Converter and flexible predictive control method thereof, computer storage medium

Abstract

This application discloses a flexible predictive control method for a converter, comprising: constructing an output current increment prediction model, which includes an inherent dynamic term reflecting the mapping relationship between historical data of the converter system and the output current increment, and an attack term reflecting external anomalies; constructing a criterion function, calculating the gradient of the criterion function with respect to the coefficient matrix, and setting the gradient to zero to obtain an estimate of the coefficient matrix at the current time; using the error between the reference current of the next cycle and the actual output current of the current cycle as the target output of the output current increment prediction model, rewriting the output current increment prediction model, and obtaining a control law for the voltage increment; based on the control law for the voltage increment, substituting the coefficient matrix estimate at the current time to obtain the voltage control increment, and selecting the converter switching state based on the voltage control increment. This method can achieve parameter-independent predictive control of the converter, resist spurious data attacks, reduce switching frequency, and improve system robustness.

Description

Technical Field

[0001] This application relates to the field of power electronics technology, and in particular to a converter and its flexible predictive control method, and a computer storage medium. Background Technology

[0002] With the continuous expansion of the scale of new energy grid connection, the control performance of converters, as the core interface, directly determines the stability and power quality of the power system. Among them, Finite Set Model Predictive Control (FCS-MPC) has become a research hotspot due to its superior dynamic performance. However, this method is highly dependent on accurate mathematical models and faces the problem of reduced control robustness due to parameter mismatch in practical applications. At the same time, FCS-MPC does not consider switching losses during the optimization process, which directly leads to high-frequency operation of switching devices and a sharp increase in losses. In addition, in the digital interconnection environment, converters are extremely vulnerable to spoofed data injection attacks, and traditional control strategies lack effective defense mechanisms, posing a serious threat to system security. Summary of the Invention

[0003] This application provides a converter and its elastic predictive control method, as well as a computer storage medium. The method can achieve completely parameter-independent predictive control of the converter, reduce the switching frequency and improve system robustness while resisting false data injection attacks.

[0004] Firstly, this application provides a flexible predictive control method for a converter, the flexible predictive control method comprising:

[0005] An output current increment prediction model is constructed, which includes an inherent dynamic term reflecting the mapping relationship between historical data of the converter system and output current increment, and an attack term reflecting external anomalies.

[0006] A criterion function is constructed, which includes a tracking term that characterizes the error between the actual output current increment and the predicted output current increment, and a penalty term that characterizes the change in the coefficient matrix of the inherent dynamic term. The predicted output current increment is obtained through the output current increment prediction model.

[0007] Calculate the gradient of the criterion function with respect to the coefficient matrix, and set the gradient to zero to obtain an estimate of the coefficient matrix at the current time.

[0008] The error between the reference current of the next cycle and the actual output current of the current cycle is used as the target output of the output current increment prediction model. The output current increment prediction model is then rewritten to obtain the control law for the voltage increment.

[0009] Based on the control law of voltage increment, the coefficient matrix of the current moment is substituted to estimate the voltage control increment, and the switching state of the converter is selected based on the voltage control increment.

[0010] In one embodiment, the resilient predictive control method further includes the following steps:

[0011] A neural network model is constructed to establish a mapping relationship between the attack items and the system state, where the system state is the output current of the converter;

[0012] Design a state estimator that combines the output of the neural network model with the output of the output current increment prediction model as the prediction term, and introduces a feedback correction term based on the output current estimation error to achieve system state tracking. The output current estimation error is the difference between the output current sample value and the system state estimate output by the state observer.

[0013] Based on the state estimator, the update law of the weight coefficient matrix of the neural network model is configured according to the stability criterion, so that the output current estimation error converges to zero under the action of the update law. The update law includes a gradient descent term related to the output current estimation error.

[0014] Based on the output current estimation error of the current cycle, the weight coefficient matrix is ​​updated using the update law. Then, using the updated weight coefficient matrix and the current system state, the attack term estimate for the next control cycle is calculated.

[0015] In one embodiment, the control method further includes:

[0016] Based on the preset control sampling period, the state observer and update law are discretized using the first-order Euler method to generate a discretized prediction model of the attack term. Based on the output current estimation error of the current period, the estimated value of the attack term for the next control period is calculated using the discretized prediction model.

[0017] In one embodiment, the control method further includes:

[0018] The voltage control increment is compensated based on the attack term estimate.

[0019] In one embodiment, the system state vector is constructed using the historical output current sequence of the converter, and the neural network model adopts a radial basis function neural network model. The activation function of the neural network model is a Gaussian radial basis function with the system state vector as the variable.

[0020] In one embodiment, an event-triggered control voltage control increment update is employed, and the event-triggered control mechanism includes:

[0021] Define event triggering error Event triggering error Current tracking error Tracking error compared to the last trigger time The difference between the reference output current and the actual output current is the tracking error.

[0022] The voltage control increment is triggered to update when the following conditions are met:

[0023] ;

[0024] In the formula, This indicates a preset safety threshold. , This represents the tracking error at time k. The prediction residual error under ideal control, This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. Represents the gain coefficient vector. , This represents the scaling factor for positive constants. This represents the regularization constant.

[0025] In one embodiment, the control law for the voltage increment is expressed by the following formula:

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, Represents the gain coefficient vector. This represents the scaling factor for positive constants. Represents the regularization constant. Represents the historical influence vector. This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. Indicates the attack item. This indicates the voltage increment.

[0030] Secondly, this application also provides a converter employing the elastic predictive control method described in the first aspect.

[0031] In one embodiment, the converter is any one of the following types of converters: two-level converter, two-level converter, modular multilevel converter.

[0032] Thirdly, this application also provides a computer storage medium storing a processing program, which executes the elastic predictive control method as described in the first aspect when executed.

[0033] The aforementioned resilient predictive control method for converters constructs an output current increment prediction model that includes inherent dynamic terms and attack terms. It then constructs a criterion function consisting of a tracking term and a coefficient matrix change penalty term. By solving for the gradient zeros of the criterion function, the coefficient matrix is ​​estimated. Based on the tracking error between the reference current and the actual current, the model is rewritten to derive the voltage increment control law. Finally, the voltage control increment is obtained by substituting the coefficient matrix estimate and used to select the converter switching state. This method achieves completely parameter-independent output current increment predictive control, improving system robustness. Furthermore, the output current increment prediction model constructed in the control method includes attack terms, effectively addressing the severe condition of false information injection attacks at the inverter level, significantly enhancing the resilience of the control method. Attached Figure Description

[0034] Figure 1 A flowchart of a resilient predictive control method for a converter in one embodiment;

[0035] Figure 2 This is a flowchart illustrating the calculation of attack term estimates for the next control cycle in one embodiment.

[0036] Figure 3 Here is a system block diagram of a flowmeter elastic predictive control method in one embodiment;

[0037] Figure 4 The current waveform output by the traditional FCS-MPC method under parameter matching conditions in one embodiment is shown.

[0038] Figure 5 The current waveform diagram is shown in one embodiment of the elastic predictive control method provided in this application under parameter matching conditions.

[0039] Figure 6 The current waveform output by the traditional FCS-MPC method under parameter mismatch conditions in one embodiment is shown.

[0040] Figure 7 The current waveform diagram is shown in one embodiment of the elastic predictive control method provided in this application under parameter mismatch conditions.

[0041] Figure 8 The current waveform output by the conventional FCS-MPC method under the worst test conditions in one embodiment is shown.

[0042] Figure 9 The current waveform diagram output by the elastic predictive control method provided in this application is shown for the worst test condition in one embodiment. Detailed Implementation

[0043] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0045] In one embodiment, such as Figure 1 As shown, this application provides a flexible predictive control method for a converter, which includes the following steps:

[0046] Step 101: Construct an output current increment prediction model. The output current increment prediction model includes an inherent dynamic term that reflects the mapping relationship between the historical data of the converter system and the output current increment, and an attack term that reflects external anomalies.

[0047] Specifically, considering the potential cyberattacks that the controlled system (such as the inverter) may suffer, the output current increment prediction model provided in this application includes an inherent dynamic term that reflects the mapping relationship between historical data of the converter system and the output current increment. and attack items reflecting external anomalies For example, considering the injected attack information, the constructed output current increment prediction model is expressed as follows:

[0048] .

[0049] Inherent dynamic terms It can reflect the inherent dynamic term that reflects the mapping relationship between the historical data of the converter system and the output current increment. The inherent dynamic term is completely driven by the historical system data of the converter.

[0050] In the formula, the historical data matrix ,in, The output current increment at adjacent sampling times; This represents the input voltage increment between adjacent sampling times; This indicates the output information of the converter system. It can be decided in the historical data vector This includes the output current increment over a number of consecutive time periods in the past. ; Indicates the dimension of system input information. It can be decided in the historical data vector This includes the input voltage increment over a number of consecutive time periods in the past. Coefficient matrix , including built-in information Let be the coefficient vector to be estimated.

[0051] Attack item It can be used to isolate network attacks such as spoofed data injection. The attack term reflects the anomalous portion of the output current increment prediction model caused by the network attack that cannot be explained by the inherent dynamics. It should be noted that the attack term is not a directly measured variable, but rather needs to be estimated and separated in real time through an online learning neural network.

[0052] Step 102: Construct a criterion function. The criterion function includes a tracking term that characterizes the error between the actual output current increment and the predicted output current increment, and a penalty term that characterizes the change in the coefficient matrix of the inherent dynamic term. The predicted output current increment is obtained through the output current increment prediction model.

[0053] Specifically, the criterion function provided in this application includes tracking items and penalty items.

[0054] The tracking term represents the actual output current increment. Error between the predicted and actual output current increments. Wherein, the actual output current increment... The increment can be obtained by sampling the actual current and calculating the increment. The predicted output current increment is calculated using an output current increment prediction model. The predicted value of the current increment at time step is obtained, i.e. In this embodiment, the tracking term is configured as the actual output current increment. The square of the L2 norm of the difference between the predicted output current increment and the actual output current increment, i.e., the tracking term, can be expressed as: Minimizing the tracking term allows the output current increment of the output current increment prediction model to approximate the actual output of the system, thereby ensuring prediction accuracy.

[0055] The penalty term represents the coefficient matrix of the intrinsic dynamic term. The change in quantity. For example, the penalty term can be expressed as: , where μ is a positive penalty coefficient. This penalty term is the coefficient matrix of the intrinsic dynamic term. Compared to the estimated value obtained at the previous time step The square of the L2 norm of the difference. The purpose of introducing the penalty term is to constrain drastic changes in the coefficient matrix, so as to ensure the smoothness and stability of the parameter estimation process, prevent drastic fluctuations during parameter updates, and ensure the stability of the estimation.

[0056] Therefore, the criterion function expression constructed in the embodiments of this application can be obtained:

[0057] .

[0058] Step 103: Calculate the gradient of the criterion function with respect to the coefficient matrix, and set the gradient to zero to obtain an estimate of the coefficient matrix at the current time.

[0059] Specifically, for the criterion function coefficient matrix Calculate the gradient. To obtain the optimal estimate, set the gradient to zero, i.e., set... By solving this equation, the coefficient matrix at the current time step can be obtained. The estimated value:

[0060] .

[0061] Current moment Estimation of the coefficient matrix It will be used for the prediction and control of the system input voltage at the next moment.

[0062] Step 104: Use the error between the reference current of the next cycle and the actual output current of the current cycle as the target output of the output current increment prediction model, rewrite the output current increment prediction model, and obtain the control law of voltage increment.

[0063] Specifically, the output current increment prediction model The structure is rewritten, that is, the control quantity to be determined at the current moment is the voltage increment. From historical data vectors The output current increment prediction model is then explicitly separated. Therefore, the rewritten model is as follows:

[0064] .

[0065] in, Represents the coefficient matrix estimated at time k. The coefficient vector of the (Ly+1)th term; Indicates voltage increment; Represents the historical influence vector. , and Both are coefficient matrices estimated at time k. Elements in; This is the current attack item.

[0066] Furthermore, the reference current for the next cycle... Compared with the actual output current of the current cycle The error between them is used as the target output of the output current increment prediction model, that is, let ,Will Substitute into the rewritten output current increment prediction model and solve. The control law for the voltage increment can then be obtained:

[0067] ;

[0068] Wherein, the gain coefficient vector is In the formula, This is the step size factor, used to control the parameter update rate. It is a very small positive number, used to prevent the denominator from being 0.

[0069] Step 105: Based on the control law of voltage increment, substitute the coefficient matrix of the current moment to obtain the voltage control increment, and select the converter switching state based on the voltage control increment.

[0070] Specifically, the voltage increment control law is as follows: This will yield an estimate of the coefficient matrix at the current time. Substitute the values ​​into the voltage increment control law to calculate the gain coefficient in the voltage increment control law. and historical influence vector Combined with the given reference current Actual current and attack items Finally, the voltage control increment for the current cycle is calculated. .

[0071] Furthermore, based on voltage control increment Select the converter switching state. Then calculate the... As the optimization objective, switching states are selected within the framework of Finite Set Model Predictive Control (FCS-MPC). The basic voltage vectors corresponding to all possible switching states of the controlled converter are traversed. For example, for a three-level neutral-point clamped (3L-NPC) converter, its 27 basic voltage vectors need to be traversed. For each candidate voltage vector, the value function is optimized (typically minimizing the output vector and voltage control increment). The evaluation is conducted with the error between the two values ​​as the objective. Ultimately, the value function is selected to best approximate the actual output to the voltage control increment. The switching state corresponding to the voltage vector of the converter acts on the power devices of the converter.

[0072] The aforementioned elastic predictive control method, in each control cycle, is no longer limited to filtering and outputting a single voltage vector, but is based on the same voltage control increment described above. Given a reference target, two or three voltage vectors are selected and vector synthesized to generate a synthesized virtual voltage vector. The switching state sequence corresponding to this virtual vector is then applied to the converter, thereby effectively improving the waveform quality of the output current while maintaining complete model-free operation and resisting network attacks.

[0073] In this embodiment, an output current increment prediction model incorporating inherent dynamic terms and attack terms is constructed. A criterion function, including a tracking term and a coefficient matrix change penalty term, is then constructed. The coefficient matrix is ​​estimated by solving for the gradient zeros of the criterion function. The output current increment prediction model is then rewritten based on the tracking error between the reference current and the actual current to derive the voltage increment control law. Finally, the voltage control increment is obtained by substituting the coefficient matrix estimate and selecting the converter switching state accordingly. This method achieves output current increment prediction control completely independent of system parameters, utilizing only the input and output data of the controlled system. Even under adverse conditions such as completely unknown system parameters and network attacks, it can still operate effectively, significantly improving the robustness and security of the converter under parameter mismatch and network attacks. Furthermore, the output current increment prediction model constructed in the control method includes an attack term, which can effectively cope with the adverse condition of false information injection attacks at the inverter level, significantly improving the flexibility of the control method.

[0074] In one embodiment, such as Figure 2 As shown, calculating the attack term estimate for the next control cycle includes the following steps:

[0075] Step 201: Construct a neural network model to establish a mapping relationship between the attack item and the system state, where the system state is the output current of the converter;

[0076] Specifically, the neural network model adopts a radial basis function structure, and the expression of the neural network model is as follows:

[0077] ;

[0078] in, This is a weight coefficient matrix of dimension j×2; The activation function is Gaussian, where the output of the m-th neuron is... , For the first The center vector of each neuron; The width of the basis functions; To minimize estimation error, this neural network model is established based on system state (output current of the converter). ) to attack item Nonlinear mapping.

[0079] Based on the above neural network model, the state of the controlled system (defined as the output current of the converter in this embodiment) can be expressed as: .

[0080] in, This represents the output predicted by the output current increment prediction model.

[0081] Step 202: Design a state estimator. The state estimator superimposes the output of the neural network model and the output of the output current increment prediction model as the prediction term, and introduces a feedback correction term based on the output current estimation error to achieve system state tracking. The output current estimation error is the difference between the output current sample value and the system state estimate output by the state observer.

[0082] Specifically, the state estimator provided in this application includes: a prediction term and a tracking term.

[0083] In this embodiment, the output of the neural network model is... Output of the output current increment prediction model Linear superposition as a prediction term can be expressed as: .

[0084] in, Mappings derived from neural networks are used for fitting attacks; The output is derived from the output current increment prediction model and is used to characterize the inherent dynamics of the system.

[0085] This application embodiment also defines a feedback term based on the output current estimation error. For example, the feedback term can be represented as: .in, It is a positive constant gain. Output current estimation error. The system state estimate is defined as the output of the state estimator. With output current sampling value The difference, that is = Introduce a feedback correction term. The purpose is to achieve system status tracking.

[0086] Based on the above description, the expression for the state estimator provided in this application embodiment can be obtained as follows: .

[0087] Step 203: Based on the state estimator, configure the update law of the weight coefficient matrix of the neural network model according to the stability criterion, so that the output current estimation error converges to zero under the action of the update law. The update law includes a gradient descent term related to the output current estimation error.

[0088] Specifically, based on the state estimator To force the output current estimation error To converge to zero, the weight coefficient matrix of the neural network needs to be derived and configured based on stability criteria such as Lyapunov stability. The update law of the weight coefficient matrix. The continuous-time form of the update law of the weight coefficient matrix is:

[0089] .

[0090] The update law of the weight coefficient matrix includes the key gradient descent term. The gradient descent term With output current estimation error Directly related, its function is to use the output current estimation error information for backpropagation to adjust the weighting coefficient matrix in real time. This drives the output of the neural network model to approximate the actual attack term. Another term in the update law is -γ2. This is a damping term used to ensure the stability of the parameter estimation process.

[0091] Step 204: Based on the output current estimation error of the current cycle, update the weight coefficient matrix using the update law. Using the updated weight coefficient matrix and the current system state, calculate the attack term estimate for the next control cycle.

[0092] Specifically, obtain the output current estimation error for the current cycle. Current estimation error The system state estimate output by the state estimator With output current sampling value The difference, that is = .

[0093] Furthermore, the weighting coefficient matrix is ​​updated using the update law of the weighting coefficient matrix, which is the current cycle output current estimation error. With the current system status Substituting the values ​​into the update law of the weight coefficient matrix, we can calculate the updated weight coefficient matrix for the next cycle. .

[0094] The updated weight coefficient matrix The system state at the next moment Substituting it into the neural network model, the result can be calculated. Estimated value of the attack term at time .

[0095] In this embodiment, online neural network technology is innovatively introduced to achieve real-time extraction and separation of attack information. This enables real-time identification and separation of network attacks such as spoofed data injection at the inverter level, providing the converter with accurate disturbance feedforward compensation information. Thus, without compromising stability, it enhances the proactive defense capability and anti-interference resilience against malicious network attacks, and improves the flexibility of the control method. Based on the estimated value of this attack item... This allows for the modification of the aforementioned voltage increment control law, resulting in a more precise voltage control increment, which is beneficial for the accurate control of the converter.

[0096] In one embodiment, based on a preset control sampling period, the state observer and the update law of the weight coefficient matrix are discretized using the first-order Euler method to generate a discretized prediction model of the attack term. Based on the output current estimation error of the current period, the estimated value of the attack term for the next control period is calculated using the discretized prediction model. This discretization process is beneficial for practical engineering applications.

[0097] Specifically, based on a preset control sampling period T s State estimator for continuous time domain Update law of weight coefficient matrix The first-order Euler method is used for discretization. The result of this discretization process is the discretized prediction model of the attack term:

[0098] .

[0099] Obtain the output current estimation error for the current cycle = The current cycle output current estimation error. System state in the current cycle Substituting into the update law of the discretized weight coefficient matrix, the update law of the weight coefficient matrix for the next control cycle is calculated. .

[0100] Update law using the updated weight coefficient matrix The system state at the next moment The calculation is performed using a discretized attack item calculation model, and the output is... Estimated value of the attack term at time .

[0101] In one embodiment, voltage control increments are compensated based on attack term estimates.

[0102] Specifically, based on the accurate estimation of attack items using a neural network model, the estimated values ​​of the attack items are... Substituting this back into the control law of the voltage increment, an accurate voltage control increment is obtained that eliminates interference from attack terms. This gives the voltage command acting on the controlled object inherent anti-attack capability, improving accuracy and robustness against cyberattacks. Furthermore, based on the neural network model proposed in this application, "spoofing attacks" targeting converter sensor measurements or control commands can be accurately predicted, achieving effective protection against cyberattacks at the inverter level.

[0103] After obtaining the voltage control increment, the system state can be predicted based on the actual current using a discretized state estimator (i.e., ...). Furthermore, the output current estimation error for the next control cycle can be calculated based on the predicted system state. And estimate the error based on the output current of the next control cycle. Iterative update of weight coefficient matrix and the estimated value of the attack item This enables online learning and optimization of neural network models, ultimately improving the accuracy of voltage control increments.

[0104] In one embodiment, an event-triggered control voltage is used to control the incremental updates.

[0105] Specifically, in this embodiment, the converter controls the increment based on the latest voltage only when the triggering constraint condition is met. Calculate and update the switching state; otherwise, the converter will retain the switching state at the previous trigger moment, which can effectively reduce switching losses.

[0106] As an optional implementation, embodiments of this application provide an event-triggered mechanism that integrates into a data-driven structure, enabling event-triggered control without using any parameters within the controlled system. The event-triggered control mechanism includes:

[0107] Define event triggering error Event triggering error Current tracking error Tracking error compared to the last trigger time The difference between the reference output current and the actual output current is the tracking error.

[0108] The voltage control increment is triggered to update when the following conditions are met:

[0109]

[0110] In the formula, This indicates a preset safety threshold. , This represents the tracking error at time k. The prediction residual error under ideal control, This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. Represents the gain coefficient vector. , This represents the scaling factor for positive constants. This represents the regularization constant.

[0111] Specifically, to reduce switching losses, an event-triggered control mechanism is used to manage voltage control increments. The update is performed only during each sampling period, rather than every sampling period. The specific design of this mechanism is as follows:

[0112] Define event triggering error Triggering error For the current moment Tracking error Compared to the time of the last event trigger Tracking error The difference is:

[0113] ;

[0114] Among them, tracking error Defined as the reference output current at time k The actual output current at time k The difference, that is .

[0115] Actual voltage control increment applied to the system Compared with ideal voltage control increment The relationship between them can be represented as:

[0116] .

[0117] In the ideal voltage control increment The prediction residual error under the action is:

[0118] .

[0119] Substituting the output current increment prediction model and the voltage increment control law into the error equation, the dynamic equation of the prediction error can be obtained as follows:

[0120] .

[0121] To ensure system convergence, i.e., to guarantee that the square of the tracking error is monotonically decreasing. Using the basic inequality Scaling the dynamic equation of the prediction error yields:

[0122] .

[0123] To satisfy the stability condition It only needs to satisfy:

[0124] .

[0125] Define the system stability margin as Ultimately, the following event triggering conditions are established:

[0126] ;

[0127] In the formula, This indicates a preset safety threshold. This represents the tracking error at time k. The prediction residual error under ideal control, This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. Represents the gain coefficient vector. , This represents the scaling factor for positive constants. This represents the regularization constant.

[0128] The event triggering mechanism includes two conditions. The first condition... These are safety protection conditions that ensure tracking errors under any circumstances. It will not exceed the permitted boundaries.

[0129] The second condition is based on trigger error. With dynamic threshold The comparison shows that this dynamic threshold takes into account the current tracking error. Size, system dynamic characteristics (through and (Reflects) and theoretical residual error information. Only when the error is triggered. Control updates are only triggered when the increase is significant enough to potentially affect control performance.

[0130] If neither of the above two conditions is met, the voltage control increment at the previous triggering time will remain unchanged, that is... This significantly reduces unnecessary switching actions, thereby lowering the average switching frequency and switching losses.

[0131] Based on the above description, the event-triggered control mechanism provided in this application embodiment explicitly derives the event trigger threshold without relying on the system parameter model (including but not limited to parameters such as resistance, inductance, and capacitance), effectively reducing the switching frequency and thus reducing switching losses. Furthermore, the event-triggered mechanism only performs control updates when the tracking error exceeds a preset safety threshold, avoiding complex online optimization calculations in most cycles, thereby reducing the computational resource consumption and communication burden of the converter.

[0132] As a preferred embodiment, the system block diagram of the converter's elastic predictive control method is as follows: Figure 3 As shown, the system block diagram includes a system layer and a digital control layer:

[0133] The system layer (converter main circuit) is the controlled three-phase converter main circuit, including the DC power supply. Three-phase fully controlled switching devices (switching state determined by signals) , , The three-phase bridge arm and filter inductor (consisting of control) and equivalent resistance Furthermore, there are instances of false information injection attacks interfering with the converter's output current. , , .

[0134] In the control layer: False information injection attacks interfere with the converter's output current. , , After the abc / αβ coordinate transformation, the output current of the converter in the αβ coordinate system is obtained. , The attack information extraction module receives the converter output current that has been subjected to a false information injection attack. , Attack items are extracted using an online-learning neural network model. .

[0135] Data-driven predictive controller receives reference current , Furthermore, based on the existence of false information, an attack was injected to interfere with the output current of the converter. , The attack information extraction module outputs the predicted value of the converter's output current. Data-driven predictive controller receives reference current , Predicted output current of the converter and attack items Based on the constructed output current increment prediction model containing inherent dynamic terms and attack terms, the optimal voltage control increment and its corresponding optimal reference voltage command are calculated by minimizing the coefficient matrix of the criterion function, which includes tracking error and parameter change penalty terms, and based on the prediction model and reference current. , .

[0136] The event triggering module receives the current tracking error and, based on adaptive triggering conditions independent of circuit parameters, outputs a trigger signal only when the conditions are met. Upon receiving the trigger signal, the switching signal generation module sends the optimal reference voltage command... , A new switching signal is generated by space vector pulse width modulation. , , The output is updated, otherwise the switching state of the previous moment remains unchanged, thereby effectively suppressing the switching frequency and reducing the computation and communication burden while ensuring control performance.

[0137] To further illustrate the elastic predictive control method provided in this application, a comprehensive performance comparison experiment was conducted between the elastic predictive control method for the converter provided in this application and the traditional finite set model elastic predictive control method (FCS-MPC). All experiments were conducted under the same set of 3L-NPC converter system parameters listed in Table 2 to ensure the fairness of the comparison. The system parameters in Table 2 are as follows:

[0138]

[0139] The results of the comparative experiment are as follows:

[0140] like Figure 4 and Figure 5 As shown, under parameter matching conditions, the total harmonic distortion (THD) of the current output by the traditional FCS-MPC method is 2.81%. The THD of the current output by the elastic predictive control method provided in this application is 1.59%. Therefore, under ideal conditions where the system model parameters are precisely known, the current waveform quality of the elastic predictive control method provided in this application is superior to that of the traditional FCS-MPC method.

[0141] To test robustness, the filter inductance value was changed from 10mH to 8mH (i.e., -20% parameter mismatch). For example... Figure 6 and Figure 7As shown, the current THD of the traditional FCS-MPC method deteriorates to 4.21%, indicating that its performance is heavily dependent on the accuracy of the model, and the control performance drops significantly when parameters mismatch occurs. The current THD of the elastic predictive control method provided in this application is 1.96%, and the performance degradation is much smaller than that of the traditional FCS-MPC method. Therefore, the elastic predictive control method provided in this application, through a completely parameterless data-driven framework, eliminates the dependence on fixed physical model parameters and exhibits extremely strong robustness in the face of parameter mismatch.

[0142] Furthermore, under the most severe test conditions, the following two conditions are simultaneously applied: inductance parameter mismatch -20%; and a spurious nonlinear attack signal is injected into the current sensor, the expression of which is: .

[0143] like Figure 8 and Figure 9 As shown, the traditional FCS-MPC method suffers from severely degraded control performance and significant distortion of output current due to the lack of a dedicated mechanism for handling parameter mismatch and network attacks. The resilient predictive control method provided in this application, thanks to its embedded neural network-based real-time attack information extraction and separation module, can effectively identify and isolate this spoofed data injection attack. Under this combined attack, the THD of the system output current remains strictly within the 5% grid-connected standard, maintaining stable control performance.

[0144] Therefore, the elastic predictive control method provided in this application achieves effective online protection against false information injection attacks at the inverter control level, solving the network security weaknesses of traditional methods in a digital interconnected environment.

[0145] Finally, the proposed event-triggered strategy was validated. Without the event-triggered mechanism, the average switching frequency of the data-driven predictive control was 1381 Hz. After enabling the event-triggered mechanism of the flexible predictive control method provided in this application, the average switching frequency of the system decreased to 1086 Hz. This represents a 21.4% reduction in switching frequency. This demonstrates that the event-triggered control mechanism designed in this application can reduce unnecessary switching actions without relying on any system physical parameters, thereby reducing switching losses of power devices and improving system efficiency.

[0146] Compared to existing technologies such as traditional FCS-MPC, this application eliminates the dependence on physical parameters such as resistors and inductors, achieving true model-free control and maintaining excellent performance even when parameters are completely unknown or significantly mismatched. Furthermore, based on neural network-based online attack estimation and feedforward compensation mechanisms, it provides the converter with intrinsic security capabilities against network attacks such as spoofed data injection. The proposed event-triggered control provides a clear theoretical design basis for the trigger threshold, achieving a reduction in switching frequency and losses while ensuring dynamic performance.

[0147] Based on the same concept, this application provides a converter that employs the elastic predictive control method described above.

[0148] This converter can operate in either off-grid or grid-connected mode. The grid-connected converter can acquire system input and output data using voltage and current sensors at the grid-connected interface. It updates the output current increment prediction model online through a completely model-free data-driven architecture and uses an online neural network to estimate in real time the attacks that the grid-connected side may suffer, such as spoofing. Furthermore, it generates voltage commands through event-triggered control laws and selects the optimal switching state to achieve accurate and robust tracking of the grid-connected current.

[0149] The elastic predictive control method provided in this application can be directly applied to grid-connected converters, and can still achieve safe, stable and efficient power conversion and grid-connected operation under adverse conditions such as unknown parameters and network attacks.

[0150] In addition, the converter can be any of the following types of converters: two-level converter, two-level converter, modular multilevel converter.

[0151] Based on the same concept, this application also provides a computer storage medium storing a processing program, which executes the elastic predictive control method described above when executed.

[0152] Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0153] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0154] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method for elastic predictive control of a converter, characterized in that, The elastic predictive control method includes: An output current increment prediction model is constructed, which includes an inherent dynamic term reflecting the mapping relationship between historical data of the converter system and the output current increment, and an attack term reflecting external anomalies. A criterion function is constructed, which includes a tracking term characterizing the error between the actual output current increment and the predicted output current increment, and a penalty term characterizing the change in the coefficient matrix of the inherent dynamic term. The predicted output current increment is obtained through the output current increment prediction model. Calculate the gradient of the criterion function with respect to the coefficient matrix, and set the gradient to zero to obtain an estimate of the coefficient matrix at the current time. The error between the reference current of the next cycle and the actual output current of the current cycle is used as the target output of the output current increment prediction model. The output current increment prediction model is then rewritten to obtain the control law for the voltage increment. Based on the control law of the voltage increment, the coefficient matrix of the current moment is estimated to obtain the voltage control increment, and the converter switching state is selected based on the voltage control increment.

2. The elastic predictive control method for a converter according to claim 1, characterized in that, The elastic predictive control method further includes the following steps: A neural network model is constructed to establish a mapping relationship between the attack item and the system state, wherein the system state is the output current of the converter; Design a state estimator that superimposes the output of the neural network model with the output of the output current increment prediction model as a prediction term, and introduces a feedback correction term based on the output current estimation error to achieve system state tracking. The output current estimation error is the difference between the output current sample value and the system state estimate output by the state observer. Based on the state estimator, the update law of the weight coefficient matrix of the neural network model is configured according to the stability criterion, so that the output current estimation error converges to zero under the action of the update law, and the update law includes a gradient descent term related to the output current estimation error. Based on the output current estimation error of the current cycle, the weight coefficient matrix is ​​updated using the update law. Using the updated weight coefficient matrix and the current system state, the attack term estimate for the next control cycle is calculated.

3. The elastic predictive control method for a converter according to claim 2, characterized in that, The control method further includes: Based on a preset control sampling period, the state observer and the update law are discretized using the first-order Euler method to generate a discretized prediction model of the attack term. Based on the output current estimation error of the current period, the attack term estimate for the next control period is calculated using the discretized prediction model.

4. The elastic predictive control method for a converter according to claim 2 or 3, characterized in that, The control method further includes: The voltage control increment is compensated based on the estimated attack term.

5. The elastic predictive control method for a converter according to claim 2, characterized in that, The system state vector is constructed using the historical output current sequence of the converter. The neural network model adopts a radial basis function neural network model, and the activation function of the neural network model is a Gaussian radial basis function with the system state vector as the variable.

6. The elastic predictive control method for a converter according to claim 1, characterized in that, The voltage control increment is updated using an event-triggered control mechanism, which includes: Define event triggering error The event triggering error Current tracking error Tracking error compared to the last trigger time The tracking error is the difference between the reference output current and the actual output current. The voltage control increment is updated when the following conditions are met: ； In the formula, This indicates a preset safety threshold. , This represents the tracking error at time k. The prediction residual error under ideal control, This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. Represents the gain coefficient vector. , This represents the scaling factor for positive constants. This represents the regularization constant.

7. The elastic predictive control method for a converter according to claim 1, characterized in that, The control law for the voltage increment is expressed by the following formula: ； ； ； In the formula, Represents the gain coefficient vector. This represents the scaling factor for positive constants. Represents the regularization constant. Represents the historical influence vector. This represents the (Ly+1)th coefficient vector in the coefficient matrix estimated at time k. This refers to the attack item. This indicates the voltage increment.

8. A converter, characterized in that, The converter employs the elastic predictive control method as described in any one of claims 1 to 7.

9. The converter according to claim 8, characterized in that, The converter can be any one of the following types of converters: two-level converter, two-level converter, or modular multilevel converter.

10. A computer storage medium, characterized in that, The computer storage medium stores a processing program, which, when executed, performs the elastic predictive control method as described in any one of claims 1 to 7.

Images