Converter load type identification method based on convolutional neural network

Abstract

The invention discloses a converter load type identification method based on a convolutional neural network. The method comprises the following steps: determining state variables participating in converter load type identification; acquiring state variable data of converter operation; and analyzing the load type of the converter by using a convolutional neural network. According to the method, system state variables reflecting the load type of the converter are firstly screened, and then the load type of the converter is identified by using the convolutional neural network. The designed method can quickly and accurately identify the load type of the converter, and has the advantage of strong anti-interference performance.

Description

Technical Field

[0001] The invention relates to a method for identifying a load type of a converter based on a convolutional neural network. Background Art

[0002] The converter model predictive control algorithm is the main form of the current converter optimization control algorithm. In the design and optimization process of the converter model predictive control algorithm, the online calculation amount of the algorithm is the main factor that constrains the further optimization of the algorithm. In order to solve the contradiction between the design of the model predictive control algorithm and the calculation amount of the algorithm, the industry has proposed a converter model predictive control algorithm based on a neural network, which uses a neural network to fit the external characteristics of the model predictive control algorithm and calculates the optimal control amount of the converter online based on the neural network.

[0003] In the neural network-based converter model predictive control algorithm, the control performance of the system depends on the neural network control model adopted. Different neural network control models correspond to different loads of the converter system. Therefore, in order to improve the adaptability of the neural network control model, it is necessary to identify the load type of the converter in a timely manner.

[0004] In the traditional load type identification method based on the relationship between voltage and current, at least one cycle of voltage and current signals needs to be obtained, which has problems such as complex calculation and low precision, and is easily interfered. Based on this, in order to quickly and accurately identify the load type of the converter, this scheme first screens the system state variables that reflect the load type of the converter, and then calculates the sampled state variables through a convolutional neural network to obtain the load type of the converter. The designed method has the advantages of quickly and accurately identifying the load type of the converter and strong anti-interference. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for identifying converter loads using a convolutional neural network.

[0006] The technical solution of the present invention to solve the above problem is: a method for identifying the load type of a converter based on a convolutional neural network, comprising the following steps: S1. Determine the state variables involved in identifying the converter load type; S2. Obtaining state variable data of the converter operation; S3. Analysis of converter load type using convolutional neural network.

[0007] The effect of the present invention is that: for the identification problem of the converter load type in the converter model prediction control system based on neural network, compared with the traditional method, the present invention provides a method that can quickly and accurately identify the converter load type and has strong anti-interference ability. The advantages of this method are: the principal component analysis method is used to determine the converter state variables involved in the load type identification, and a convolutional neural network is constructed to analyze the converter load type, which improves the accuracy of the converter load type identification, solves the problem of online identification of the converter load type, and provides a basis for further improving the performance of the neural network control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 Flow chart of the load identification method in the present invention.

[0009] Figure 2 Structure diagram of the converter model predictive control system based on neural network in the present invention.

[0010] Figure 3 Structural diagram of a three-phase voltage-type inverter with an LC filter in the present invention.

[0011] Figure 4 A structural diagram of the convolutional neural network in the present invention.

[0012] Figure 5 Confusion matrix diagram of the convolutional neural network classification calculation in the present invention.

[0013] Figure 6 Figure 2. Result diagram of load type recognition based on convolutional neural network in the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] The present invention provides a method for identifying the load type of a converter based on a convolutional neural network, and its flow chart is as follows: Figure 1 As shown, it includes the following steps: S1. Determine the state variables involved in the identification of the converter load type.

[0016] The inverter load type identification refers to the inverter load type determining the neural network control model used in the neural network-based inverter model predictive control system. Timely and accurate identification of the load type can improve the adaptability and reliability of the control system. Figure 2 The figure shows the structure of the neural network-based converter model predictive control system, which provides information for switching the neural network control model by online identifying the converter load type. Figure 2In the example, the load types that can be identified include: no load, resistive-inductive light load, resistive-inductive heavy load, and nonlinear load, which are numbered 1, 2, 3, and 4 respectively, and are used to switch the corresponding neural network control model parameters.

[0017] The state variables refer to the state quantities of the converter control system. For a three-phase voltage-source inverter with an LC filter, the state variables include load current Io, output voltage Uo, inductor current Is, output voltage control error dUo, etc. Figure 3 The figure shows a three-phase voltage-type inverter with an LC filter. In this embodiment, the control target of the three-phase inverter is to achieve tracking control of its output voltage.

[0018] The state variables involved in the identification of the load type of the converter refer to system state variables that change significantly when the load type of the converter changes. The state variables involved in the identification of the load type of the converter refer to analyzing the correlation between the system state variables and the load type using the principal component analysis method to obtain state variables that are strongly correlated with the load type. For a three-phase voltage inverter with an LC filter, the state variables involved in the identification of the load type obtained by principal component analysis are: load current Io, output voltage Uo, and inductor current Is.

[0019] S2. Obtain the state variable data of the converter operation.

[0020] The state variable data of the inverter operation is to obtain the state variable data involved in the identification of the inverter load type. For a three-phase voltage inverter with an LC filter, the state variable data obtained include: three-phase load current, three-phase output voltage, and three-phase inductor current. The state variable data is to obtain the three-phase load current, output voltage, and inductor current at a certain time k by sampling, and coordinate transformation of the three-phase signal to obtain α - β The specific transformation process of the vector in the coordinate system is: x = x r + jx i =2 / 3[1, a , a 2 ]·[ x A , x B , x C ] T (1) In the formula, x r , xi is a state variable x After the coordinate transformation, the real and imaginary parts are x A , x B , x C is a state variable x The ABC three-phase components, operator a =exp( j ·2π / 3).

[0021] For a three-phase voltage-source inverter with an LC filter, the system state variables at time k obtained through sampling and coordinate transformation are: the real part of the load current Ior(k), the imaginary part of the load current Ioi(k), the real part of the output voltage Uor(k), the imaginary part of the output voltage Uoi(k), the real part of the inductor current Isr(k), and the imaginary part of the inductor current Isi(k).

[0022] S3. Analysis of converter load type using convolutional neural network.

[0023] The convolutional neural network constructed is as follows Figure 4 As shown, it includes 6 layers, and the specific structure is as follows: (1) The first layer is the input layer, which inputs six state variables {Ior(k), Ioi(k), Uor(k), Uoi(k), Isr(k), Isi(k)} through 6 neurons; (2) The second layer is convolution layer 1, which includes convolution, activation and pooling operations. The convolution kernel size of this layer is set to 3×1, the number is 16, and the pooling layer adopts maximum pooling with a kernel size of 2×1 and a stride of 2. (3) The third layer is convolution layer 2, which includes convolution and activation operations. The convolution kernel size is 3×1 and the number is 32. The activation functions in both convolution layers are ReLU activation functions. In addition, normalization is performed between the convolution and activation functions to avoid overfitting. (4) The fourth layer is a fully connected layer that connects neurons in all previous layers; (5) The fifth layer is the Softmax layer, which converts the output of the fully connected layer into a probability distribution so that the sum of the probabilities of each category is 1; (6) The sixth layer is the classification output layer, which takes the class with the highest statistical probability as the output and contains a total of 4 neurons.

[0024] The converter load type includes four load types, namely: no load, resistive-inductive load light load, resistive-inductive load heavy load, and nonlinear load, which are numbered 1, 2, 3, and 4 in the classification output of the convolutional neural network. Among the four load types, no load means that the output end of the three-phase inverter is not connected to a load. Resistive-inductive load light load means that the output end of the three-phase inverter is connected to a resistive-inductive load, and the power factor of the load is higher than a certain value. In this embodiment, the value is set to 0.85. Resistive-inductive load heavy load means that the output end of the three-phase inverter is connected to a resistive-inductive load, and the power factor of the load is lower than a certain value. In this embodiment, the value is set to 0.6. Nonlinear load means that the output end of the three-phase inverter is connected to an uncontrolled rectifier and then connected to a resistive-capacitive load.

[0025] The process of using a convolutional neural network to analyze the load type of a converter is as follows: first, converter operation status variable data including four load types are obtained to form a sample data set, and the constructed convolutional neural network is trained; in the actual operation process of the converter, the converter operation status data at time k is obtained by sampling, and the load type at that time is calculated by the trained convolutional neural network; the converter operation status data is sampled multiple times at intervals of a certain time length n, and the corresponding load type is calculated by the convolutional neural network; when the load types calculated by the convolutional neural network multiple times are consistent, the load type is determined to be the converter load type of the time period.

[0026] In this embodiment, the constructed convolutional neural network is used to identify the load type of the three-phase inverter. Figure 5 The figure shows the confusion matrix diagram of the convolutional neural network classification. As can be seen from the figure, the convolutional neural network has the best recognition accuracy for both light and heavy loads of resistive and inductive loads, with a recognition accuracy of 100%; the recognition accuracy for no-load is 99%, and the recognition accuracy for nonlinear loads is 97%. Figure 6 The figure shows the corresponding convolutional neural network load type recognition result diagram. The number of test samples for each type of load is 100. In the recognition of the first type of load, there was one error. In the recognition of the fourth type of load, there were three errors. In the recognition of the second and third types of loads, there were no errors.

[0027] The use of a convolutional neural network to identify the load type of a converter also includes offline training and online calculation processes of the convolutional neural network. Offline training refers to a convolutional neural network constructed by training a sample data set in a host computer. Online calculation refers to deploying the trained convolutional neural network on an embedded platform to calculate the sampled converter operating status data to obtain the load type of the converter.

Claims

1. A method for identifying converter load type based on convolutional neural network, characterized in that: The following steps are involved: S1. Determine the state variables involved in identifying the converter load type; The inverter load type identification refers to the load type of the inverter in the neural network-based inverter model prediction control system. The load type of the inverter determines the neural network control model used. Timely and accurate identification of the load type can improve the adaptability and reliability of the control system; the state variable refers to the state quantity of the inverter control system. For a three-phase voltage inverter with an LC filter, the state variables include load current Io, output voltage Uo, inductor current Is, output voltage control error dUo, etc.; the state variable involved in the inverter load type identification refers to the system state variable that changes significantly when the inverter load type changes; the state variable determined to participate in the inverter load type identification refers to the analysis of the correlation between the system state variable and the load type by the principal component analysis method to obtain the state variable that is strongly correlated with the load type; for a three-phase voltage inverter with an LC filter, the state variables involved in the load type identification obtained by principal component analysis are: load current Io, output voltage Uo, inductor current Is; S2. Obtaining state variable data of the converter operation; The state variable data of the converter operation is to obtain the state variable data involved in the identification of the converter load type. For a three-phase voltage inverter with an LC filter, the state variable data obtained include: three-phase load current, three-phase output voltage, and three-phase inductor current. The state variable data is to obtain the three-phase load current, output voltage, and inductor current at a certain time k by sampling, and coordinate transformation of the three-phase signal to obtain α - β The specific transformation process of the vector in the coordinate system is: x = x r + jx i =2 / 3[1, a , a 2 ]·[ x A , x B , x C ] T (1) In the formula, x r , x i is a state variable x After the coordinate transformation, the real and imaginary parts are x A , x B , x C is a state variable x The ABC three-phase components, operator a =exp( j 2π / 3); For a three-phase voltage source inverter with an LC filter, the system state variables at time k obtained through sampling and coordinate transformation are: the real part of the load current Ior(k), the imaginary part of the load current Ioi(k), the real part of the output voltage Uor(k), the imaginary part of the output voltage Uoi(k), the real part of the inductor current Isr(k), and the imaginary part of the inductor current Isi(k); S3. Analysis of converter load type using convolutional neural network.

2. The method for identifying the load type of a converter based on a convolutional neural network according to claim 1, characterized in that: In step S3, the structure of the constructed convolutional neural network has 6 layers in total, and the specific structure is as follows: (1) The first layer is the input layer, which inputs six state variables {Ior(k), Ioi(k), Uor(k), Uoi(k), Isr(k), Isi(k)} through 6 neurons; (2) The second layer is convolution layer 1, which includes convolution, activation and pooling operations. The convolution kernel size of this layer is set to 3×1, the number is 16, and the pooling layer adopts maximum pooling with a kernel size of 2×1 and a stride of 2. (3) The third layer is convolution layer 2, which includes convolution and activation operations. The convolution kernel size is 3×1 and the number is 32. The activation functions in both convolution layers are ReLU activation functions. In addition, normalization is performed between the convolution and activation functions to avoid overfitting. (4) The fourth layer is a fully connected layer that connects neurons in all previous layers; (5) The fifth layer is the Softmax layer, which converts the output of the fully connected layer into a probability distribution so that the sum of the probabilities of each category is 1; (6) The sixth layer is the classification output layer, which takes the class with the highest statistical probability as the output and contains a total of 4 neurons; The converter load type includes four load types, namely: no load, resistive inductive load light load, resistive inductive load heavy load, and nonlinear load, which are numbered 1, 2, 3, and 4 in the convolutional neural network classification output respectively; The process of using the convolutional neural network to analyze the load type of the converter is as follows: first, the converter operation state variable data including four load types are obtained to form a sample data set, and the constructed convolutional neural network is trained; in the actual operation process of the converter, the converter operation state data at time k is obtained by sampling, and the load type at that time is calculated by the trained convolutional neural network; with a certain time length n as an interval, the converter operation state data is sampled multiple times continuously, and the corresponding load type is calculated by the convolutional neural network; when the load types calculated by the convolutional neural network multiple times continuously are consistent, the load type is determined to be the converter load type of the time period; The use of a convolutional neural network to identify the load type of a converter also includes offline training and online calculation processes of the convolutional neural network. Offline training refers to a convolutional neural network constructed by training a sample data set in a host computer. Online calculation refers to deploying the trained convolutional neural network on an embedded platform to calculate the sampled converter operating status data to obtain the load type of the converter.

Images