore, realizing a dynamic decoupling control that can respond to power disturbances and has self-correcting capabilities is a technical problem that urgently needs to be solved in the current field. Summary of the Invention

[0005] Purpose of the invention: To provide a power feedforward decoupling control method for a PWM rectifier in hydrogen electrolysis under renewable energy fluctuations, in order to solve one of the aforementioned problems in the existing technology.

[0006] Technical solution: A power feedforward decoupling control method for a PWM rectifier in hydrogen electrolysis under renewable energy fluctuations, comprising:

[0007] Obtain the operating data of the PWM rectifier;

[0008] Based on operational data, the future power change is determined using a pre-built prediction model, and the feedforward compensation component is calculated.

[0009] The feedback adjustment component is calculated based on the difference between the pre-configured set value of the DC bus voltage and the measured DC bus voltage obtained from sampling.

[0010] Based on future power changes, the predicted trajectory of DC bus voltage is calculated.

[0011] Based on the deviation between the predicted trajectory and the measured trajectory of the DC bus voltage obtained from the operating data, a prediction correction component is generated, the deviation is accumulated in multiple control cycles, and a dynamic confidence factor for adjusting the intervention depth of the feedforward path is determined based on the historical statistical characteristics of the deviation.

[0012] By combining the feedback adjustment component, the feedforward compensation component, and the prediction correction component, and by using a dynamic confidence factor to weight the feedforward compensation component and the prediction correction component, a reference value for the d-axis current is synthesized.

[0013] Beneficial Effects: To address the response lag problem of traditional control, this invention introduces a forward-looking predictive feedforward control channel. This channel utilizes a Long Short-Term Memory (LSTM) network and, through deep learning of the dynamic evolution of historical operating data, can predict future input power fluctuations several control cycles in advance, thereby generating a feedforward compensation current to offset the disturbance. This predictive control method bypasses the inherent delay caused by bandwidth limitations in traditional voltage feedback loops, achieving active suppression of power disturbances rather than a passive response. Attached Figure Description

[0014] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0015] Figure 1 This is an overall flowchart of a power feedforward decoupling control method for an electrolytic hydrogen production PWM rectifier under renewable energy fluctuations, as described in an embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating the steps for obtaining the operating data of the PWM rectifier in an embodiment of this application.

[0017] Figure 3 This is a flowchart illustrating the steps involved in determining future power changes using a pre-built prediction model, as described in this application embodiment.

[0018] Figure 4 This is a schematic diagram of three different control strategies in the embodiments of this application. Detailed Implementation

[0019] Example 1

[0020] This embodiment describes the overall process of a power feedforward decoupling control method for a PWM rectifier in hydrogen electrolysis under renewable energy fluctuations, such as... Figure 1 As shown, the specific steps are as follows:

[0021] Step 101: Obtain the operating data of the PWM rectifier.

[0022] In this embodiment, the operating data of the PWM rectifier refers to the collection of physical quantities that reflect the instantaneous electrical state of the rectifier in each control cycle. Specifically, this operating data is acquired by deploying high-precision sensors on both the AC input and DC output sides of the rectifier system. The AC side includes at least three-phase voltage sensors and three-phase current sensors for real-time sampling of the instantaneous values ​​of the three-phase AC voltage and current; the DC side includes at least a DC voltage sensor for real-time sampling of the instantaneous value of the DC bus voltage. These sensors sample synchronously at the switching frequency set by the control system, obtaining a series of discrete time-series data. This raw sampled data collectively forms the data basis for subsequent calculation steps, providing a real-time and unique input source for system state observation, prediction, and control.

[0023] Step 102: Based on the operating data, use the pre-built prediction model to determine the future power change and calculate the feedforward compensation component.

[0024] This embodiment constitutes a predictive feedforward path in the control framework, which combines feedback regulation with disturbance pre-compensation to achieve real-time compensation for power fluctuations.

[0025] The pre-built prediction model is obtained through offline training. Specifically, historical operating data containing events of large fluctuations in renewable energy power is acquired, and multiple sets of training samples are constructed from them. Each set of training samples includes a sliding window feature matrix as input features and the future actual power change as a ground truth label.

[0026] By optimizing the algorithm to minimize the error between the predicted power output of the Long Short-Term Memory (LSTM) network and the true label, the internal network parameters of the LSTM network are iteratively adjusted to obtain a pre-built prediction model.

[0027] Specifically, the pre-built predictive model is a deep learning model trained offline using a large amount of historical operating data. Preferably, it can be a Long Short-Term Memory (LSTM) network, which excels at extracting dynamic evolution patterns from time-series data. This model receives a feature sequence formed from processed operating data as input and outputs a predicted value for the change in input power on the renewable energy side over one or more future control cycles—that is, the future power change. After obtaining this predicted value, the system converts it into an equivalent current command that can preemptively offset the power disturbance, i.e., a feedforward compensation component, based on the PWM rectifier's own power balance equation. The purpose of this path is to generate corresponding control actions before the power disturbance actually affects the DC bus voltage, thereby improving the system's dynamic response speed.

[0028] Step 103: Calculate the feedback adjustment component based on the difference between the pre-configured set value of the DC bus voltage and the measured DC bus voltage obtained from sampling.

[0029] As an optional implementation, this embodiment utilizes a traditional feedback path in the control framework to achieve steady-state accuracy of the DC bus voltage. Specifically, this implementation can be a proportional-integral (PI) controller. This controller continuously compares the given target setpoint of the DC bus voltage with the current actual measured value obtained from the operating data; the deviation between the two is the setpoint tracking error, reflecting the gap between the current control effect and the desired target. The PI controller performs proportional and integral calculations on this deviation to generate a current regulation command used to eliminate the steady-state error, i.e., the feedback regulation component. Unlike the feedforward path, this feedback path is a lag regulation mechanism, responsible for correcting residual errors that the feedforward path failed to compensate for, as well as other unmodeled disturbances, providing a fundamental guarantee for stable system operation.

[0030] Step 104: Based on future power changes, calculate the predicted trajectory of the DC bus voltage; based on the deviation between the predicted trajectory and the measured trajectory of the DC bus voltage obtained from the operating data, generate a prediction correction component, accumulate the deviation over multiple control cycles, and determine the dynamic confidence factor for adjusting the depth of feedforward path intervention based on the historical statistical characteristics of the deviation.

[0031] Accordingly, this embodiment constructs a feedforward self-correcting closed loop. Based on the energy conservation relationship between future power changes and DC bus capacitance, the system can calculate the theoretical value that the DC bus voltage should reach in the next control cycle, if the prediction is accurate. This theoretical value is the predicted trajectory of the DC bus voltage. The system compares this predicted trajectory with the actual measured DC bus voltage in the next cycle, i.e., the measured trajectory. The deviation between the two is defined as the prediction tracking error in this invention, and this error signal characterizes the accuracy of the prediction model itself.

[0032] Furthermore, based on this prediction tracking error, this embodiment performs two operations in parallel: First, through a fast compensation stage, a current component is generated to instantly offset the disturbance introduced by a single inaccurate prediction, i.e., the prediction correction component; Second, on a longer time scale, statistical analysis is performed on the historical sequence of the prediction tracking error, such as calculating its mean and standard deviation, to obtain an index for calculating the overall reliability of the prediction model under the current operating conditions, i.e., the dynamic confidence factor.

[0033] Step 105: Combine the feedback adjustment component, the feedforward compensation component, and the prediction correction component, and use the dynamic confidence factor to weight the feedforward compensation component and the prediction correction component to synthesize the d-axis current reference value.

[0034] Based on this, this embodiment is the final output stage of the entire control method, organically combining the three independently calculated current components to form a unified control command. Specifically, the final d-axis current reference value is composed of three superimposed parts: the feedback adjustment component from the traditional feedback path, and the sum of the feedforward compensation component and the prediction correction component after being weighted by the dynamic confidence factor. The dynamic confidence factor acts as an intelligent regulating valve: when the prediction model performs well and the prediction tracking error is small, the factor approaches 1, allowing the feedforward and correction paths to play a dominant role, resulting in a rapid system response; when the prediction model performs poorly and the prediction tracking error increases, the factor automatically decreases and approaches 0, smoothly reducing the intervention of the feedforward and correction paths, causing the system to automatically degenerate to a mode dominated by traditional feedback control, ensuring the stability and robustness of the system. This dynamically weighted synthesis method ensures that the system can adopt a control strategy adapted to the current operating conditions under different operating conditions and different prediction accuracies.

[0035] Example 2

[0036] This embodiment describes the specific technical implementation process for obtaining the operating data of a PWM rectifier.

[0037] In one possible implementation, the operating data of the PWM rectifier is acquired, such as... Figure 2 As shown, it specifically includes:

[0038] Step 201: Sample the three-phase AC voltage and three-phase AC current cycle by cycle, and project them onto the synchronous rotating coordinate system through Park transformation to extract the d-axis voltage component, d-axis current component and q-axis current component.

[0039] As an optional implementation, in this embodiment, the three-phase AC electrical quantities on the grid side are converted into easily controllable DC components. Specifically, the control system uses voltage and current sensors installed on the AC input side of the PWM rectifier, and a sampling rate synchronized with the rectifier switching frequency, for example, 20kHz, to collect the instantaneous voltage and current values ​​of the three-phase AC phases a, b, and c, frame by frame. Simultaneously, the system tracks the phase angle θ of the grid voltage in real time through a phase-locked loop (PLL). After obtaining the three-phase AC quantities and the synchronous phase angle, the system performs a standard Park transformation. This transformation is a well-known technique in the field of power electronics control; its function is to convert the sinusoidal AC quantities in the three-phase stationary coordinate system to a dq coordinate system that rotates synchronously with the grid voltage vector as a reference, obtaining an approximate DC d-axis voltage component V. d d-axis current component i d and q-axis current component i q Under a grid voltage-oriented control strategy, the d-axis is defined as being in the same direction as the grid voltage vector, and the q-axis leads the d-axis by 90°. Correspondingly, the d-axis current component i... d It is mainly used to control the flow of active power, while the q-axis current component i q It is mainly used to control reactive power, by controlling i q A power factor of zero allows for unity power factor operation.

[0040] Step 202: Normalize the d-axis voltage component, d-axis current component, q-axis current component, and synchronously sampled measured DC bus voltage, and concatenate them into a feature vector as running data.

[0041] Based on this, this embodiment constructs an input data structure that can be used by subsequent prediction models. Specifically, in each control cycle, the system will input the d-axis voltage component V... d d-axis current component i d q-axis current component i q And the measured DC bus voltage V obtained by synchronous sampling through the DC bus side voltage sensor. dc The above four physical quantities are then concatenated. Before concatenation, normalization is required to eliminate the adverse effects on model training caused by differences in dimensions and numerical ranges between different physical quantities. Preferably, a maximum-minimum normalization method can be used. For example, for the d-axis current component i... d Its normalized calculation formula is as follows:

[0042] i d_norm =(i d -i d_min ) / (i d_max -i d_min );

[0043] Among them, id_norm i is the normalized d-axis current value. d i is the current sampled value. d_max and i d_min These are the global maximum and minimum values ​​of the d-axis current, statistically derived from a large amount of historical data during the offline analysis phase.

[0044] After all four physical quantities have undergone similar processing, the resulting four normalized values ​​between 0 and 1 are arranged in a fixed order, for example (V... d_norm i d_norm i q_norm V dc_norm These data are concatenated to form a four-dimensional feature vector. This feature vector represents the operational data for the current control cycle, describing the system's operational state at that moment in a standardized form.

[0045] In some alternative implementations, the feature vector may also include other derived features to provide information about power change trends to the prediction model. For example, the AC-side active power P for the current cycle may be additionally calculated. ac and reactive power Q ac This data is then normalized and added to the eigenvector, making it a six-dimensional eigenvector. AC-side active power P ac With reactive power Q ac The calculation of is a well-known technique in the art, and its formula is as follows:

[0046] P ac =1.5*(V d *i d +V q *i q );

[0047] Q ac =1.5*(V q *i d -V d *i q );

[0048] Among them, V q This represents the q-axis voltage component.

[0049] Furthermore, this embodiment also includes the process of constructing a sliding window feature matrix. The system internally maintains a first-in, first-out (FIFO) buffer. The depth N of this buffer is a preset time step length, for example, N=150. If the control period T... s If the value is 50μs, corresponding to a sampling frequency of 20kHz, then the corresponding time window length is 7.5ms.

[0050] In some optional implementations, if the system uses a lower frequency outer loop control cycle (e.g., 1 ms), then N can be set to 150, corresponding to a time window length of 150 ms. The selection of the control cycle should comprehensively consider the balance between prediction accuracy requirements and real-time computing power. In each control cycle, the latest feature vector is pushed into the tail of the buffer, while the oldest feature vector at the head of the buffer is popped. This buffer stores a sequence of historical feature vectors dating back N control cycles from the current moment. Stacking N four-dimensional or six-dimensional feature vectors in chronological order forms a sliding window feature matrix of dimension (N,4) or (N,6). This matrix not only contains the system state at the current moment but also contains dynamic evolution information over a short period of time, which will serve as the complete input to the LSTM prediction model.

[0051] Example 3

[0052] In one possible implementation, the future power change is determined using a pre-built prediction model, such as... Figure 3 As shown, it specifically includes:

[0053] Optionally, a sliding window feature matrix is ​​constructed based on the running data in a time step sequence; the sliding window feature matrix is ​​input into a pre-constructed prediction model, wherein the prediction model is a long short-term memory network; and the long short-term memory network processes the sliding window feature matrix along the time step sequence to output the future power change.

[0054] In this embodiment, achieving high-precision prediction hinges on the selection of the prediction model. Long Short-Term Memory (LSTM) networks are preferred because the power fluctuation sequences of renewable energy exhibit time-dependent and nonlinear characteristics. Traditional linear models or simple feedforward neural networks struggle to capture complex dynamic relationships. LSTM, through its unique input, forget, and output gate structures, can selectively memorize long-term dependencies from historical information and filter out irrelevant short-term disturbances, resulting in higher prediction accuracy for time series problems with historical correlations, such as power fluctuations.

[0055] As an optional implementation, the pre-built prediction model can have a stacked network structure consisting of two LSTM layers and one fully connected output layer. For example, the first LSTM layer can have 128 neurons, the second LSTM layer can have 64 neurons, and finally, the fully connected layer maps the high-dimensional features to the final single-value output, i.e., the future power change. The physical meaning of this output value is the predicted increment of the input power relative to the current power from the current moment to a specific future time range, such as the next 3 to 5 control cycles.

[0056] In a preferred embodiment of the present invention, the Long Short-Term Memory (LSTM) network is pre-trained offline using the following method: A large amount of operational data from the PWM rectifier during its historical operation is collected, covering various typical power fluctuation scenarios (such as sudden gusts of wind speed and cloud cover obstructing photovoltaic power). A large number of training samples are constructed from the historical data. Each training sample contains a sliding window feature matrix as input features, for example, with dimensions (150,4), and the corresponding future actual power change as the ground truth label. Supervised learning is employed, using optimization algorithms such as Adam, with the goal of minimizing the root mean square error (RMSE) between the predicted output and the ground truth label. All weights and bias parameters within the LSTM network are iteratively adjusted using the backpropagation algorithm.

[0057] In a preferred training configuration, the initial learning rate is set to 0.001, and a cosine annealing strategy is used to automatically decay the learning rate as training progresses. The mini-batch size for each training iteration is set to 64 samples, and the total number of training epochs is set to 200. An early stopping mechanism is triggered to prevent overfitting when the model's performance on the independent validation set no longer improves after 20 consecutive epochs. The training set to validation set ratio is 8:2. This training process is performed on a high-performance computing platform (such as a GPU server) until the model's performance on the independent validation set converges and no longer improves. The resulting set of fixed network parameters constitutes the pre-built prediction model, which is then deployed online in the rectifier's real-time control system.

[0058] Optionally, the feedforward compensation component is calculated, specifically based on the algebraic relationship P satisfied by the active power on the AC side of the PWM rectifier and the d-axis current. ac =(3 / 2)*V d *i d The future power change is taken as the change in active power on the AC side, and combined with the d-axis component of the grid voltage obtained from the operating data, according to i ff =(2*ΔP pred ) / (3*V d The feedforward compensation component is calculated; where i ff For the feedforward compensation component, ΔP pred V represents the future change in power. d This represents the d-axis component of the grid voltage.

[0059] Specifically, this embodiment implements the prediction-to-control conversion process, transforming the physical quantity (power) output by the prediction model into an electrical quantity (current) that the control system can execute. In the grid voltage-oriented dq coordinate system, the active power P on the AC side of the PWM rectifier... ac With d-axis current i dThere exists an approximately linear steady-state relationship that can be described by algebraic equations. This is when the change in input power ΔP is predicted. pred To maintain DC-side voltage stability, the control system needs to adjust the d-axis current in advance to generate an equal and opposite power to counteract the disturbance. The required adjustment of the d-axis current is the feedforward compensation component i. ff This can be obtained by linearizing the above power balance equation. The specific calculation formula is as follows:

[0060] i ff =(2*ΔP pred ) / (3*V d );

[0061] Among them, i ff The desired feedforward compensation component is expressed in amperes (A); ΔP pred The future power change output by the LSTM network, in W; V d This represents the d-axis component of the grid voltage, measured in volts, obtained from the current cycle's operating data. The calculation is purely algebraic, involving no integration or differentiation, resulting in a fast response time, limited only by the controller's computational delay, typically on the microsecond level. This allows feedforward compensation commands to quickly apply to the inner current loop after the prediction result is generated, avoiding the inherent response lag of traditional outer voltage loop PI controllers that rely on the integral accumulation of voltage deviation, thus improving the system's dynamic response speed.

[0062] As an example, suppose that during the control cycle, the system controller obtains the current grid voltage d-axis component V from the operating data. d The voltage is 311V, corresponding to an effective value of 220V AC. Simultaneously, the LSTM network predicts that the input power will have a change of ΔP within the next 5ms. pred =+10,000W step increase. At this point, the controller calculates the required feedforward compensation component:

[0063] i ff =(2*10000) / (3*311)≈21.4A;

[0064] The controller will use the calculated 21.4A current command as a feedforward compensation component and add it to the subsequent current reference value synthesis stage to increase the input current of the rectifier in advance and prepare for power growth.

[0065] According to one aspect of this application, the predicted trajectory of the DC bus voltage is calculated as follows: based on the energy conservation relationship of the DC bus capacitor, the future power change is taken as the net power flowing into the DC bus capacitor during the control cycle, and combined with the current measured DC bus voltage obtained from the operating data, as well as the pre-configured control cycle and DC bus capacitor value, according to V dc_pred(k+1)=V dc_actual (k)+(ΔP pred ×T s ) / (C dc ×V dc_actual (k) calculates the predicted voltage value for the next control cycle, which serves as the predicted trajectory of the DC bus voltage; where V dc_pred (k+1) is the predicted voltage value for the next control cycle, V dc_actual (k) represents the current measured DC bus voltage, ΔP pred T represents the future power change. s To control the cycle, C dc This is the DC bus capacitance value.

[0066] Optionally, this embodiment serves as the logical starting point for the feedforward self-correcting closed loop, providing a reference benchmark for evaluating the prediction accuracy of subsequent correction stages. The physical principle is that, within a sufficiently short control cycle, the net power change flowing into or out of the DC bus capacitor will cause a corresponding change in the voltage across the capacitor. When the LSTM network predicts a future power change ΔP... pred Based on this, it can be deduced that, if the prediction is accurate, the DC bus voltage should reach its value at the end of the next control cycle. The calculated theoretical voltage value is the predicted trajectory of the DC bus voltage. This can be achieved using the following formula:

[0067] V dc_pred (k+1)=V dc_actual (k)+(ΔP pred *T s ) / (C dc *V dc_actual (k));

[0068] Among them, V dc_pred (k+1) represents the predicted DC bus voltage for the next control cycle, i.e., the predicted trajectory; V dc_actual (k) represents the measured DC bus voltage obtained from the operating data of the current cycle; ΔP pred T represents the future change in power. s For the pre-configured control cycle, e.g., 50 μs; C dc The capacitance value of the pre-configured DC bus capacitor, for example, 2200μF.

[0069] It is understandable that this formula has sufficiently high accuracy when the control period is much smaller than the system electrical time constant. The V calculated in this embodiment... dc_pred (k+1) is temporarily stored as an intermediate variable for use in the prediction and tracking error calculation stage when the next control cycle arrives.

[0070] Example 4

[0071] As an alternative implementation, the prediction error of power fluctuations is non-stationary. For example, the power surge caused by wind gusts is close to a step signal in the time domain, while the power change caused by clouds slowly shading photovoltaic panels is close to a ramp signal. Predictive models trained for a single mode will produce persistent and systematic prediction biases when switching between fluctuating modes. Correspondingly, traditional control systems have inherent time scale conflicts. The injection rate of feedforward disturbances is once per control cycle, typically on the order of microseconds or sub-milliseconds; while the response time of the voltage outer-loop PI controller responsible for correcting its errors is typically greater than 50ms. This time scale mismatch causes a lag in the feedback loop's adjustment of the feedforward injection error, leading to continuous oscillations in the DC bus voltage.

[0072] As an optional implementation, a prediction correction component is generated based on the deviation between the predicted trajectory and the measured trajectory of the DC bus voltage obtained from the operating data. This further includes: defining the deviation as a prediction tracking error, which is calculated in each control cycle. The prediction tracking error is the difference between the predicted trajectory calculated in the previous control cycle and the measured trajectory sampled in the latest control cycle, and a prediction correction component is generated based on the prediction tracking error.

[0073] Specifically, at the arrival of each control cycle k, the controller performs the following calculation, expressed as follows:

[0074] e pred (k)=V dc_pred (k)-V dc_actual (k);

[0075] Among them, e pred (k) represents the prediction tracking error for the current k-th control cycle; V dc_pred (k) is the calculated and temporarily stored voltage prediction value for the current period k at the end of the previous period, i.e., period k-1; V dc_actual (k) represents the latest measured DC bus voltage sampled in the current period k.

[0076] Accordingly, the predicted tracking error e pred Tracking error V with set value ref -V dc_actual They differ in their generation mechanism and time scale. Setpoint tracking error measures the gap between the control result and the desired target, driving the feedback controller to eliminate the gap and ensure steady-state accuracy.

[0077] Optionally, generating a prediction correction component based on the prediction tracking error includes: inputting the prediction tracking error into a fast correction compensator, the fast correction compensator performing proportional and integral operations on the prediction tracking error, synthesizing and outputting the prediction correction component.

[0078] As an optional implementation, this embodiment constructs a fast time-scale layer in the dual-time-scale correction module to rapidly compensate for transient voltage disturbances caused by single prediction errors at millisecond or even sub-millisecond time scales. Specifically, this fast correction compensator can preferably be implemented using a proportional-integral (PI) controller structure, with its input being the prediction tracking error e. pred The output of this compensator is the predicted correction component i. corr The calculation process is as follows:

[0079] i corr (k)=Kp corr *e pred (k)+I corr (k);

[0080] Among them, i corr (k) represents the prediction correction component for the current period; Kp corr To quickly correct the proportional gain of the compensator; e pred (k) represents the prediction tracking error for the current period; I corr (k) is the integral accumulation term for the current period.

[0081] Integral Cumulative Term I corr The recursive calculation method for (k) is as follows:

[0082] I corr (k)=I corr (k-1)+Ki corr *e pred (k)*T s ;

[0083] Among them, I corr (k-1) is the cumulative integral value of the previous period; Ki corr To quickly correct the integral gain of the compensator; T s To control the cycle. Proportional term Kp corr *e pred (k) serves to provide an immediate response to prediction errors occurring at the current instant, and its response speed is limited only by computational latency. Integral term I corr The role of (k) is to eliminate the steady-state prediction error accumulated by the persistent bias in the output of the LSTM model. For example, when the wind and solar power fluctuation mode changes, causing the predicted value to be systematically higher or lower for a period of time, the integral term can continuously accumulate the systematic bias and generate a continuous correction current to suppress the bias.

[0084] Optionally, the bandwidth of the fast-correction compensator is designed to be 5 to 10 times the bandwidth of the controller used to calculate the feedback adjustment component, ensuring that the fast-correction compensator can complete the correction of the predictive tracking error before the controller used to calculate the feedback adjustment component responds.

[0085] Furthermore, this embodiment further defines the characteristics of the fast correction layer from a performance design perspective. Bandwidth is an indicator of controller response speed. Traditional voltage outer-loop PI controllers used to calculate feedback regulation components typically have their bandwidth limited to a low level to ensure system stability margin; for example, a response time greater than 50ms corresponds to a frequency below 20Hz. However, the fast correction compensator in this embodiment does not aim for the stability of the entire system, but only for the error of the feedforward channel itself, thus allowing for more flexible parameter tuning. Preferably, its bandwidth can be designed to be 100Hz~200Hz, with a response time on the order of 5~10ms.

[0086] Accordingly, this timescale separation ensures functional decoupling of the two control channels: when high-frequency, sudden prediction errors occur, the fast correction layer, with its high bandwidth characteristics, intervenes and completes the correction before the main feedback PI controller even detects the voltage deviation. For low-frequency, steady-state errors, the main feedback PI controller remains responsible for the final elimination.

[0087] In some alternative implementations, the fast-correction compensator can also employ a more complex controller structure, such as a proportional-resonant (PR) controller or a PID controller with a derivative element. When the prediction error contains harmonic components of a specific frequency, a PR controller can achieve zero steady-state error tracking of the error at that frequency. Introducing a derivative element can further predict the trend of error changes and provide a faster dynamic response, but requires proper handling of high-frequency noise.

[0088] According to one aspect of this application, a dynamic confidence factor is determined based on the historical statistical characteristics of the deviation, specifically including: maintaining a sliding statistical window with a pre-configured window length and storing the predicted tracking error calculated for each control cycle into the sliding statistical window; and at each update time of the statistical window, calculating the mean error reflecting the systematic offset and the standard deviation of the error reflecting the degree of random fluctuation, as historical statistical characteristics.

[0089] Optionally, this embodiment constructs a slow time-scale layer in the dual-time-scale correction module to evaluate the overall performance degradation of the prediction model under the current operating conditions from a statistical perspective. Specifically, the system maintains a sliding statistical window of fixed length, the time span of which can be pre-configured to 100-200 milliseconds. When the control system adopts an inner and outer loop separation architecture, if the statistical layer collects error samples at a low frequency (e.g., 1ms / cycle), then a 200-point window corresponds to 200ms; if the statistical layer operates at a high-frequency control cycle (e.g., 50μs / cycle), then to achieve an equivalent statistical depth of 100-200ms, the number of window points should be set to 2000-4000. Those skilled in the art can adjust the number of window points accordingly based on the statistical update rate of the actual control system. In each control cycle, the latest prediction tracking error e pred The data is stored in the window. When the window is updated, that is, every control cycle, the system performs a statistical calculation on all error samples contained in the window (for example, 200 error data points for a 200ms window).

[0090] The calculated statistical indicators include at least two items: one is the arithmetic mean, i.e., the mean error μ. e This reflects the prediction of whether there is a persistent, consistent systematic shift within this time window. The second is the standard deviation, i.e., the standard deviation of error σ. e This reflects the degree of random fluctuation in the prediction results, that is, the stability of the prediction. The above two statistics together constitute a description of the historical statistical characteristics of the prediction tracking error.

[0091] Optionally, determining the dynamic confidence factor based on historical statistical characteristics further includes: adding the absolute value of the mean error to the standard deviation of the error, and normalizing the sum using a pre-configured tolerance error threshold; and using the normalized result as a negative exponent, and calculating the prediction quality index between 0 and 1 through an exponential mapping function.

[0092] In this embodiment, the raw statistics obtained in the previous step are mapped to a performance index with physical meaning. Specifically, the predicted quality index MQ is calculated using the following formula:

[0093] MQ=exp(-(abs(μ e )+σ e ) / e th );

[0094] Where MQ is the desired predictive quality index; exp is the natural exponential function; abs is the absolute value function; μ e and σ e These are the mean and standard deviation of the error, respectively; e thThe pre-configured tolerance threshold is selected based on the range of DC bus voltage fluctuations allowed by the system. For example, it can be the absolute value of the voltage corresponding to 2% of the rated DC voltage.

[0095] Correspondingly, the design of this exponential mapping function exhibits nonlinear characteristics. When the prediction is accurate, both the mean and standard deviation of the error are very small, much smaller than e. th When the absolute value of the exponential term is close to 0, the value of MQ will stabilize at a level close to 1, indicating high prediction quality. However, once the prediction error increases, whether it is a systematic bias (μ) or not, the absolute value of the exponential term is close to 0, and the value of MQ will stabilize at a level close to 1, indicating high prediction quality. e Increase) or random fluctuation (σ) e Increasing the value of the exponent (MQ) will cause the absolute value of the exponent to increase rapidly, and the value of MQ will decay quickly and approach 0, indicating that the prediction quality has deteriorated. Compared with linear mapping, it has strong robustness to small errors and high sensitivity to suppressing significant errors, enabling the system to maintain the stability of the feedforward channel when the prediction performance fluctuates slightly, and to take protective measures when the performance deteriorates.

[0096] Optionally, determining the dynamic confidence factor includes: applying a first-order low-pass filter to the predicted quality index, smoothing the predicted quality index using pre-configured filter coefficients, generating the dynamic confidence factor, and suppressing its abrupt changes caused by disturbances in the results of a single statistical window.

[0097] Building upon this, this embodiment represents the final stage of the slow adaptation layer, aiming to enhance the smoothness and robustness of decision-making and prevent drastic system responses due to single, sporadic prediction errors. Specifically, the final dynamic confidence factor α is updated by applying a first-order low-pass filter to the prediction quality index MQ, and its recursive calculation formula is as follows:

[0098] α(k)=τ*α(k-1)+(1-τ)*MQ(k);

[0099] Where α(k) is the dynamic confidence factor after the current window update; α(k-1) is the value of the previous update; MQ(k) is the prediction quality index calculated for the current window; and τ is the pre-configured filter coefficient, which preferably ranges from 0.8 to 0.95. A larger τ results in a stronger filtering effect, smoother changes in α, and better resistance to interference from occasional prediction error windows, but also a slower perception of the continuous degradation of prediction ability. Conversely, a smaller τ results in a more sensitive response to α.

[0100] In renewable energy hydrogen production scenarios, the switching of power fluctuation modes usually occurs on a time scale of seconds. Therefore, a relatively high filter coefficient is selected to avoid the feedforward channel from exiting prematurely due to brief disturbances, thereby improving the utilization rate of feedforward control while ensuring safety.

[0101] Optionally, the dynamic confidence factor is also used to modulate the integral operation within the fast correction compensator; when the prediction tracking error is integrated and accumulated in the fast correction compensator, the rate of integration accumulation is proportional to the dynamic confidence factor, and the integration accumulation is frozen when the dynamic confidence factor approaches zero, thereby achieving anti-integral saturation.

[0102] As an optional implementation, this embodiment describes a cross-timescale coupling mechanism connecting the slow adaptation layer and the fast correction layer. Specifically, the calculation formula for the integral accumulation term of the fast correction compensator is further optimized as follows:

[0103] I corr (k)=I corr (k-1)+α(k)*Ki corr *e pred (k)*T s ;

[0104] Where α(k) is the dynamic confidence factor most recently updated by the slow adaptation layer.

[0105] By introducing a dynamic confidence factor α(k) to modulate the integral stage, intelligent linkage between the two layers is achieved. When the prediction model is reliable and α(k) is close to 1, the integral stage of the fast correction compensator operates normally, eliminating persistent prediction bias. Conversely, when the slow adaptation layer detects an overall and persistent decline in prediction capability, causing α(k) to gradually approach 0, the rate of integral accumulation is also synchronously controlled.

[0106] The above design can prevent integral saturation problems and avoid the situation where the integral term continues to accumulate indefinitely due to the persistent prediction error, even when the prediction is no longer reliable and the feedforward channel is about to be or has already been suppressed, which would eventually lead to controller output saturation and cause system oscillation.

[0107] To further illustrate the working process of the dual-timescale correction module, a specific numerical calculation example is given below:

[0108] Assuming the current control cycle is k, the system parameters are as follows: Fast correction compensator proportional gain Kp corr =0.05A / V, integral gain Ki corr =500A / (V·s), control period T s =50μs, the statistical layer sampling period is set to 1ms, and error samples are collected once every 20 high-frequency control cycles. The sliding statistical window length is 200 sampling points, corresponding to 200ms, which is consistent with the time span specified above. The allowable error threshold is e. th =14V, 2% of the rated voltage of 700V, low-pass filter coefficient τ=0.9.

[0109] Scenario setting: In the (k-1)th control cycle, the predicted voltage value V dc_pred (k) = 698.5V;

[0110] When the k-th control cycle arrives, the actual sampled DC bus voltage V dc_actual (k) = 695.2V.

[0111] Step 1: Calculate the predicted tracking error e pred (k)=V dc_pred (k)-V dc_actual (k) = 698.5 - 695.2 = 3.3V;

[0112] This positive value indicates that the prediction model overestimates the voltage, meaning that the actual power flowing out of the DC bus is greater than the predicted value.

[0113] Step 2: Quickly correct the compensator output. Assume the previous cycle's integral accumulation term I... corr (k-1)=0.8A, and the current dynamic confidence factor α(k)=0.868. Therefore:

[0114] I corr (k)=I corr (k-1)+α(k)×Ki corr ×e pred (k)×T s =0.8 + 0.868 × 500 × 3.3 × 50 × 10 -6 =0.8 + 0.071 = 0.871A;

[0115] i corr (k)=Kp corr ×e pred (k)+I corr (k)=0.05×3.3+0.871=0.165+0.871=1.036A.

[0116] Step 3: Slow Adaptation Layer Update. Assume that within the statistical window of the most recent 200 control cycles, the mean of the prediction tracking error is μ. e =2.1V, standard deviation σ e =1.8V. Therefore:

[0117] MQ=exp(-(|2.1|+1.8) / 14)=exp(-3.9 / 14)=exp(-0.2786)≈0.757;

[0118] Assuming the dynamic confidence factor α(k-1) was 0.88 in the last update, then:

[0119] α(k)=τ×α(k-1)+(1-τ)×MQ=0.9×0.88+0.1×0.757=0.792+0.076=0.868.

[0120] Step 4: Three-component synthesis. Assume the current feedback adjustment component i... pi =2.5A, feedforward compensation component i ff =21.4A, then:

[0121] i d_ref =i pi +α×(i ff +i corr )=2.5+0.868×(21.4+1.036)=2.5+0.868×22.436=2.5+19.474=21.974A;

[0122] The above calculations show that, under the condition of acceptable prediction accuracy (α≈0.87), the system primarily relies on feedforward compensation, contributing approximately 19.5A, supplemented by feedback regulation, contributing approximately 2.5A, thus achieving rapid and stable control. If the prediction accuracy continues to deteriorate until α approaches 0, then i d_ref Approaching i pi The system automatically reverts to a pure feedback control mode.

[0123] The dual-timescale correction module constructed in this embodiment forms a complete defense system with complementary functions and separated timescales through its internal fast correction layer and slow adaptation layer. The fast correction layer is responsible for transient prediction errors, ensuring the real-time accuracy of feedforward compensation. The slow adaptation layer is responsible for trend-based model performance degradation, determining the overall intervention depth of feedforward control by dynamically adjusting the confidence factor. The dual-timescale correction module is organically combined through a coupling mechanism, ensuring that the entire control system can maintain a control strategy that balances speed and stability across the continuous spectrum from prediction reliability to prediction failure.

[0124] Example 5

[0125] This embodiment integrates the control intentions from three different functional paths—the traditional feedback path, the predictive feedforward path, and the feedforward correction path—into a unified final control command that can be issued to the current inner loop for execution.

[0126] Optionally, the d-axis current reference value is synthesized by combining the feedback adjustment component, the feedforward compensation component, and the prediction correction component, and by weighting the feedforward compensation component and the prediction correction component using a dynamic confidence factor. Specifically, it is calculated using the following formula:

[0127] i d_ref =i pi +α*(i ff +i corr).

[0128] In this embodiment, the final d-axis current reference value i d_ref It consists of three core components. The first component is the feedback regulation component i. pi The first component, generated by a traditional voltage outer-loop PI controller, is calculated solely based on the difference between the setpoint and the measured value. It primarily handles the system's steady-state accuracy and the final correction for unmodeled disturbances. The second and third components are the feedforward compensation components i... ff and predicted correction component i corr Together, they form a feedforward control channel with predictive and self-correcting capabilities. Before being superimposed, these two components are multiplied by a dynamic confidence factor α calculated by the slow adaptation layer.

[0129] The final synthesis formula is as follows:

[0130] i d_ref =i pi +α*(i ff +i corr );

[0131] Among them, i d_ref The synthesized d-axis current reference value will be used as the input command for the inner current loop controller; i pi For feedback adjustment components; α is the dynamic confidence factor; i ff For feedforward compensation components; i corr To predict the correction component.

[0132] The adaptive and robust nature inherent in this synthetic architecture enables smooth and automatic switching of control strategies through a dynamic confidence factor α. Specifically, it can analyze two typical operating conditions:

[0133] Operating Condition 1: Renewable energy power fluctuations are stable, or the fluctuation pattern matches the training data of the prediction model. Under this condition, the LSTM model has very high prediction accuracy, leading to a prediction tracking error e. pred It remains at a low level. Correspondingly, the dynamic confidence factor α calculated by the slow adaptation layer will approach 1. At this point, the synthesis formula approximately degenerates into i. d_ref ≈i pi +i ff +i corr Because the prediction was accurate, i ff It can cancel out power disturbances in advance, and i corr Because of input e pred The output is also close to zero. The DC bus voltage hardly fluctuates, resulting in minimal feedback regulation component i. pi It is also close to zero. Therefore, the control action of the system is almost entirely determined by the feedforward component i. ffIt dominates and achieves the suppression of disturbances, with the dynamic response speed reaching the optimal level.

[0134] Operating Condition 2: Abrupt changes occur in the power fluctuation pattern of renewable energy sources, such as a sudden shift from stable, cloudless sunshine to gusty winds, causing the prediction model to fail for a period of time. Under this condition, the model's predicted output deviates significantly from the actual physical process, resulting in a prediction tracking error e. pred The level remains consistently high. Correspondingly, the slow-adaptation layer rapidly reduces the dynamic confidence factor α to near zero. At this point, the synthesis formula approximately degenerates into i. d_ref ≈i pi Feedforward channel i ff and fast correction channel i corr Although the output may be large, its influence on the final control command is suppressed because its weight α approaches zero. The entire control system automatically and smoothly switches back to the mode relying solely on PI feedback control. Although this mode has a slower response speed, its stability has been verified through long-term engineering, ensuring that even in the worst-case scenario of predictive function failure, the risk of system oscillation or instability is reduced, guaranteeing basic operational safety.

[0135] This three-component dynamic weighted synthesis architecture is not merely a simple arithmetic superposition of three control signals, but also an intelligent control allocation mechanism. The control method in this embodiment can maintain system robustness while pursuing rapid dynamic response, adapting to practical application scenarios where renewable energy power prediction accuracy fluctuates.

[0136] In some optional implementations, the system also includes the following initialization and protection strategies:

[0137] During system power-on startup, the outputs of both the prediction feedforward channel and the correction channel are unreliable because the sliding window buffer of the LSTM prediction model is not yet filled with valid historical feature vectors, and the sliding statistical window of the slow adaptation layer has not accumulated enough prediction tracking error samples. Therefore, during initialization, the dynamic confidence factor α is forcibly set to 0, and the system operates in pure feedback control mode. Only after the sliding window buffer is filled (e.g., after running N=150 control cycles) and the statistical window has undergone at least one complete update, is the system allowed to allow the dynamic confidence factor to evolve automatically based on actual prediction performance.

[0138] In addition, to prevent the LSTM model from outputting outliers under extreme conditions, feedforward compensation can be applied to the i-th component. ff and predicted correction component i corr Output clamping and limiting are set on each of the above, and the limiting value can be set to ±50% of the rated d-axis current. If the final synthesized d-axis current reference value i... d_refIf the current exceeds the rectifier's allowable operating current range, the system will automatically trigger the limiting protection and set the dynamic confidence factor α to zero, thus switching the system to a safe feedback control mode.

[0139] Example 6

[0140] To conduct simulation verification, in this embodiment, a simulation model of a wind-solar-hydrogen production coupled system, including a three-phase power grid, a PWM rectifier, a DC bus capacitor, and an equivalent downstream hydrogen production load, was built on the Matlab / Simulink software platform. The system parameters are set as follows: the grid line voltage is 380V, the frequency is 50Hz; the rated power of the PWM rectifier is 100kW; and the DC bus capacitor C... dc Set to 2200μF; Rated DC bus voltage setting value V ref The voltage is 700V. The switching frequency and sampling frequency of the control system are both set to 20kHz.

[0141] To simulate the power fluctuations of renewable energy, this embodiment designs extreme test conditions: after the simulation starts, the system operates stably at 30% of the rated power, i.e., 30 kW; at t=0.5 seconds, the input power linearly rises to 80% of the rated power, i.e., 80 kW, within 0.1 seconds, and the power change rate reaches 500% / s of the rated power, which is far beyond the level that can usually be stably handled.

[0142] For this test condition, such as Figure 4 As shown, this embodiment compares three different control strategies:

[0143] Strategy A: The traditional dual-closed-loop PI control method, which does not contain any feedforward channels.

[0144] Strategy B: Feedforward control based on ideal prediction, i.e., assuming LSTM prediction is accurate, ΔP pred It is equal to the actual power change and uses open-loop feedforward compensation, but does not include the dual time-scale correction module of this invention.

[0145] Strategy C: The control method proposed in this invention includes an LSTM prediction module and a dual time-scale correction module.

[0146] Simulation results show that:

[0147] Under Strategy A and traditional PI control, when power increases, the DC bus voltage dips due to the limited bandwidth of the outer voltage loop and delayed response, reaching a minimum of approximately 610V, with a voltage fluctuation amplitude of 12.8%. Simultaneously, this voltage fluctuation is transmitted to the AC side through the control loop, causing distortion of the AC input current, with the total harmonic distortion (THD) peaking at 8.5% during the fluctuation period.

[0148] Under Strategy B and the ideal open-loop feedforward, the dynamic response of the system is improved due to the introduction of feedforward compensation. However, due to factors such as slight mismatch in model parameters and system nonlinearity, even accurate power prediction cannot achieve compensation. The DC bus voltage still experiences a transient drop of approximately 655V, with a fluctuation range of about 6.4%. In other words, while relying solely on open-loop feedforward can improve performance, it still cannot eliminate coupled oscillations.

[0149] Under strategy C and the method of this invention, the system exhibits optimal control performance. When the power begins to climb, the LSTM prediction module anticipates the power change trend and generates a feedforward compensation component i. ff During the power ramp-up process, the dual-time-scale correction module compares the predicted trajectory with the measured trajectory in real time, and quickly corrects component i. corr The system compensates for minor biases in the predictions and model mismatches. Throughout the process, the lowest point of the DC bus voltage only dropped to approximately 685V, and the maximum voltage fluctuation was suppressed to within 2.1%. Simultaneously, due to the stability of the DC side, the AC side current waveform was smooth, and its THD remained below 2.5% throughout the dynamic process, ensuring power quality.

[0150] The simulation comparison experiments demonstrated that the proposed dual-time-scale feedforward decoupling control method based on prediction error can reduce the DC bus voltage fluctuation amplitude under power fluctuations and improve the AC side current quality. This method addresses the problem of strong AC-DC coupling oscillations under renewable energy fluctuations, enhancing the stability and robustness of the hydrogen production system under grid disturbances.

[0151] This invention constructs a prediction error-driven dual-timescale closed-loop correction architecture to address the problem that traditional feedforward control cannot handle model inaccuracies due to its open-loop characteristics. This architecture calculates the real-time accuracy of the prediction model by predicting and tracking the error signal. Based on this, a high-bandwidth fast correction layer is responsible for instantly offsetting transient prediction deviations within milliseconds; simultaneously, a slow adaptation layer performs statistical evaluation of the prediction error on a hundred-millisecond timescale, generating a dynamic confidence factor. This factor can smoothly adjust the intervention depth of the entire feedforward channel, leveraging its speed when predictions are accurate, and automatically suppressing its effect when predictions are inaccurate, causing the system to degenerate into a stable feedback control mode.

[0152] Meanwhile, through the organic combination of the aforementioned prediction, feedforward, and dual-time-scale self-correction, this invention upgrades the open-loop feedforward compensation into a closed-loop intelligent feedforward system with self-verification and self-adjustment capabilities. By stabilizing the DC bus voltage and cutting off the feedback path that triggers strong AC-DC coupling oscillations, this invention achieves a comprehensive technical effect: suppressing DC voltage fluctuations and ensuring AC power quality under renewable energy power fluctuations.

[0153] In the embodiments of this invention, the parameter values, process steps, etc., given are merely exemplary implementation methods and are not intended to limit the scope of protection of this invention. Based on the core technical logic of this invention, the parameters and implementation methods can be equivalently replaced and adjusted according to the actual application scenario.

[0154] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0155] The embodiments or examples disclosed herein are not exhaustive, but merely illustrative of some embodiments or examples, and are not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment or example can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment or example can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment or example can be arbitrarily interchanged. Furthermore, optional methods or examples in a particular embodiment or example can be arbitrarily combined; moreover, embodiments or examples can be arbitrarily combined. For example, some or all steps of different embodiments or examples can be arbitrarily combined, and a particular embodiment or example can be arbitrarily combined with optional methods or examples of other embodiments or examples.

[0156] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0157] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for power feedforward decoupling control of a PWM rectifier for hydrogen electrolysis under renewable energy fluctuations, characterized in that, include: Obtain the operating data of the PWM rectifier; Based on operational data, the future power change is determined using a pre-built prediction model, and the feedforward compensation component is calculated. The feedback adjustment component is calculated based on the difference between the pre-configured set value of the DC bus voltage and the measured DC bus voltage obtained from sampling. Based on future power changes, the predicted trajectory of DC bus voltage is calculated. Based on the deviation between the predicted trajectory and the measured trajectory of the DC bus voltage obtained from the operating data, a prediction correction component is generated, the deviation is accumulated in multiple control cycles, and a dynamic confidence factor for adjusting the intervention depth of the feedforward path is determined based on the historical statistical characteristics of the deviation. By combining the feedback adjustment component, the feedforward compensation component, and the prediction correction component, and by using a dynamic confidence factor to weight the feedforward compensation component and the prediction correction component, a reference value for the d-axis current is synthesized.

2. The method according to claim 1, characterized in that, Determining future power changes using a pre-built prediction model includes: Construct a sliding window feature matrix with time steps as the sequence based on the runtime data; The sliding window feature matrix is ​​input into a pre-built prediction model, where the prediction model is a long short-term memory network; The sliding window feature matrix is ​​processed by a long short-term memory network along the time step sequence to output the future power change.

3. The method according to claim 1, characterized in that, The feedforward compensation component is calculated as follows: Based on the algebraic relationship between the active power on the AC side of the PWM rectifier and the d-axis current, P ac =(3 / 2)*V d *i d The future power change is taken as the change in active power on the AC side, and combined with the d-axis component of the grid voltage obtained from the operating data, according to i ff =(2*ΔP pred ) / (3*V d The feedforward compensation component is calculated. Among them, i ff For the feedforward compensation component, ΔP pred V represents the future change in power. d This represents the d-axis component of the grid voltage.

4. The method according to claim 1, characterized in that, The predicted trajectory of the DC bus voltage is calculated as follows: Based on the energy conservation principle of the DC bus capacitor, the future power change is taken as the net power flowing into the DC bus capacitor during the control cycle. This is combined with the current measured DC bus voltage obtained from operating data, the pre-configured control cycle, and the DC bus capacitor value, according to V... dc_pred (k+1)=V dc_actual (k)+(ΔP pred ×T s ) / (C dc ×V dc_actual (k) Calculate the predicted voltage value for the next control cycle, which serves as the predicted trajectory of the DC bus voltage; Among them, V dc_pred (k+1) is the predicted voltage value for the next control cycle, V dc_actual (k) represents the current measured DC bus voltage, ΔP pred T represents the future power change. s To control the cycle, C dc This is the DC bus capacitance value.

5. The method according to claim 1, characterized in that, Based on the deviation between the predicted trajectory and the measured trajectory of the DC bus voltage obtained from the operational data, a prediction correction component is generated, further including: The deviation is defined as the prediction tracking error. The prediction tracking error is calculated in each control cycle. Its value is the difference between the predicted trajectory calculated in the previous control cycle and the latest sampled measured trajectory in the current control cycle. A prediction correction component is generated based on the prediction tracking error.

6. The method according to claim 5, characterized in that, Based on the prediction tracking error, a prediction correction component is generated, specifically including: The predicted tracking error is input into the fast correction compensator, which performs proportional and integral operations on the predicted tracking error, synthesizes and outputs the predicted correction component.

7. The method according to claim 5, characterized in that, The dynamic confidence factor is determined based on the historical statistical characteristics of the bias, specifically including: Maintain a sliding statistical window with a pre-configured window length, and store the prediction tracking error calculated for each control cycle into the sliding statistical window; At each update time of the statistical window, all prediction tracking error samples stored in the sliding statistical window are calculated to obtain the mean error reflecting the systematic shift and the standard deviation error reflecting the degree of random fluctuation, which are used as historical statistical characteristics.

8. The method according to claim 7, characterized in that, Determining dynamic confidence factors based on historical statistical characteristics also includes: The absolute value of the mean error is added to the standard deviation of the error, and the sum is normalized using a pre-configured tolerance error threshold. The result of the normalization process is used as a negative exponent, and the prediction quality index between 0 and 1 is obtained by using the exponential mapping function.

9. The method according to claim 8, characterized in that, Determine the dynamic confidence factor, including: A first-order low-pass filter is applied to the predicted quality index, and the predicted quality index is smoothed using pre-configured filter coefficients to generate a dynamic confidence factor, thereby suppressing its abrupt changes caused by disturbances in the results of a single statistical window.

10. The method according to claim 1, characterized in that, Obtain the operating data of the PWM rectifier, specifically including: The three-phase AC voltage and three-phase AC current are sampled in each control cycle and projected onto the synchronous rotating coordinate system through Park transformation to extract the d-axis voltage component, d-axis current component and q-axis current component. The d-axis voltage component, d-axis current component, q-axis current component, and synchronously sampled measured DC bus voltage are normalized and concatenated into a feature vector, which is used as the running data.