Photovoltaic scheduling optimization method and system based on inverse optimal control and recursive high-order neural network

Abstract

A photovoltaic scheduling optimization method based on inverse optimal control and a recursive high-order neural network comprises the following steps: S1, constructing a multi-machine photovoltaic power generation system, and performing inversion power supply based on a received control signal output by a primary control layer; s2, constructing a basic optimization model based on an inverse optimal control method in combination with a recursive high-order neural network; s3, constructing a grid-connected optimization model by taking the active power and the reactive power of the micro-grid as the input of the basic optimization model, constructing an offline optimization model by taking the frequency and the voltage of the micro-grid as the input of the basic optimization model, combining the grid-connected optimization model and the offline optimization model to obtain a secondary control layer, and taking the output of the secondary control layer as an optimal current reference value; and S4, constructing a primary control layer by taking the optimal current reference value and the current and voltage state vectors as input of a basic optimization model, wherein the primary control layer outputs a control signal to the multi-machine photovoltaic power generation system. According to the design, optimization control can be carried out according to different power consumption operation scenes, and the power supply stability and continuity are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of distributed power optimization scheduling method, more particularly to photovoltaic scheduling optimization method and system based on inverse optimal control and recursive high-order neural network. BACKGROUND

[0002] The power supply mode of traditional power grid is mainly centralized large power grid power supply, in this large interrelated power system, some small disturbances or faults can cause great chain reaction, leading to large-area power failure, causing huge losses. In recent years, with the rapid development of renewable energy technology, distributed photovoltaic system is widely used in power system.

[0003] The quality of distributed photovoltaic system generation mainly depends on the research of photovoltaic grid-connected inverter control, the common outer ring control method mainly includes: constant power control, constant voltage constant frequency control, droop control. Constant power control is essentially decoupled active and reactive power control, distributed photovoltaic power supply cannot maintain the microgrid system frequency and voltage by this way; Constant voltage constant frequency control aims to maintain the bus voltage amplitude and microgrid frequency constant, which is mainly used in island operation mode, similar to the balance node of conventional power system; Droop control simulates the static characteristic of generator set, which can distribute load power among distributed photovoltaic power sources, but it will cause large fluctuations in microgrid voltage and frequency. To solve these problems, this paper adopts neural inverse optimal (NIOC) distributed collaborative control to realize the stability of voltage and frequency of microgrid in parallel and off-grid.

[0004] In summary, when microgrid is parallel and off-grid, the following constraints exist:

[0005] 1. Parallel / Off-grid switching response time ≤0.2s, fault switching response time ≤0.2s;

[0006] 2. Frequency deviation absolute value ≤0.5Hz;

[0007] 3. Voltage deviation absolute value ≤ (grid voltage rated value).

[0008] Under the above conditions, the energy scheduling of each unit needs to be optimized, which can improve the performance of single photovoltaic grid-connected inverter, prolong the service life and optimize the power quality; On the other hand, it can improve the energy utilization level of new energy multiple units, and improve the stability and continuity of power supply. SUMMARY

[0009] The purpose of the present application is to overcome the problems of parallel and off-grid scheduling in the prior art, and to provide a photovoltaic scheduling optimization method and system based on inverse optimal control and recursive high-order neural network for optimizing parallel and off-grid scheduling.

[0010] To achieve the above objectives, the technical solution of the present invention is:

[0011] In a first aspect, the present invention provides a photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks, the scheduling method comprising the following steps:

[0012] S1 constructs a multi-machine photovoltaic power generation system, which is used to perform inverter power supply based on the control signal output by the primary control layer;

[0013] S2 is based on the inverse optimal control method combined with a recursive high-order neural network to construct a basic optimization model;

[0014] S3 constructs a grid-connected optimization model based on the active and reactive power of the microgrid as the input of the optimization model, and constructs an offline optimization model based on the frequency and voltage of the microgrid as the input of the optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain a secondary control layer. The output of the secondary control layer is the optimal current reference value.

[0015] S4 constructs a primary control layer based on the optimal current reference value and the current and voltage state vectors as inputs to the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

[0016] The inverse optimal control method is combined with a recursive high-order neural network to construct a basic optimization model;

[0017] Define a discrete-time dynamical system:

[0018] ;

[0019] In the formula, Let k be the state variable at time k. For control signals, , State variables The relationship between two unknown functions. ;

[0020] The inverse optimal control law is expressed as:

[0021] ;

[0022] In the formula, It is the calculated optimal control law. For a real symmetric positive definite weighted control parameter matrix, State variables A functional relationship, where T represents transpose. For a real symmetric positive definite weighted control parameter matrix, State variables Another functional relationship, It is the error of the state variable at time k.

[0023] For the j-th neuron in the i-th layer of a recursive high-order neural network, the intermediate variable is calculated at time k. :

[0024] ;

[0025] In the formula, Given an M×n dimensional input weight vector, For an M×1 dimensional input scalar, For n×1 dimensional bias terms, the superscript T indicates matrix transpose;

[0026] Perform recursive neuron output calculation based on intermediate variables Recursive state at the previous moment The output of the recurrent neuron at time k is obtained by activation using the hyperbolic tangent function. :

[0027] ;

[0028] In the formula, The recursive weight matrix is ​​n×n dimensional. Let the previous recursive state vector be n×1 dimensional; the hyperbolic tangent function can be expanded into exponential form:

[0029] ;

[0030] Calculate the higher-order product terms, the higher-order product terms of the i-th level. The output of N recurrent neurons in this layer (j=1,2,…,N) Element-wise multiplication yields:

[0031] ;

[0032] In the formula, It is an n×1 dimensional vector. Indicates product operation;

[0033] Perform network output calculation, predicting the output of the recursive high-order neural network at time k. From higher-order product terms Output weight vector and output bias Linear combination yields:

[0034] ;

[0035] In the formula, The output weight vector is n×1 dimensional. For scalar output bias,

[0036] The extended Kalman filter is used to train a recursive high-order neural network until the error of the state variable at time k meets the set requirements. Then the corresponding inverse optimal control law is output.

[0037] Training process: All weights and bias parameters of the recursive high-order neural network are integrated as parameters to be trained into a state vector L, which is used for the state estimation of the extended Kalman filter: L = [ , , , , ];

[0038] ;

[0039] ;

[0040] ;

[0041] in, This represents the error between the actual value and the reference value. , Let k be the reference value of the state variable output by the recursive higher-order neural network at time k. ; The Kalman matrix is... Let R be the covariance matrix, and R be a real matrix. Defined as the partial derivative of the actual state value with respect to the state vector L; based on different parameters in the state vector [ , , , , The influence path on the output is derived in five parts, showing the partial derivatives:

[0042] Partial derivatives of output weight v with output bias c:

[0043] ;

[0044] ;

[0045] Partial derivative of recursive weight u:

[0046] ;

[0047] Partial derivatives of input weights w and biases b:

[0048] ;

[0049] ;

[0050] In summary, the matrix For the concatenation of partial derivatives of each part: .

[0051] In the secondary control layer, the basic optimization model will be... , , , Use respectively , , , By substitution, we obtain the state-space equations of the quadratic control layer:

[0052] ;

[0053] ;

[0054] Off-grid mode secondary control layer state variables ;

[0055] State variables of the secondary control layer in grid-connected mode ;

[0056] In the formula, , The first The state variables of an inverter at time k are its frequency and voltage amplitude. , The first The state variables of active and reactive power of an inverter at time k. It is the first The state variables of the inverter's current d-axis and q-axis components at time k. For the first The voltage amplitude of each inverter at time k, where , The first The d-axis and q-axis voltage components of the inverter; For the second control layer The control signals of each inverter are used in the secondary control layer with the current-optimal control rate. for Output, It is the second control layer. The optimal control law of an inverter at time k. It is a node to be controlled The output vector where All are unknown functions;

[0057] No. The trajectory tracking error of the inverse optimal control of the secondary control layer of the inverter is:

[0058] ;

[0059] In the formula: For nodes Adjacent nodes The output vector, This represents the total number of adjacent nodes. This is the output vector of the microgrid leader node; For adjacency communication gain, For clamping gain;

[0060] Nonlinear mapping For unknown functions, a high-order neural network is used to identify them and map them. have:

[0061] Offline mode:

[0062] ;

[0063] Grid connection mode:

[0064] ;

[0065] in, These are recursive high-order neural network structures in offline and grid-connected modes, respectively.

[0066] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0067] To achieve trajectory tracking of the state-space equations of the quadratic control layer, the inverse optimal control method is adopted, and the optimal control law of the quadratic control layer is determined as follows:

[0068] ;

[0069] In the formula, This is the optimal control law output by the secondary control layer. It is the diagonal control law gain matrix of the second-order control layer. yes transpose, For a real symmetric positive definite weighted control parameter matrix, It is a real symmetric positive definite weighted control parameter matrix. For the first The trajectory tracking error of the inverse optimal control of the secondary control layer of an inverter.

[0070] In a first-level control layer, the basic optimization model is... , , , Use respectively , , , By substitution, we obtain the state-space equation of the first-order control layer:

[0071] ;

[0072] ;

[0073] In the formula: It is the first control layer The optimal control law of an inverter at time k. For the first control layer The state vector of an inverter at time k. , For the first The d-axis and q-axis current components of the inverter , For the first The d-axis and q-axis voltage components of the inverter , The d-axis and q-axis reference currents are input from the secondary control layer to the primary control layer. It is the output to be controlled, where All are unknown functions; For the first control layer The control signals for each inverter are controlled in the primary control layer with the optimal voltage control rate. for Output;

[0074] No. The trajectory tracking error of the inverse optimal control of the primary control layer of the inverter is:

[0075] ;

[0076] In the formula, The optimal control law input to the secondary control layer is the d-axis and q-axis reference current sent from the secondary control layer to the primary control layer.

[0077] Nonlinear mapping Unknown entities are identified using high-order neural networks; mappings are then performed. have:

[0078] ;

[0079] in, A recursive high-order neural network structure with a single control layer;

[0080] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0081] To achieve trajectory tracking of the state-space equations of the primary control layer, an inverse optimal control method is employed. The optimal control law of the primary control layer is as follows:

[0082] ;

[0083] In the formula, To achieve optimal control, the obtained d-axis and q-axis voltage components are voltage synthesized and SVPWM pulse modulated to obtain a modulation signal, which is then input to the inverter. It is the diagonal control law gain matrix of the first-order control layer. yes transpose, For real symmetric positive definite weighted control parameters, It is a real symmetric positive definite weighted control parameter. For the first The trajectory tracking error of the inverse optimal control of the primary control layer of an inverter.

[0084] A multi-unit photovoltaic power generation system is constructed, which includes multiple inverters with the same circuit structure. Each inverter has a common connection point for grid connection, and the control signal input terminal of the inverter is used to receive the control signal output by the primary control layer.

[0085] Secondly, the present invention provides a photovoltaic scheduling optimization system based on inverse optimal control and recursive high-order neural network. The system is used to execute the aforementioned photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network, specifically including: a multi-machine photovoltaic power generation system construction module, a basic optimization model construction module, a secondary control layer construction module, and a primary control layer construction module.

[0086] Multi-machine photovoltaic power generation system construction module: used to construct a multi-machine photovoltaic power generation system, which is used to perform inverter power supply based on the control signal output by the primary control layer;

[0087] Basic Optimization Model Construction Module: Used to construct basic optimization models based on inverse optimal control methods combined with recursive high-order neural networks;

[0088] Secondary control layer construction module: used to construct a grid-connected optimization model based on the input of the microgrid active power and reactive power optimization model, and to construct an offline optimization model based on the input of the microgrid frequency and voltage optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain the secondary control layer. The output of the secondary control layer is the optimal current reference value.

[0089] Primary control layer construction module: used to construct a primary control layer based on the optimal current reference value and current and voltage state vectors as inputs of the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

[0090] Thirdly, the present invention provides a photovoltaic scheduling optimization device based on inverse optimal control and recursive high-order neural network, including a memory and a processor, wherein the memory is used to store computer program code and transmit the computer program code to the processor;

[0091] The processor is configured to execute the aforementioned photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks according to the instructions in the computer program code.

[0092] Fourthly, the present invention provides a computer program product, including a computer program that is executed by a processor of the aforementioned photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks.

[0093] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0094] 1. The present invention proposes a photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network, which designs a neural inverse optimal distributed cooperative control strategy to control the voltage, frequency and power output of the distributed photovoltaic system, thereby improving the stability and continuity of power supply.

[0095] 2. This invention proposes a novel neural inverse optimal distributed cooperative secondary control scheme in a photovoltaic dispatch optimization method based on inverse optimal control and recursive high-order neural networks. The scheme consists of a secondary control layer, a primary control layer, and a physical layer. The secondary control layer defines the current trajectory used as the input reference signal for the primary control layer. Its logic is as follows: in grid-connected mode, the current trajectory is determined based on active and reactive power; in off-grid mode, the current trajectory is determined based on grid voltage and frequency. The obtained current trajectory is used by the primary control layer to derive the voltage control signal for the grid-connected inverter.

[0096] 3. The distributed secondary control scheme proposed in the photovoltaic dispatch optimization method based on inverse optimal control and recursive high-order neural network of the present invention establishes the unknown model of DG based on the extended Kalman filter (EKF) trained recursive high-order neural network (RHONN). It can estimate the oscillation of DG current, microgrid voltage and frequency during the switching process, and reduce their influence through inverse optimal controller, thereby ensuring the smooth switching.

[0097] 4. The photovoltaic scheduling optimization system based on inverse optimal control and recursive high-order neural networks of the present invention includes a multi-machine photovoltaic power generation system construction module, a basic optimization model construction module, a secondary control layer construction module, and a primary control layer construction module. This system is used to implement the steps of the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks provided in any of the above technical solutions. Therefore, this system simultaneously includes all the beneficial effects of the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks provided in any of the above technical solutions, which will not be elaborated further here.

[0098] 5. The photovoltaic scheduling optimization device based on inverse optimal control and recursive high-order neural network of the present invention includes a processor and a memory. The memory is used to store computer program code and transmit the computer program code to the processor. The processor is used to execute the method provided in any of the above-mentioned technical solutions according to the instructions in the computer program code. Therefore, this device simultaneously includes all the beneficial effects of the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network provided in any of the above-mentioned technical solutions, which will not be repeated here.

[0099] 6. The present invention provides a computer program product, which, when executed by a processor, implements the steps of the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks provided in any of the above-described technical solutions. Therefore, this computer program product simultaneously includes all the beneficial effects of the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks provided in any of the above-described technical solutions, which will not be elaborated further here. Attached Figure Description

[0100] Figure 1 This is a flowchart of the steps of the present invention.

[0101] Figure 2 This is a diagram of the neural inverse optimal distributed cooperative control scheme of the present invention.

[0102] Figure 3 This is a control strategy diagram of the method of the present invention.

[0103] Figure 4 This is a schematic diagram of the recursive high-order neural network operation process of the secondary control layer.

[0104] Figure 5 This is a schematic diagram of a recursive high-order neural network operation process in the control layer.

[0105] Figure 6 This is a system schematic diagram of the present invention.

[0106] Figure 7 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0107] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0108] Example 1:

[0109] See Figures 1 to 5 A photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks, the scheduling method comprising the following steps:

[0110] S1 constructs a multi-machine photovoltaic power generation system, which is used to perform inverter power supply based on the control signal output by the primary control layer;

[0111] S2 is based on the inverse optimal control method combined with a recursive high-order neural network to construct a basic optimization model;

[0112] S3 constructs a grid-connected optimization model based on the active and reactive power of the microgrid as the input of the optimization model, and constructs an offline optimization model based on the frequency and voltage of the microgrid as the input of the optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain a secondary control layer. The output of the secondary control layer is the optimal current reference value.

[0113] S4 constructs a primary control layer based on the optimal current reference value and the current and voltage state vectors as inputs to the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

[0114] Construct a multi-inverter photovoltaic power generation system, which includes multiple inverters with the same circuit structure. Each inverter has a common connection point connected to the grid and is used to receive control signals output from the primary control layer.

[0115] The inverter consists of a voltage source converter connected in series with a filter circuit.

[0116] See Figures 1 to 3 , , The first The q-axis and d-axis components of the inverter output voltage , The first The q-axis and d-axis components of the inverter output current. , The first The active power and reactive power output of each inverter. , , The first The frequency, angular frequency, and voltage amplitude of each inverter; among which... The calculation formula is:

[0117] (1);

[0118] (2);

[0119] , The first The reference values ​​of the q and d-axis voltages output by the primary control layer; , The first The reference values ​​for the q-axis and d-axis currents output by the secondary control layer. , These are the active power and reactive power predicted by the recursive high-order neural network, respectively. , These represent the frequency and voltage amplitude predicted by the recursive high-order neural network, respectively. , These are the q-axis and d-axis currents predicted by a high-order neural network, respectively. These are the weight parameters in a high-order neural network.

[0120] The inverse optimal control method is combined with a recursive high-order neural network to construct a basic optimization model;

[0121] Define a discrete-time dynamical system:

[0122] (3);

[0123] In the formula, Let k be the state variable at time k. For control signals, , State variables The relationship between two unknown functions. ;

[0124] Using optimization theory, the cost function is defined by the HJB equation (Hamilton-Jacobi-Bellman equation):

[0125] (4);

[0126] In the formula, It is the error of the state variable at time k. The actual value of the state variable at time k. Let be the expected value of the state variable at time k; It is a positive semi-definite function. For real symmetric positive definite weighted control parameters, For the transpose of the control signal; when In the optimal time, It will be used as a Lyapunov function; in order to satisfy the Bellman principle. Set as a time-invariant function:

[0127] (5);

[0128] In the formula, It is a time-invariant function at time k+1;

[0129] We obtain the Hamiltonian expression:

[0130] (6);

[0131] A necessary condition for an optimal control law to satisfy is: partial derivatives. The solution to this equation is not easy to find;

[0132] Using an alternative approach, based on inverse optimal control of the Lyapunov control function, the Lyapunov control function (CLF) is chosen as follows:

[0133] (7);

[0134] In the formula, For the transpose of the error, For real symmetric positive definite weighted control parameters, and ;

[0135] Substitute the Lyapunov control function into the Hamiltonian expression, and then... Taking the partial derivative, the inverse optimal control law is obtained as follows:

[0136] (8);

[0137] In the formula, It is the calculated optimal control law. For a real symmetric positive definite weighted control parameter matrix, State variables A functional relationship, where T represents transpose. For a real symmetric positive definite weighted control parameter matrix, State variables Another functional relationship, It is the error of the state variable at time k.

[0138] See Figure 4 , Figure 5 To ensure stable control, dynamic power regulation, and anti-interference operation of microgrids under various operating conditions, including grid-connected and off-grid operation, this invention proposes a method of training a recurrent high-order neural network (EKF-RHONN) using an extended Kalman filter. For the j-th neuron in the i-th layer of the recurrent high-order neural network, intermediate variables are calculated at time k. :

[0139] (20);

[0140] In the formula, Given an M×n dimensional input weight vector, For an M×1 dimensional input scalar, For n×1 dimensional bias terms, the superscript T indicates matrix transpose;

[0141] Perform recursive neuron output calculation based on intermediate variables Recursive state at the previous moment The output of the recurrent neuron at time k is obtained by activation using the hyperbolic tangent function. :

[0142] (twenty one);

[0143] In the formula, The recursive weight matrix is ​​n×n dimensional. Let the previous recursive state vector be n×1 dimensional; the hyperbolic tangent function can be expanded into exponential form:

[0144] (twenty two);

[0145] Calculate the higher-order product terms, the higher-order product terms of the i-th level. The output of N recurrent neurons in this layer (j=1,2,…,N) Element-wise multiplication yields:

[0146] (twenty three);

[0147] In the formula, It is an n×1 dimensional vector. Indicates product operation;

[0148] Perform network output calculation, predicting the output of the recursive high-order neural network at time k. From higher-order product terms Output weight vector and output bias Linear combination yields:

[0149] (twenty four);

[0150] In the formula, The output weight vector is n×1 dimensional. For scalar output bias, In off-network mode , In grid connection mode ;

[0151] Therefore, the overall computation process of the recursive high-order neural network is as follows:

[0152] (25);

[0153] The extended Kalman filter is used to train a recursive high-order neural network until the error of the state variable at time k meets the set requirements. Then the corresponding inverse optimal control law is output.

[0154] Training process: The weights and biases of the recurrent higher-order neural network are used as the state variables of the extended Kalman filter. All weights and bias parameters of the recurrent higher-order neural network are integrated as parameters to be trained into a state vector L, which is used for state estimation of the extended Kalman filter: L = [ , , , , ];

[0155] (26);

[0156] (27);

[0157] (28);

[0158] in, This represents the error between the actual value and the reference value. , Let k be the reference value of the state variable output by the recursive higher-order neural network at time k. ; The Kalman matrix is... Let R be the covariance matrix, and R be a real matrix. Defined as the partial derivative of the actual state value with respect to the state vector L; based on different parameters in the state vector [ , , , , The influence path on the output is derived in five parts, showing the partial derivatives:

[0159] Partial derivatives of output weight v with output bias c:

[0160] (29);

[0161] (30);

[0162] Partial derivative of recursive weight u:

[0163] (31);

[0164] Partial derivatives of input weights w and biases b:

[0165] (32);

[0166] (33);

[0167] In summary, the matrix For the concatenation of partial derivatives of each part: .

[0168] In the secondary control layer, the basic optimization model will be... , , , Use respectively , , , By substitution, we obtain the state-space equations of the quadratic control layer:

[0169] (9);

[0170] (10);

[0171] Assume that the frequency and voltage of the microgrid in off-grid mode have unknown discrete-time mathematical models, and the active power and reactive power in grid-connected mode have unknown discrete-time mathematical models.

[0172] Off-grid mode secondary control layer state variables ;

[0173] State variables of the secondary control layer in grid-connected mode ;

[0174] In the formula, , The first The state variables of an inverter at time k are its frequency and voltage amplitude. , The first The state variables of active and reactive power of an inverter at time k. It is the first The state variables of the inverter's current d-axis and q-axis components at time k. For the first The voltage amplitude of each inverter at time k, where , The first The d-axis and q-axis voltage components of the inverter; For the second control layer The control signals of each inverter are used in the secondary control layer with the current-optimal control rate. for Output, It is the second control layer. The optimal control law of an inverter at time k. It is a node to be controlled The output vector where All are unknown functions;

[0175] No. The trajectory tracking error of the inverse optimal control of the secondary control layer of the inverter is:

[0176] (11);

[0177] In the formula: For nodes Adjacent nodes The output vector (obtained through inter-module communication). This represents the total number of adjacent nodes. This is the output vector of the microgrid leader node; For adjacency communication gain, For clamping gain;

[0178] Nonlinear mapping For unknown functions, a high-order neural network is used to identify them and map them. have:

[0179] Offline mode:

[0180] (12);

[0181] Grid connection mode:

[0182] (13);

[0183] in, These are recursive high-order neural network structures in offline and grid-connected modes, respectively.

[0184] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0185] To achieve trajectory tracking of the state-space equations of the quadratic control layer, the inverse optimal control method is adopted, and the optimal control law of the quadratic control layer is determined as follows:

[0186] (14);

[0187] In the formula, This is the optimal control law output by the secondary control layer. It is the diagonal control law gain matrix of the second-order control layer. yes transpose, For a real symmetric positive definite weighted control parameter matrix, It is a real symmetric positive definite weighted control parameter matrix. For the first The trajectory tracking error of the inverse optimal control of the secondary control layer of an inverter.

[0188] In a first-level control layer, the basic optimization model is... , , , Use respectively , , , By substitution, we obtain the state-space equation of the first-order control layer:

[0189] (15);

[0190] (16);

[0191] In the formula: It is the first control layer The optimal control law of an inverter at time k. For the first control layer The state vector of an inverter at time k. , For the first The d-axis and q-axis current components of the inverter , For the first The d-axis and q-axis voltage components of the inverter , The d-axis and q-axis reference currents are input from the secondary control layer to the primary control layer. It is the output to be controlled, where All are unknown functions; For the first control layer The control signals for each inverter are controlled in the primary control layer with the optimal voltage control rate. for Output;

[0192] No. The trajectory tracking error of the inverse optimal control of the primary control layer of the inverter is:

[0193] (17);

[0194] In the formula, The optimal control law input to the secondary control layer is the d-axis and q-axis reference current sent from the secondary control layer to the primary control layer.

[0195] Nonlinear mapping Unknown entities are identified using high-order neural networks; mappings are then performed. have:

[0196] (18);

[0197] in, A recursive high-order neural network structure with a single control layer;

[0198] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0199] To achieve trajectory tracking of the state-space equations of the primary control layer, an inverse optimal control method is employed. The optimal control law of the primary control layer is as follows:

[0200] (19);

[0201] In the formula, To achieve optimal control, the obtained d-axis and q-axis voltage components are voltage synthesized and SVPWM pulse modulated to obtain a modulation signal, which is then input to the inverter. It is the diagonal control law gain matrix of the first-order control layer. yes transpose, For real symmetric positive definite weighted control parameters, It is a real symmetric positive definite weighted control parameter. For the first The trajectory tracking error of the inverse optimal control of the primary control layer of an inverter.

[0202] Example 2:

[0203] See Figure 6 A photovoltaic scheduling optimization system based on inverse optimal control and recursive high-order neural network specifically includes: a multi-machine photovoltaic power generation system construction module, a basic optimization model construction module, a secondary control layer construction module, and a primary control layer construction module;

[0204] Multi-machine photovoltaic power generation system construction module: used to construct a multi-machine photovoltaic power generation system, which is used to perform inverter power supply based on the control signal output by the primary control layer;

[0205] A multi-unit photovoltaic power generation system is constructed, which includes multiple inverters with the same circuit structure. Each inverter has a common connection point for grid connection, and the control signal input terminal of the inverter is used to receive the control signal output by the primary control layer.

[0206] Basic Optimization Model Construction Module: Used to construct basic optimization models based on inverse optimal control methods combined with recursive high-order neural networks;

[0207] The inverse optimal control method is combined with a recursive high-order neural network to construct a basic optimization model;

[0208] Define a discrete-time dynamical system:

[0209] ;

[0210] In the formula, Let k be the state variable at time k. For control signals, , State variables The relationship between two unknown functions. ;

[0211] The inverse optimal control law is expressed as:

[0212] ;

[0213] In the formula, It is the calculated optimal control law. For a real symmetric positive definite weighted control parameter matrix, State variables A functional relationship, where T represents transpose. For a real symmetric positive definite weighted control parameter matrix, State variables Another functional relationship, It is the error of the state variable at time k.

[0214] For the j-th neuron in the i-th layer of a recursive high-order neural network, the intermediate variable is calculated at time k. :

[0215] ;

[0216] In the formula, Given an M×n dimensional input weight vector, For an M×1 dimensional input scalar, For n×1 dimensional bias terms, the superscript T indicates matrix transpose;

[0217] Perform recursive neuron output calculation based on intermediate variables Recursive state at the previous moment The output of the recurrent neuron at time k is obtained by activation using the hyperbolic tangent function. :

[0218] ;

[0219] In the formula, The recursive weight matrix is ​​n×n dimensional. Let the previous recursive state vector be n×1 dimensional; the hyperbolic tangent function can be expanded into exponential form:

[0220] ;

[0221] Calculate the higher-order product terms, the higher-order product terms of the i-th level. The output of N recurrent neurons in this layer (j=1,2,…,N) Element-wise multiplication yields:

[0222] ;

[0223] In the formula, It is an n×1 dimensional vector. Indicates product operation;

[0224] Perform network output calculation, predicting the output of the recursive high-order neural network at time k. From higher-order product terms Output weight vector and output bias Linear combination yields:

[0225] ;

[0226] In the formula, The output weight vector is n×1 dimensional. For scalar output bias,

[0227] The extended Kalman filter is used to train a recursive high-order neural network until the error of the state variable at time k meets the set requirements. Then the corresponding inverse optimal control law is output.

[0228] Training process: All weights and bias parameters of the recursive high-order neural network are integrated as parameters to be trained into a state vector L, which is used for the state estimation of the extended Kalman filter: L = [ , , , , ];

[0229] ;

[0230] ;

[0231] ;

[0232] in, This represents the error between the actual value and the reference value. , Let k be the reference value of the state variable output by the recursive higher-order neural network at time k. ; The Kalman matrix is... Let R be the covariance matrix, and R be a real matrix. Defined as the partial derivative of the actual state value with respect to the state vector L; based on different parameters in the state vector [ , , , , The influence path on the output is derived in five parts, showing the partial derivatives:

[0233] Partial derivatives of output weight v with output bias c:

[0234] ;

[0235] ;

[0236] Partial derivative of recursive weight u:

[0237] ;

[0238] Partial derivatives of input weights w and biases b:

[0239] ;

[0240] ;

[0241] In summary, the matrix For the concatenation of partial derivatives of each part: .

[0242] Secondary control layer construction module: used to construct a grid-connected optimization model based on the input of the microgrid active power and reactive power optimization model, and to construct an offline optimization model based on the input of the microgrid frequency and voltage optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain the secondary control layer. The output of the secondary control layer is the optimal current reference value.

[0243] In the secondary control layer, the basic optimization model will be... , , , Use respectively , , , By substitution, we obtain the state-space equations of the quadratic control layer:

[0244] ;

[0245] ;

[0246] Off-grid mode secondary control layer state variables ;

[0247] State variables of the secondary control layer in grid-connected mode ;

[0248] In the formula, , The first The state variables of an inverter at time k are its frequency and voltage amplitude. , The first The state variables of active and reactive power of an inverter at time k. It is the first The state variables of the inverter's current d-axis and q-axis components at time k. For the first The voltage amplitude of each inverter at time k, where , The first The d-axis and q-axis voltage components of the inverter; For the second control layer The control signals of each inverter are used in the secondary control layer with the current-optimal control rate. for Output, It is the second control layer. The optimal control law of an inverter at time k. It is a node to be controlled The output vector where All are unknown functions;

[0249] No. The trajectory tracking error of the inverse optimal control of the secondary control layer of the inverter is:

[0250] ;

[0251] In the formula: For nodes Adjacent nodes The output vector, This represents the total number of adjacent nodes. This is the output vector of the microgrid leader node; For adjacency communication gain, For clamping gain;

[0252] Nonlinear mapping For unknown functions, a high-order neural network is used to identify them and map them. have:

[0253] Offline mode:

[0254] ;

[0255] Grid connection mode:

[0256] ;

[0257] in, These are recursive high-order neural network structures in offline and grid-connected modes, respectively.

[0258] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0259] To achieve trajectory tracking of the state-space equations of the quadratic control layer, the inverse optimal control method is adopted, and the optimal control law of the quadratic control layer is determined as follows:

[0260] ;

[0261] In the formula, This is the optimal control law output by the secondary control layer. It is the diagonal control law gain matrix of the second-order control layer. yes transpose, For a real symmetric positive definite weighted control parameter matrix, It is a real symmetric positive definite weighted control parameter matrix. For the first The trajectory tracking error of the inverse optimal control of the secondary control layer of an inverter.

[0262] Primary control layer construction module: used to construct a primary control layer based on the optimal current reference value and current and voltage state vectors as inputs of the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

[0263] In a first-level control layer, the basic optimization model is... , , , Use respectively , , , By substitution, we obtain the state-space equation of the first-order control layer:

[0264] ;

[0265] ;

[0266] In the formula: It is the first control layer The optimal control law of an inverter at time k. For the first control layer The state vector of an inverter at time k. , For the first The d-axis and q-axis current components of the inverter , For the first The d-axis and q-axis voltage components of the inverter , The d-axis and q-axis reference currents are input from the secondary control layer to the primary control layer. It is the output to be controlled, where All are unknown functions; For the first control layer The control signals for each inverter are controlled in the primary control layer with the optimal voltage control rate. for Output;

[0267] No. The trajectory tracking error of the inverse optimal control of the primary control layer of the inverter is:

[0268] ;

[0269] In the formula, The optimal control law input to the secondary control layer is the d-axis and q-axis reference current sent from the secondary control layer to the primary control layer.

[0270] Nonlinear mapping Unknown entities are identified using high-order neural networks; mappings are then performed. have:

[0271] ;

[0272] in, A recursive high-order neural network structure with a single control layer;

[0273] Mapping Yes: will Set as control law gain matrix Let be a constant;

[0274] To achieve trajectory tracking of the state-space equations of the primary control layer, an inverse optimal control method is employed. The optimal control law of the primary control layer is as follows:

[0275] ;

[0276] In the formula, To achieve optimal control, the obtained d-axis and q-axis voltage components are voltage synthesized and SVPWM pulse modulated to obtain a modulation signal, which is then input to the inverter. It is the diagonal control law gain matrix of the first-order control layer. yes transpose, For real symmetric positive definite weighted control parameters, It is a real symmetric positive definite weighted control parameter. For the first The trajectory tracking error of the inverse optimal control of the primary control layer of an inverter.

[0277] Example 3:

[0278] See Figure 7 A photovoltaic scheduling optimization device based on inverse optimal control and recursive high-order neural network, characterized in that it includes a memory and a processor, wherein the memory is used to store computer program code and transmit the computer program code to the processor;

[0279] The processor is configured to execute the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network described in Embodiment 1 according to the instructions in the computer program code.

[0280] Example 4:

[0281] A computer program product includes a computer program that is executed by a processor of the aforementioned photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks.

Claims

1. A photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks, characterized in that: The scheduling method includes the following steps: S1 constructs a multi-machine photovoltaic power generation system, which is used to perform inverter power supply based on the control signal output by the primary control layer; S2 is based on the inverse optimal control method combined with a recursive high-order neural network to construct a basic optimization model; S3 constructs a grid-connected optimization model based on the active and reactive power of the microgrid as the input of the optimization model, and constructs an offline optimization model based on the frequency and voltage of the microgrid as the input of the optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain a secondary control layer. The output of the secondary control layer is the optimal current reference value. S4 constructs a primary control layer based on the optimal current reference value and the current and voltage state vectors as inputs to the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

2. The photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network according to claim 1, characterized in that: The inverse optimal control method is combined with a recursive high-order neural network to construct a basic optimization model; Define a discrete-time dynamical system: ； In the formula, Let k be the state variable at time k. For control signals, , State variables The relationship between two unknown functions. ; The inverse optimal control law is expressed as: ； In the formula, It is the calculated optimal control law. For a real symmetric positive definite weighted control parameter matrix, State variables A functional relationship, where T represents transpose. For a real symmetric positive definite weighted control parameter matrix, State variables Another functional relationship, It is the error of the state variable at time k.

3. The photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network according to claim 1, characterized in that: For the j-th neuron in the i-th layer of a recursive high-order neural network, the intermediate variable is calculated at time k. : ； In the formula, Given an M×n dimensional input weight vector, For an M×1 dimensional input scalar, For n×1 dimensional bias terms, the superscript T indicates matrix transpose; Perform recursive neuron output calculation based on intermediate variables Recursive state at the previous moment The output of the recurrent neuron at time k is obtained by activation using the hyperbolic tangent function. : ； In the formula, The recursive weight matrix is ​​n×n dimensional. Let the previous recursive state vector be n×1 dimensional; the hyperbolic tangent function can be expanded into exponential form: ； Calculate the higher-order product terms, the higher-order product terms of the i-th level. The output of N recurrent neurons in this layer (j=1,2,…,N) Element-wise multiplication yields: ； In the formula, It is an n×1 dimensional vector. Indicates product operation; Perform network output calculation, predicting the output of the recursive high-order neural network at time k. From higher-order product terms Output weight vector and output bias Linear combination yields: ； In the formula, The output weight vector is n×1 dimensional. For scalar output bias, The extended Kalman filter is used to train a recursive high-order neural network until the error of the state variable at time k meets the set requirements. Then the corresponding inverse optimal control law is output.

4. The photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network according to claim 3, characterized in that: Training process: All weights and bias parameters of the recursive high-order neural network are integrated as parameters to be trained into a state vector L, which is used for the state estimation of the extended Kalman filter: L=[ , , , , ]; ； ； ； in, This represents the error between the actual value and the reference value. , Let k be the reference value of the state variable output by the recursive high-order neural network at time k. ; The Kalman matrix is... Let R be the covariance matrix, and R be a real matrix. Defined as the partial derivative of the actual state value with respect to the state vector L; based on different parameters in the state vector [ , , , , The influence path on the output is derived in five parts, showing the partial derivatives: Partial derivatives of output weight v with output bias c: ； ； Partial derivative of recursive weight u: ； Partial derivatives of input weights w and biases b: ； ； In summary, the matrix For the concatenation of partial derivatives of each part: .

5. The photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network according to claim 1, characterized in that: In the secondary control layer, the basic optimization model will be... , , , Use respectively , , , By substitution, we obtain the state-space equations of the quadratic control layer: ； ； Off-grid mode secondary control layer state variables ; State variables of the secondary control layer in grid-connected mode ; In the formula, , The first The state variables of an inverter at time k are its frequency and voltage amplitude. , The first The state variables of active and reactive power of an inverter at time k. It is the first The state variables of the inverter's current d-axis and q-axis components at time k. For the first The voltage amplitude of each inverter at time k, where , The first The d-axis and q-axis voltage components of the inverter; For the second control layer The control signals of each inverter are used in the secondary control layer with the current-optimal control rate. for Output, It is the second control layer. The optimal control law of an inverter at time k. It is a node to be controlled The output vector where All are unknown functions; No. The trajectory tracking error of the inverse optimal control of the secondary control layer of the inverter is: ； In the formula: For nodes Adjacent nodes The output vector, This represents the total number of adjacent nodes. This is the output vector of the microgrid leader node; For adjacency communication gain, For clamping gain; Nonlinear mapping For unknown functions, a high-order neural network is used to identify them and map them. have: Offline mode: ； Grid connection mode: ； in, These are recursive high-order neural network structures in offline and grid-connected modes, respectively. Mapping Yes: will Set as control law gain matrix Let be a constant; To achieve trajectory tracking of the state-space equations of the quadratic control layer, the inverse optimal control method is adopted, and the optimal control law of the quadratic control layer is determined as follows: ； In the formula, This is the optimal control law output by the secondary control layer. It is the diagonal control law gain matrix of the second-order control layer. yes transpose, For a real symmetric positive definite weighted control parameter matrix, It is a real symmetric positive definite weighted control parameter matrix. For the first The trajectory tracking error of the inverse optimal control of the secondary control layer of an inverter.

6. The photovoltaic scheduling optimization method based on inverse optimal control optimization recursive high-order neural network according to claim 1, characterized in that: In a first-level control layer, the basic optimization model is... , , , Use respectively , , , By substitution, we obtain the state-space equation of the first-order control layer: ； ； In the formula: It is the first control layer The optimal control law of an inverter at time k. For the first control layer The state vector of an inverter at time k. , For the first The d-axis and q-axis current components of the inverter , For the first The d-axis and q-axis voltage components of the inverter , The d-axis and q-axis reference currents are input from the secondary control layer to the primary control layer. It is the output to be controlled, where All are unknown functions; For the first control layer The control signals for each inverter are controlled in the primary control layer with the optimal voltage control rate. for Output; No. The trajectory tracking error of the inverse optimal control of the primary control layer of the inverter is: ； In the formula, The optimal control law input to the secondary control layer is the d-axis and q-axis reference current sent from the secondary control layer to the primary control layer. Nonlinear mapping Unknown entities are identified using high-order neural networks; mappings are then performed. have: ； in, A recursive high-order neural network structure with a single control layer; Mapping Yes: will Set as control law gain matrix Let be a constant; To achieve trajectory tracking of the state-space equations of the primary control layer, an inverse optimal control method is employed. The optimal control law of the primary control layer is as follows: ； In the formula, To achieve optimal control, the obtained d-axis and q-axis voltage components are voltage synthesized and SVPWM pulse modulated to obtain a modulation signal, which is then input to the inverter. It is the diagonal control law gain matrix of the first-order control layer. yes transpose, For real symmetric positive definite weighted control parameters, It is a real symmetric positive definite weighted control parameter. For the first The trajectory tracking error of the inverse optimal control of the primary control layer of an inverter.

7. The photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network according to claim 1, characterized in that: A multi-unit photovoltaic power generation system is constructed, which includes multiple inverters with the same circuit structure. Each inverter has a common connection point for grid connection, and the control signal input terminal of the inverter is used to receive the control signal output by the primary control layer.

8. A photovoltaic scheduling optimization system based on inverse optimal control and recursive high-order neural networks, characterized in that, The system is used to execute the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network as described in any one of claims 1 to 7, specifically including: a multi-machine photovoltaic power generation system construction module, a basic optimization model construction module, a secondary control layer construction module, and a primary control layer construction module; Multi-machine photovoltaic power generation system construction module: used to construct a multi-machine photovoltaic power generation system, which is used to provide inverter power supply based on the control signals output by the primary control layer; Basic Optimization Model Construction Module: Used to construct basic optimization models based on inverse optimal control methods combined with recursive high-order neural networks; Secondary control layer construction module: used to construct a grid-connected optimization model based on the input of the microgrid active power and reactive power optimization model, and to construct an offline optimization model based on the input of the microgrid frequency and voltage optimization model. The grid-connected optimization model and the offline optimization model are combined to obtain the secondary control layer. The output of the secondary control layer is the optimal current reference value. Primary control layer construction module: used to construct a primary control layer based on the optimal current reference value and current and voltage state vectors as inputs of the optimization model. The primary control layer outputs control signals to the multi-machine photovoltaic power generation system.

9. A photovoltaic scheduling optimization device based on inverse optimal control and recursive high-order neural networks, characterized in that, It includes a memory and a processor, wherein the memory is used to store computer program code and transfer the computer program code to the processor; The processor is configured to execute the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural network as described in any one of claims 1 to 7, according to the instructions in the computer program code.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor using the photovoltaic scheduling optimization method based on inverse optimal control and recursive high-order neural networks as described in any one of claims 1 to 7.

Images