A Distributed Finite Control Set Model Predictive Control Method and Controller

Abstract

The present application discloses a distributed finite control set model predictive control method and a controller. The method expresses the voltage of the first arm in the form of voltage levels, constructs a discrete mathematical model of the first arm expressed in voltage levels, performs calculations based on the discrete mathematical model of the first arm voltage expressed in voltage levels, reduces the number of switching states to be calculated by converting the selection of switching states into the selection of voltage levels, reduces the computational burden, and thus realizes the fast control of the modular multilevel matrix converter. Moreover, the present application further analyzes the potential categories of false data injection attacks that may occur in this control structure, explores the deficiencies of conventional detectors, and proposes a detector combining a fuzzy logic structure and a neural network to achieve the fast detection of false data injection attacks and ensure system security.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular, to a distributed finite control set model predictive control method and a controller. Background Art

[0002] Since the MMMC (Modular Multilevel Matrix Converter) has a total of nine arms, and each arm is formed by connecting a plurality of AC-AC conversion sub-modules in series. In the related art, the traditional MPC (model predictive control) control system based on switch states has complex calculations and needs to traverse the switch states of all sub-modules in each arm during operation, resulting in a huge computational burden for the model predictive control of the MMMC. Summary of the Invention

[0003] In order to solve the deficiencies of the prior art, the present application adopts the following technical solutions:

[0004] In a first aspect, the present application provides a distributed finite control set model predictive control method, which is applied to the control of a modular multilevel matrix converter. The modular multilevel matrix converter includes nine arms, and any one arm includes a plurality of AC-AC conversion sub-modules. Each arm is arranged between the input port and the output port of the modular multilevel matrix converter. The predictive control method includes:

[0005] Based on Kirchhoff's current law, model the dynamics of the current of the first arm to obtain the arm current dynamic model of the first arm;

[0006] Express the voltage of the first arm in the form of a voltage series. Based on the voltage of the first arm expressed by the voltage series, discretize the arm current dynamic model to construct the discrete mathematical model of the first arm. The discrete mathematical model includes the voltage of the first arm expressed by the voltage series;

[0007] Based on the discrete mathematical model, according to the finite voltage levels of the AC-AC conversion sub-modules in the first arm, calculate the predicted output values of the first arm at different voltage levels respectively;

[0008] Obtain the reference output value of the first arm. Based on the predicted output value and the reference output value of the first arm, construct a basic cost function corresponding to the first arm, and take the minimum of the basic cost function as the optimal switch state;

[0009] Traverse all arms and select the optimal switch states of each arm to act on the modular multilevel matrix converter.

[0010] In summary, a distributed finite control set model predictive control method provided by the present application expresses the voltage of the first arm in the form of voltage levels, constructs a discrete mathematical model of the first arm expressed in voltage levels, performs calculations based on the discrete mathematical model of the first arm voltage expressed in voltage levels, and reduces the number of switching states to be calculated and the computational burden by converting the selection of switching states into the selection of voltage levels, thereby achieving fast control of the modular multilevel matrix converter.

[0011] Further, the voltage levels of the first arm are expressed as follows:

[0012] ;

[0013] ;

[0014] ;

[0015] where V bk represents the voltage of the k-th arm, N b represents the discretized voltage levels, N level represents the number of non-negative voltage levels that can be selected after considering the balance between calculation time and accuracy, V k-s represents the voltage amplitude corresponding to each voltage level, V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first arm.

[0016] Further, the discrete mathematical model is represented by the following formula:

[0017] ;

[0018] where T s is the sampling time, represents the voltage drop of the arm viewed from the external perspective, L b represents the arm inductance, i bk (k) represents the current of the arm at time k, V com (k) represents the potential difference between the neutral points on both sides at time k.

[0019] Further, the method further includes:

[0020] obtaining the capacitor voltages of the AC-AC conversion sub-modules in the first arm, and achieving the capacitor voltage balance of the i-th AC-AC conversion sub-module in the first arm through the following equation:

[0021] ;

[0022] where is the control signal input to the modulation stage, is the gain parameter to be selected, sign(*) represents the sign function, ibk represents the current of the k-th bridge arm, N represents the number of AC-AC conversion sub-modules in each bridge arm, and V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first bridge arm, and V con is the bridge arm voltage reference.

[0023] Further, the basic cost function is the square of the error between the predicted output value of the first bridge arm and the reference output value.

[0024] Further, the method further includes:

[0025] Detecting any one of the AC-AC conversion sub-modules by using a Kalman filter to obtain the prior estimate and the measured value of the AC-AC conversion sub-module;

[0026] Constructing a single-layer neural network, including configuring network parameters, where the network parameters at least include the number of nodes in the input layer, the hidden layer, and the output layer;

[0027] Using the prior estimate, the integral of the prior estimate, the measured value, and the integral of the measured value as the input of the single-layer neural network, and determining whether the modular multilevel matrix converter is abnormal based on the output of the single-layer neural network;

[0028] Wherein, the output of the hidden layer is represented by the following formula:

[0029] ;

[0030] The output of the output layer is represented by the following formula:

[0031] ;

[0032] In the formula, represents the activation function, represents the i-th input node of the system, represents the input bias, J and K respectively represent the number of input nodes and the number of hidden nodes, represents the output of the j-th hidden layer neuron.

[0033] Further, the method further includes:

[0034] Adopting a random weight method to obtain the weights from the input layer to the hidden layer in the single-layer neural network for a preset number of times, and the biases corresponding to the weights;

[0035] The obtained weights are evaluated using a validity function, and the optimal weight and corresponding bias are selected therefrom.

[0036] Further, the validity function is expressed by the following formula:

[0037] ;

[0038] ;

[0039] where, represents the validity function for evaluating randomly input weights, is an intermediate variable, represents the error of the k-th neuron, and M are learning parameters, and , and the superscript T represents transpose.

[0040] Further, the method further includes:

[0041] Through forward propagation calculation, the weights of the output layer are optimized by minimizing the following objective function, and the objective function is expressed by the following formula:

[0042] ;

[0043] In the formula, represents the output layer weights, represents the weight corresponding to the j-th neuron in the i-th sampling, represents the output of the i-th sampling.

[0044] In a second aspect, the present application also provides a controller for a modular multilevel matrix converter, and the controller adopts the distributed finite control set model predictive control method described in any one of the above. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic diagram of the topology structure of a modular multilevel matrix converter provided by an embodiment of the present application;

[0046] Figure 2 is a flowchart of the steps for quickly controlling a modular multilevel matrix converter by the distributed finite control set model predictive control method provided by an embodiment of the present application;

[0047] Figure 3 is a schematic diagram of the simulation result of the classical MPC method provided by a comparative example of the present application;

[0048] Figure 4 is a schematic diagram of the simulation result of the sequential MPC method provided by a comparative example of the present application;

[0049] Figure 5 Schematic diagram of the simulation results of the distributed finite control set model predictive control method provided by an embodiment of the present application;

[0050] Figure 6 Schematic diagram of the strategy of applying the distributed finite control set model predictive control method provided by an embodiment of the present application to a bridge arm;

[0051] Figure 7 Flowchart of the steps for the distributed finite control set model predictive control method provided by an embodiment of the present application to detect Xu Jian data injection attacks;

[0052] Figure 8 Schematic diagram of constructing a single-layer neural network in the distributed finite control set model predictive control method provided by an embodiment of the present application;

[0053] Figure 9 Detection result graph of the distributed finite control set model predictive control method provided by an embodiment of the present application under two false data injection attacks;

[0054] Figure 10 Schematic diagram of the overall detection framework of the distributed finite control set model predictive control method provided by an embodiment of the present application. Specific embodiments

[0055] The present application will be described in detail below in conjunction with the specific embodiments shown in the drawings, but these embodiments do not limit the present application. Structural, method, or functional transformations made by those of ordinary skill in the art based on these embodiments are all included within the protection scope of the present application.

[0056] To solve the deficiencies of the prior art, the present application provides a distributed finite control set model predictive control method, which is applied to the control of a modular multilevel matrix converter. The modular multilevel matrix converter includes nine bridge arms, and any bridge arm includes a plurality of AC-AC conversion sub-modules. Each bridge arm is arranged between the input port and the output port of the modular multilevel matrix converter. The topological structure of the MMMC is as Figure 1 shown. The MMMC is composed of 9 bridge arms, and each bridge arm is connected between an input port and an output port.

[0057] As Figure 2 shown, the predictive control method includes:

[0058] Step S11, based on Kirchhoff's current law, model the dynamics of the current of the first bridge arm to obtain the bridge arm current dynamic model of the first bridge arm;

[0059] Step S12: Express the voltage of the first arm in the form of voltage levels. Based on the voltage of the first arm expressed in voltage levels, discretize the dynamic model of the arm current to construct the discrete mathematical model of the first arm. The discrete mathematical model includes the voltage of the first arm expressed in voltage levels.

[0060] Step S13: Based on the discrete mathematical model, calculate the predicted output values of the first arm at different voltage levels according to the finite voltage levels of each AC-AC conversion sub-module in the first arm.

[0061] Step S14: Obtain the reference output value of the first arm. Based on the predicted output value and the reference output value of the first arm, construct the basic cost function corresponding to the first arm, and take the minimum of the basic cost function as the optimal switching state.

[0062] Step S15: Traverse all arms and select the optimal switching states of each arm to act on the modular multilevel matrix converter.

[0063] As Figure 3 、 Figure 4 and Figure 5 shown, compare the classical MPC method, the sequential MPC method and the distributed MPC method proposed in this paper, and introduce error variables to analyze the results. Figure 3 、 Figure 4 and Figure 5 In εi ,(a) part represents the output three-phase current waveform, (b) part represents the source-side three-phase current waveform, (c) part represents the capacitor voltage of the sub-module, and (d) part represents the result of the fast Fourier transform analysis. εc represents the percentage of the average tracking error of the three-phase current,

[0064] represents the percentage of the error between the average capacitor voltage of a single arm and the reference capacitor voltage. The first arm is selected for testing in the figure. Introduce THD to evaluate the current.

[0065] According to the above description, a distributed finite control set model predictive control method provided by the present application converts the voltage of the first arm into a voltage level form, constructs a discrete mathematical model of the first arm expressed in voltage levels, and performs calculations based on the discrete mathematical model of the first arm voltage expressed in voltage levels. By converting the selection of the switching state into the selection of voltage levels, the number of switching states to be calculated is reduced, and the calculation burden is reduced, thereby achieving fast control of the modular multilevel matrix converter.

[0066] As Figure 1 shown, , and represent the voltages input to the grid side, while represents the three-phase current input to the grid side. represents the three-phase voltage on the output side, represents the output three-phase voltage. The potential difference between the neutral points on both sides is defined as .

[0067] It can be seen from Figure 1 that each arm consists of N AC-AC conversion sub-modules. Assuming that the capacitance of each AC-AC conversion sub-module is , and the reference voltage of each AC-AC conversion sub-module is , then under ideal conditions, each AC-AC conversion sub-module can output three voltage levels: , 0, and . The inductance of the arm is denoted as .

[0068] In step S11 of the embodiment of the present application, by applying Kirchhoff's current law to the current of the first arm, the dynamics of the current of the first arm are modeled, thereby obtaining the arm current dynamic model of the first arm. The modeling process can be expressed by the following formula:

[0069] (1);

[0070] In the formula, V u , V v , and V w represent the voltages input to the grid side, V r , V s , and V t represent the output three-phase voltages, V com represents the potential difference between the neutral points on both sides, I 3x3 represents the identity matrix of dimension 3, L b represents the inductance of the arm, i bk represents the current of the b k th arm, k = {1, 2,..., 9}, Vbk Denote the voltage of the b-th k bridge arm.

[0071] Combining formula (1) and Figure 1 the structure in, the currents on the input side and the output side satisfy the following relationship:

[0072] (2);

[0073] In the formula, denotes the three-phase current input to the grid side. denotes the three-phase voltage on the output side.

[0074] In the related art, in the traditional MPC control algorithm based on the switching state, the switching states of the AC-AC conversion sub-modules of the first bridge arm are represented as , where , then the voltage of any bridge arm is expressed by the following formula:

[0075] (3);

[0076] In the formula, denotes the sum of the voltages of all the AC-AC conversion sub-modules in the k-th bridge arm, V kj denotes the switching state of the AC-AC conversion sub-module, denotes the capacitor voltage of the j-th AC-AC conversion sub-module in the k-th bridge arm.

[0077] Substituting formula (3) into formula (1) and discretizing, the dynamics of the k-th bridge arm at time k can be expressed as:

[0078] (4);

[0079] In the formula, L b denotes the inductance of the bridge arm, T s denotes the sampling time, denotes the voltage drop of the k-th bridge arm seen from the external perspective.

[0080] At time k, the bridge arm current at time (k + 1) can be predicted using equation (4). Assuming that the reference current of each bridge arm has been obtained, the basic cost function of each bridge arm can be constructed.

[0081] In the traditional MPC control algorithm based on the switching state, the main idea is to traverse all the switching states to optimize the basic cost function of each bridge arm, and then, select the switching state that minimizes the current error at time k + 1. The total number of switching states that need to be calculated in each traversal process is . It can be seen that in the traditional MPC control algorithm based on the switching state, the number of iterations grows exponentially with the number of sub-modules, resulting in a huge computational load.

[0082] In step S12 of the present application, the voltage of the first arm is expressed in the form of voltage levels. As an implementation, the voltage level of the first arm is expressed in the following form:

[0083] (5);

[0084] In the formula, represents the sum of the voltages of all AC-AC conversion sub-modules in the k-th arm, and N b represents the discretized voltage level, and V k-s represents the voltage amplitude corresponding to each voltage level.

[0085] Furthermore, N b is expressed as follows:

[0086] (6);

[0087] In the formula, N b represents the discretized voltage level, represents the number of non-negative voltage levels that can be selected considering the balance between calculation time and accuracy.

[0088] When the capacitor voltage of the AC-AC conversion sub-module does not change significantly within a single time step, the measurement information of the previous time step can be used to calculate V k-s , and the calculation formula is expressed as follows:

[0089] (7);

[0090] In the formula, represents the voltage amplitude corresponding to each level, and V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first arm, represents the number of non-negative voltage levels that can be selected considering the balance between calculation time and accuracy.

[0091] Based on the voltage of the first arm expressed in voltage levels, the dynamic model of the arm current is discretized to construct the discrete mathematical model of the first arm. The discrete mathematical model includes the voltage of the first arm expressed by voltage data and is represented by the following formula:

[0092] (8);

[0093] In the formula, T s represents the sampling time, and L b represents the inductance of the arm, Indicates the voltage drop of the arm from an external perspective , V com Indicates the potential difference between the neutral points on both sides.

[0094] In the above manner, the selection of the switching states of the AC-AC conversion sub-module can be transformed into the selection of voltage levels, thereby reducing the number of switching states that need to be calculated. The number of switching states that need to be calculated is reduced from to .

[0095] As an implementation, in step S14, based on the predicted output value and the reference output value of the first arm, a basic cost function corresponding to the first arm is constructed. The basic cost function is the square of the error between the predicted output value and the reference output value of the first arm, and the formula is expressed as follows:

[0096] (9);

[0097] In the formula, represents the reference output value of the first arm at time, represents the predicted output value of the first arm at time.

[0098] Based on the finite voltage levels of the arm, optimize the basic cost function of each arm in formula (9), select the switching state with the minimum current error of each arm as the optimal switching state, and apply the optimal switching state to the modular multilevel matrix converter, thereby realizing the fast control of the modular multilevel matrix converter.

[0099] Furthermore, the effectiveness of the control method provided by this application depends on the modulation stage and the capacitor voltage balance of the AC-AC conversion sub-module. As an implementation, the control method provided by this application further includes obtaining the capacitor voltages of each AC-AC conversion sub-module in the first arm, and realizing the capacitor voltage balance of the i-th AC-AC conversion sub-module in the first arm through the following equation:

[0100] (10);

[0101] In the formula, represents the control signal input to the modulation stage, represents the gain parameter that needs to be selected, sign(*) represents the sign function, V c-ki represents the capacitor voltage of the i-th AC-AC conversion sub-module in the first arm, V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first arm, V conRepresents the arm voltage reference.

[0102] The strategy for a single arm is as Figure 6 shown, and the same configuration is applied to all other arms. The central controller is responsible for handling higher-level tasks such as overall energy balance and current control, while the distributed controller achieves more efficient management of the energy balance of the AC-AC conversion sub-module, realizing the balance of the modulation stage and the capacitor voltage of the AC-AC conversion sub-module. For the problem of sub-module capacitor voltage balance solved in this application, it is assumed that all LCs (local controllers) within the same arm can communicate effectively, so they will be able to obtain the capacitor voltage information of other SMs (sub-modules) within the arm. Under the above assumption, it is assumed that the arm voltage reference transmitted by the high-level controller is . Since the FCS-MPC omits the modulation stage, the capacitor voltage balance of the i-th SM in the k-th arm can be achieved through Equation (10). According to Equation (10), the control signal for achieving capacitor voltage balance can be obtained by separately calculating in each AC-AC conversion sub-module, and the parallel structure of multiple sub-controllers can reduce the calculation time.

[0103] Furthermore, in the distributed finite control set model predictive control method provided in this application, considering that data transmission between sub-modules depends on the transmission network under distributed control, there is a potential risk of false data injection attack, and in some cases, the existing Kalman filter fails to detect this attack. In response to this, this application further analyzes the detectability of FDIA (false data injection attack) in the MMMC system and provides corresponding detection means to ensure that the system has the detection ability for various different FDIA attacks to ensure the stability of the MMMC system.

[0104] Specifically, the analysis of the detectability of FDIA in the MMMC system in this application is as follows:

[0105] False data injection attacks may cause local controllers to receive incorrect information. Since each AC-AC conversion sub-module is decoupled, the capacitor voltage state matrix of the cluster is a diagonal matrix. For a single AC-AC conversion sub-module, its state equation can be written in scalar form as follows:

[0106] (11);

[0107] In the formula, represents the capacitor voltage, represents the control input, w(k) represents the process noise, v(k) represents the measurement noise, w(k) satisfies , and v(k) satisfies .

[0108] When a false data injection attack occurs, the measured value of the capacitor voltage will be changed first, resulting in a deviation in the control result based on these measured values. Let be the attack sequence. Under the attack, the state equation of the system can be expressed as:

[0109] (12);

[0110] In the formula, represents the attack sequence, represents the system control input after being attacked, represents the system output after being attacked, represents the system state after being attacked.

[0111] The Kalman filter detects the AC-AC conversion sub-module to obtain the prior estimate and measured value of the AC-AC conversion sub-module. The detection process can be expressed as follows:

[0112] (13);

[0113] In the formula, represents the predicted state, represents the predicted covariance matrix, represents the Kalman gain, represents the corrected system state, represents the corrected covariance, represents the measurement output.

[0114] Define as the error between the estimated value and the reference value, as the state estimation error, as the residual, which can be expressed as:

[0115] (14);

[0116] In the formula, represents the actual voltage value of this sub-module.

[0117] Combining formula (13) and formula (14), it can be obtained that is the same as e(k), and the state matrix of the AC-AC conversion sub-module does not contain unstable eigenvalues. Thus, in the MMMC system, the occurrence of is avoided while z(k) remains finite. Since there is no capacitor voltage reference value in the local controller, the false data injection attack on the MMMC system can be simplified to the following form:

[0118] (15);

[0119] In the formula, represents the error between the estimated value and the reference value, and represents the residual.

[0120] The existing Kalman filter detection methods mainly rely on residual-based monitoring, and the expression is as follows:

[0121] (16);

[0122] In the formula, represents the detection function.

[0123] At each sampling moment, calculate Gi. When Gi exceeds the preset threshold , the detector triggers an alarm, indicating that a false data injection attack is detected. At this time, set the preset threshold to , where , and design so that the detector does not trigger an alarm at any time. Before the false data injection attack occurs, the parameters P(k) and K(k) of the Kalman filter have converged, that is:

[0124] (17);

[0125] In the formula, P(k) represents the covariance matrix, and K(k) represents the Kalman gain.

[0126] Through the above description, the system equation of the residual under the false data injection attack can be derived as follows:

[0127] (18);

[0128] In the formula, represents the residual, and represents the attack sequence.

[0129] Combining formula (15) and formula (16), the following conclusion can be drawn: If no alarm is triggered at any time k, then satisfies . Assuming , then it can be observed that . Thus, the following conclusion can be derived:

[0130] (19);

[0131] In the formula, represents the detection threshold.

[0132] Set the start and end times of the false data injection attack to be , respectively, and use Denote the attack sequence. Then, a simple stealth attack algorithm that is undetectable by the Kalman filter can be designed based on the constraint condition (19), and its characteristic is that the gradient satisfies the requirements of formula (19).

[0133] Based on the above analysis and description of the attack form, as Figure 7 shown, the distributed finite control set model predictive control method provided by this application further includes the following steps to efficiently detect various different FDIA attacks including the false data injection attack described above, including the following steps:

[0134] Step S21: Use the Kalman filter to detect any AC-AC conversion sub-module to obtain the prior estimate and measurement value of the AC-AC conversion sub-module;

[0135] Step S22: Construct a single-layer neural network, including configuring network parameters, where the network parameters at least include the number of nodes in the input layer, hidden layer, and output layer;

[0136] Step S23: Use the prior estimate, the integral of the prior estimate, the measurement value, and the integral of the measurement value as the input of the single-layer neural network, and determine whether the modular multilevel matrix converter is abnormal based on the output of the single-layer neural network.

[0137] The distributed finite control set model predictive control method provided by this application directly obtains the detection result by processing the output of the Kalman filter. As Figure 8 shown, construct a single-layer neural network and configure network parameters. In the figure, is the activation function, h is the output of the hidden layer, represents the weight matrix, and b is the bias of the input layer. The network parameters at least include the number of nodes in the input layer, hidden layer, and output layer. The output of the hidden layer is expressed by the following formula:

[0138] (20);

[0139] In the formula, represents the jth hidden layer, represents the activation function, represents the ith input node of the system, represents the input bias. The output of the output layer is expressed by the following formula:

[0140] (21);

[0141] In the formula, represents the output of the output layer, J and K respectively represent the number of input nodes and hidden nodes, represents the output of the jth hidden layer neuron.

[0142] The AC-AC conversion sub-module is detected by a Kalman filter to obtain the prior estimate and measurement value of the AC-AC conversion sub-module. The prior estimate, the integral of the prior estimate, the measurement value, and the integral of the measurement value are used as the inputs of a single-layer neural network. Through the calculation of the single-layer neural network, it is determined whether the modular multilevel matrix converter is abnormal based on the output result of the single-layer neural network. The input sequence length for attack detection is set to N, denoted as x, and . The corresponding detection result is , where . When , it indicates that a false data injection attack has occurred. When, it indicates that no false data injection attack has occurred.

[0143] As Figure 9 shown, the detector proposed in this paper is simulated and verified under the step false data injection attack and the covert false data injection attack proposed in this paper. The capacitor voltage values of the AC-AC conversion sub-module received by the local controller and the true capacitor voltage values of the AC-AC conversion sub-module when the two attacks occur respectively are as Figure 9 in (a) and Figure 9 in (b) shown. It can be seen that the false data injection attack can cause a large gap between the information obtained on the local controller side and the true value, resulting in a decline in the system control effect. The results of the two detectors are as Figure 9 in (c) shown. The conventional Kalman detector has a slightly faster detection speed for the step false data injection attack than the proposed detector due to its simple structure, but it cannot detect the covert attack in this paper, while the detector proposed in this application can effectively detect both attacks.

[0144] As an implementation, in step S23, it further includes:

[0145] Step S231, using the random weight method to obtain the weights from the input layer to the hidden layer in the single-layer neural network for a preset number of times, and the biases corresponding to the weights;

[0146] Step S232, using the effectiveness function to evaluate the weights obtained each time, and selecting the optimal weights and corresponding biases from them.

[0147] Specifically, in step S231, the weights of the input layer are obtained by the random weight method to reduce the computational burden of online detection. The random weight method is expressed as follows:

[0148] (22);

[0149] In the formula, represents the weight, and Range parameters representing weights and biases.

[0150] If the number of neurons in the output layer is set to L, the definition of the output error is expressed as follows:

[0151] (23);

[0152] Where, Represents the sampled value of the output within N steps.

[0153] In step S232, for each hidden layer neuron, the input weights are randomly generated times and evaluated by selecting the validity function to ensure that the randomly selected input layer weights can converge quickly during training.

[0154] As an implementation, the selected validity function is expressed as follows:

[0155] (24);

[0156] Where, Represents the validity function used to evaluate the random input weights, is an intermediate variable, represents the error of the k-th neuron, and M are learning parameters, and , the superscript T represents transpose.

[0157] Among them:

[0158] (25);

[0159] Where, represents transpose and M is a learning parameter.

[0160] As an implementation, through forward propagation calculation, the weights of the output layer are optimized by minimizing the objective function.

[0161] Specifically, by selecting the best from randomly generated values, its optimized form is:

[0162] (26);

[0163] Where, represents the optimal input bias.

[0164] After obtaining the input layer weights, the output layer weights are optimized by forward propagation calculation through minimizing the following objective function:

[0165] (27);

[0166] Wherein, represents the output layer weight, represents the weight corresponding to the j-th neuron in the i-th sampling, represents the output of the i-th sampling.

[0167] By solving this optimization problem:

[0168] (28);

[0169] By selecting an appropriate sample data set and training according to the above method, real-time detection of the AC-AC conversion table sub-module is achieved. In the above description, the input of the input layer of the single-layer neural network is selected as , where represents the state estimated by the Kalman filter (prior estimate), and the output of the single-layer neural network is . As Figure 10 shown, by adjusting the parameters and training under various different FDIA attacks, the accuracy of the online detection ability of this control method can be further improved.

[0170] According to the above description, a distributed finite control set model predictive control method provided by the present application converts the voltage of the first bridge arm into a voltage series form, constructs a discrete mathematical model of the first bridge arm expressed in voltage series, calculates based on the discrete mathematical model of the first bridge arm voltage expressed in voltage series, and reduces the number of switching states to be calculated and the calculation burden by converting the selection of the switching state into the selection of voltage levels, thereby realizing the fast control of the modular multilevel matrix converter; and, by analyzing the false data injection attack target and constraints of the Kalman filter, a single-layer neural network is constructed to improve the comprehensiveness of the detection of false data injection attacks and ensure the stable operation of the modular multilevel matrix converter.

[0171] In a second aspect, the present application also provides a controller for a modular multilevel matrix converter. The controller adopts the distributed finite control set model predictive control method described above to realize the fast control of the modular multilevel matrix converter and ensure the stable operation of the modular multilevel matrix converter.

[0172] It will be understood that the term "exemplary" as used herein means "serving as an example, instance, or illustration". Any embodiment described as "exemplary" is not necessarily preferred or superior to other embodiments and / or does not preclude the combination of features with other embodiments. It should be understood that certain features of the present application described in the context of separate embodiments for clarity may also be provided in combination in a single embodiment. Conversely, the various features of the present application described in the context of a single embodiment for clarity may also be provided separately or in any suitable combination or as any other described embodiment of the present application.

[0173] The foregoing disclosure is only the preferred embodiment of the present application, but it is not intended to limit the scope of the rights of the present application. Those of ordinary skill in the art can understand that within the spirit and scope of the present application and the appended claims, changes, modifications, substitutions, combinations, and simplifications should all be equivalent replacement methods and still fall within the scope covered by the invention.

Claims

1. A distributed finite control set model predictive control method applied to the control of a modular multilevel matrix converter. The modular multilevel matrix converter includes nine arms, and any one of the arms includes a plurality of AC-AC conversion sub-modules. Each arm is disposed between the input port and the output port of the modular multilevel matrix converter, and is characterized in that The predictive control method includes: Based on Kirchhoff's current law, model the dynamics of the current of the first arm to obtain the arm current dynamic model of the first arm; Express the voltage of the first arm in the form of a voltage series. Based on the voltage of the first arm expressed by the voltage series, discretize the arm current dynamic model to construct the discrete mathematical model of the first arm. The discrete mathematical model includes the voltage of the first arm expressed by the voltage series; Based on the discrete mathematical model, calculate the predicted output values of the first arm at different voltage levels respectively according to the finite voltage levels of each AC-AC conversion sub-module in the first arm; Obtain the reference output value of the first arm. Based on the predicted output value and the reference output value of the first arm, construct a basic cost function corresponding to the first arm, and take the minimum of the basic cost function as the optimal switching state; Traverse all arms and select the optimal switching states of each arm to act on the modular multilevel matrix converter.

2. The distributed finite control set model predictive control method according to claim 1, wherein The voltage level of the first arm is expressed as follows: ； ； ； where, V bk represents the voltage of the k-th bridge arm, N b represents the discretized voltage level, N level represents the number of non-negative voltage levels that can be selected considering the balance between calculation time and accuracy, V k-s represents the voltage amplitude corresponding to each voltage level, V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first bridge arm.

3. The distributed finite control set model predictive control method according to claim 2, wherein The discrete mathematical model is represented by the following formula: ； Where, T s is the sampling time, represents the voltage drop across the arm viewed from the external perspective, L b represents the arm inductance, i bk (k) represents the current of the arm at time k, V com (k) represents the potential difference between the neutral points on both sides at time k.

4. The distributed finite control set model predictive control method according to claim 3, characterized in that The method further includes: Obtain the capacitor voltages of each AC-AC conversion sub-module in the first arm, and achieve the capacitor voltage balance of the i-th AC-AC conversion sub-module in the first arm through the following equation: ； wherein, is the control signal input to the modulation stage, is the gain parameter to be selected, sign(*) represents the sign function, ibk represents the current of the k-th bridge arm, N represents the number of AC-AC conversion sub-modules in each bridge arm, V c-kj represents the capacitor voltage of the j-th AC-AC conversion sub-module in the first bridge arm, V con is the bridge arm voltage reference.

5. The distributed finite control set model predictive control method according to claim 1, wherein The basic cost function is the square of the error between the predicted output value and the reference output value of the first arm.

6. The distributed finite control set model predictive control method according to claim 1, characterized in that The method further includes: Use a Kalman filter to detect any one of the AC-AC conversion sub-modules to obtain the prior estimate and measurement value of the AC-AC conversion sub-module; Construct a single-layer neural network, including configuring network parameters, and the network parameters at least include the number of nodes in the input layer, hidden layer, and output layer; Use the prior estimate, the integral of the prior estimate, the measurement value, and the integral of the measurement value as the input of the single-layer neural network, and judge whether the modular multilevel matrix converter is abnormal based on the output of the single-layer neural network; Among them, the output of the hidden layer is represented by the following formula: ； The output of the output layer is represented by the following formula: ； In the formula, represents the activation function, represents the i-th input node of the system, represents the input bias, J and K respectively represent the number of input nodes and the number of hidden nodes, represents the output of the j-th hidden layer neuron.

7. The distributed finite control set model predictive control method according to claim 6, characterized in that The method further includes: Use the random weight method to obtain the weights from the input layer to the hidden layer in the single-layer neural network for a preset number of times, and the biases corresponding to the weights; Use an effectiveness function to evaluate the weights obtained each time, and select the optimal weights and corresponding biases therefrom.

8. The distributed finite control set model predictive control method according to claim 7, wherein The effectiveness function is represented by the following formula: ； ； Among them, represents the validity function, which is used to evaluate the random input weights, is an intermediate variable, represents the error of the k-th neuron, and M are learning parameters, and , the superscript T represents transpose.

9. The distributed finite control set model predictive control method according to claim 7, characterized in that The method further includes: Through forward propagation calculation, optimize the weights of the output layer by minimizing the following objective function, and the objective function is represented by the following formula: ； In the formula, represents the output layer weight, represents the weight corresponding to the j-th neuron in the i-th sampling, represents the output of the i-th sampling.

10. A controller for a modular multilevel matrix converter, characterized in that, The controller adopts the distributed finite control set model predictive control method described in any one of claims 1 to 9.

Images