MPPT (Maximum Power Point Tracking) control method based on dynamic optimization of lightweight convolutional neural network

Abstract

The invention relates to an MPPT (Maximum Power Point Tracking) control method based on dynamic optimization of a lightweight convolutional neural network, and belongs to the field of artificial intelligence control power tracking. The method comprises the following steps: S0, constructing a convolution-fuzzy neural network model which at least comprises an input layer, a convolution layer, a fuzzy control layer, a deep reinforcement learning layer and an output layer; s1, key parameters of photovoltaic equipment are collected in real time to serve as input data of an input layer; s2, performing convolution operation on the input data by using a convolution layer, extracting a feature vector, and transmitting the feature vector to a fuzzy control layer; s3, combining a fuzzy control layer with a deep reinforcement learning layer, and dynamically adjusting an optimization process of a fuzzy rule through a reward signal; and S4, the fuzzy control layer generates a control signal, and the control signal is transmitted to a DC / DC converter of the photovoltaic system through the output layer. The control method is suitable for a dynamic environment of a photovoltaic power generation system, can be expanded to a grid-connected system, and provides a new thought and technical support for efficient operation of photovoltaic power generation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence controlled power tracking, and in particular to an MPPT control method based on dynamic optimization of a lightweight convolutional neural network. Background Art

[0002] The Maximum Power Point Tracking (MPPT) system is an electrical system that adjusts the working state of the electrical module to enable the photovoltaic panel to output more electrical energy. It can effectively store the direct current generated by the solar panel in the battery, and can effectively solve the living and industrial electricity consumption in remote areas and tourist areas that are not covered by conventional power grids without causing environmental pollution.

[0003] In the early days of photovoltaic power generation system development, traditional maximum power point tracking (MPPT) methods had difficulty achieving maximum power output under changing environmental conditions due to problems such as high algorithm complexity and slow dynamic response speed. These methods usually rely on fixed rules and have difficulty adapting to rapid changes in external conditions such as light intensity and temperature, resulting in insufficient tracking accuracy, long response delays, and poor robustness under noise interference and non-ideal conditions. With the increase in the proportion of distributed photovoltaic power generation and the widespread application of microgrids, these deficiencies have become the main obstacles to improving the efficiency of photovoltaic power generation systems.

[0004] With the advancement of artificial intelligence and control technology, intelligent control methods have gradually become the key means to solve traditional MPPT problems. The MPPT method based on the combination of convolution-fuzzy neural network and deep reinforcement learning (DRL) has attracted much attention due to its efficient feature extraction ability, adaptive adjustment characteristics and optimization ability of reinforcement learning. Convolutional neural network (CNN) extracts features from system input parameters (such as light intensity, temperature, voltage and current, etc.), and uses fuzzy controllers to infer fuzzy rules on these features to predict the optimal operating point of the system. Fuzzy control is deeply integrated with CNN, and the convolution layer is used to optimize the membership function and improve the formulation of fuzzy rules, making it more robust and accurate in complex environments.

[0005] Furthermore, deep reinforcement learning optimizes fuzzy control rules by introducing a reward mechanism, and can actively explore and utilize the best control strategy when the environment changes dynamically. By adjusting the width of the membership function and the weight of the fuzzy rules, DRL improves the system's self-learning ability and adaptability to changing environments. Combining these three technologies, the system can more efficiently reduce tracking errors, while significantly improving the response speed and tracking accuracy of MPPT under rapidly changing lighting conditions, effectively reducing energy losses, and enhancing the overall performance and reliability of the photovoltaic power generation system.

[0006] Compared with the traditional MPPT method, this intelligent control method overcomes the sensitivity of fixed algorithms to noise and outliers, and shows higher robustness and reliability in complex environments. However, the existing MPPT method based on the combination of BP neural network and fuzzy control still has the following shortcomings: the BP neural network has a high computational complexity when processing high-dimensional features, which leads to longer training and inference time and affects the dynamic response performance; its feature extraction ability is limited, and it is difficult to capture the subtle dynamics of illumination changes, which reduces the tracking accuracy; at the same time, the BP neural network is easily disturbed by noise and outliers, and the tracking efficiency may decrease in unstable environments.

[0007] For example, in the patent document CN 105279558A "A multi-peak photovoltaic MPPT method based on BP neural network", a BP neural network is used to map the global maximum power point voltage of the photovoltaic system with the light intensity and ambient temperature of each component, eliminating the need for the search process of the traditional MPPT method and improving the response speed.

[0008] However, this method still has the following defects: the BP neural network has a high computational complexity when processing high-dimensional features, which may lead to prolonged training and inference time and insufficient response speed; the BP network has limited feature extraction capabilities for input data and has difficulty capturing subtle changes in illumination, thus affecting tracking accuracy; the BP neural network is easily affected by noise and outliers and has poor robustness, which may lead to decreased tracking efficiency in unstable environments.

[0009] In subsequent studies, many studies have also improved it. For example, patent document CN118350405A discloses a ship photovoltaic MPPT method based on an improved whale optimization algorithm to optimize the BP neural network. In this method, the improved whale optimization algorithm is used to improve the problems of long search time, slow convergence speed, easy to fall into local optimality, and poor generalization ability of the BP neural network.

[0010] However, this method is still relatively complicated, and it only considers the voltage error, the control parameters of the linear controller PI are lagging, and the system response speed is slow. In addition, the method does not disclose the specific PWM duty cycle adjustment rules. Summary of the invention

[0011] The present invention aims to overcome the shortcomings of the prior art and proposes an MPPT control method based on lightweight convolutional neural network dynamic optimization to meet the needs of photovoltaic power generation systems for rapid detection, selective protection and dynamic environmental response, while significantly improving the tracking accuracy and system efficiency. The solution to the technical problem of the present invention is as follows: A MPPT control method based on dynamic optimization of lightweight convolutional neural network includes the following steps: S0: constructing a convolution-fuzzy neural network model, wherein the convolution-fuzzy neural network model at least includes an input layer, a convolution layer, a fuzzy control layer, a deep reinforcement learning layer and an output layer; S1: collecting key parameters of photovoltaic equipment in real time as input data of the input layer, wherein the key parameters include light intensity, temperature, total radiation irradiance, direct irradiance and scattered irradiance; S2: using the convolution layer to perform convolution operation on the input data, extracting feature vectors, and transferring the feature vectors to the fuzzy control layer for real-time adjustment of the weights of fuzzy rules; S3: combining the fuzzy control layer with the deep reinforcement learning layer, and dynamically adjusting the optimization process of the fuzzy rules through reward signals; S4: the fuzzy control layer generates a control signal according to the current maximum power point voltage prediction value and the state of the system, and the control signal is transmitted to the DC / DC converter of the photovoltaic system through the output layer.

[0012] In one embodiment of the present invention, the convolution-fuzzy neural network model also includes a fully connected layer, which establishes a mapping relationship between the feature vector output by the convolution layer and the shape, width and center position of the membership function; the learning and updating of the membership function are optimized through the error back propagation of the convolution layer.

[0013] Furthermore, the output of the fuzzy control layer is obtained by weighted averaging, and the output of each rule is weighted by the membership function to obtain a final control signal, where the weighted average formula is: The output of the fuzzy control layer is obtained by weighted averaging, and the output of each rule is weighted by the membership function to obtain a final control signal, where the weighted average formula is: ,in, is the value of the membership function, n is the number of fuzzy rules, is the output of the rule, D is the duty cycle of the DC / DC converter.

[0014] Furthermore, the membership function is a Gaussian membership function: ,in, c is the center of the Gaussian function, σ is the standard deviation.

[0015] Furthermore, the loss function L is used to calculate the error between the target value and the network output. The formula is: ,in, is the network prediction output, is the target value; N is the total number of samples used to calculate the loss; According to the back propagation formula, calculate the parameters of the membership function , and the gradient of the fuzzy rule weight ; The calculation formula of its membership function parameters is: , , the gradient of its fuzzy rule weight The calculation formula is: ,in is the membership value corresponding to the i-th fuzzy rule, n is the number of fuzzy rules, and y is the final output value after defuzzification; the output layer defuzzification adopts weighted average, and its formula is: ,in and are the jth rule weight and fuzzy membership value respectively; n is the number of fuzzy rules; the gradient is calculated as follows: , and finally apply the gradient descent method or its variant to update the parameters, the formula is: ,in is the updated value of the model parameters at time t+1, is the current value of the model parameter at time t, is the learning rate.

[0016] In one embodiment of the present invention, the state space of the deep reinforcement learning layer consists of key parameters and errors of the photovoltaic system, and the action space consists of a duty cycle D, which is used to adjust the maximum power point of the photovoltaic system.

[0017] Furthermore, the reward signal R is defined as the performance of the system at each time step, measured by the proximity to the maximum power point: , where R t is the reward signal at time t, is the current power, is the maximum power point power; the deep reinforcement learning layer maximizes the cumulative reward , r is the discount factor.

[0018] In one embodiment of the present invention, the key parameters collected in S2 are normalized and preprocessed, and the preprocessed data is clustered and analyzed to set two types of clusters. The average value, standard deviation, maximum value and minimum value of each feature are calculated, and the cluster distance d(x, y) is calculated using the Euclidean metric. The distance calculation formula is as follows: , where: is the value of the mth feature of the day to be tested; is the same feature value of each possible similar day; T is the total number of features; the cluster with the highest similarity is selected as the similar day training data of the day to be tested.

[0019] In one embodiment of the present invention, the convolution layer is used to extract features from the input data. Multiple convolution kernels are used to perform convolution operations on the input data to extract spatial correlation features. Each input parameter will undergo a convolution operation to generate a different feature map. Assuming that the input data is X and the convolution kernel is W, the convolution operation is: ,in, represents the convolution operation, Y is the output feature map; The output of the convolutional layer is processed through a ReLU activation function.

[0020] The present invention also provides a grid-connected system, which uses the MPPT control method based on dynamic optimization of lightweight convolutional neural network as described above to perform grid connection and power scheduling. The grid-connected system uses the MPPT control method based on dynamic optimization of lightweight convolutional neural network as described in any one of claims 1-9 to perform grid connection and power scheduling.

[0021] Compared with the prior art, the present invention has the following beneficial effects: through the deep integration of convolutional neural network and fuzzy control, there is no need to introduce traditional disturbance voltage, which reduces system loss and improves stability; fuzzy control rules are combined with deep reinforcement learning to achieve dynamic optimization, enhance the self-learning and self-adaptation ability of the system in the change of external environment, and effectively solve the power oscillation and misjudgment problems near the maximum power point; in a dynamic environment, through the photovoltaic MPPT method combined with intelligent control, the output power of photovoltaic cells is significantly improved, while the system response speed and overall efficiency are improved. The present invention is suitable for the dynamic environment of photovoltaic power generation systems and can be extended to grid-connected systems, providing new ideas and technical support for the efficient operation of photovoltaic power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention is further described below in conjunction with the accompanying drawings.

[0023] Figure 1 Flowchart of the convolutional-fuzzy neural network algorithm.

[0024] Figure 2 The convolutional-fuzzy neural network structure diagram.

[0025] Figure 3 The data preprocessing flow chart.

[0026] Figure 4 Structural diagram of the intelligent control method. DETAILED DESCRIPTION

[0027] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0028] The present invention relates to artificial intelligence control technology, power tracking technology, data processing technology and microcomputer control technology, and is specifically an MPPT control method based on dynamic optimization of lightweight convolutional neural network. The method integrates convolutional neural network (CNN), fuzzy control and deep reinforcement learning (DRL) to realize the global maximum power point tracking (MPPT) of photovoltaic system in dynamic environment. The input parameters of photovoltaic system (such as light intensity, temperature, voltage and current) are extracted by CNN, and the adaptive adjustment characteristics of fuzzy control are used to predict the optimal working point; at the same time, the fuzzy control layer is combined with DRL to optimize the membership function and control rules through reward signals to achieve more efficient dynamic adjustment. The back propagation mechanism further optimizes the learning process of feature extraction and fuzzy rules, enhances the real-time learning ability and adaptive performance of the system, and the method can effectively solve the power oscillation and local optimal problems caused by the traditional perturbation observation method, realize efficient power dispatching and utilization, and provide technical support for the intelligent control of photovoltaic power generation.

[0029] See also Figure 1 The present invention discloses an MPPT control method based on dynamic optimization of convolutional neural network, comprising the following steps: S0: constructing a convolutional-fuzzy neural network model, wherein the convolutional-fuzzy neural network model at least comprises an input layer, a convolutional layer, a fuzzy control layer, a deep reinforcement learning layer and an output layer. In one embodiment of the present invention, the convolutional-fuzzy neural network structure diagram is as follows: Figure 2 shown.

[0030] S1: Real-time collection of key parameters of photovoltaic equipment as input data of the input layer, the key parameters include light intensity, temperature, total radiation irradiance, direct irradiance, and diffuse irradiance; S2: Use the convolution layer to perform convolution operation on the input data, extract the feature vector, and pass the feature vector to the fuzzy control layer for real-time adjustment of the weight of the fuzzy rule; S3: Combine the fuzzy control layer with the deep reinforcement learning layer to dynamically adjust the optimization process of the fuzzy rules through reward signals; S4: The fuzzy control layer generates a control signal based on the current maximum power point voltage prediction value and the system status, and the control signal is transmitted to the DC / DC converter of the photovoltaic system through the output layer.

[0031] The convolution-fuzzy neural network module is used to predict the maximum power point voltage and power of the photovoltaic system, and output the MPPT control signal to dynamically adjust the duty cycle of the DC-DC converter. When the system detects a drastic change in the external environment (such as light intensity or temperature), the convolution-fuzzy neural network will adjust the maximum power point prediction model through real-time learning capabilities, and make rapid adjustments based on new inputs to ensure that the system can respond quickly to changes and continue to track the maximum power point to achieve maximum power tracking.

[0032] The neural network is a convolution-fuzzy neural network, and fuzzy control is deeply integrated with the convolutional neural network, and the membership function learning of fuzzy control is carried out together through the convolutional layer of the convolutional neural network. While extracting the characteristics of the input parameters of the photovoltaic system (such as light intensity, temperature, voltage and current, etc.), the convolutional neural network optimizes the rules and feature extraction process of fuzzy control through back propagation, thereby achieving accurate prediction of the system state. On this basis, the fuzzy control layer is combined with deep reinforcement learning (DRL) to dynamically adjust the optimization process of fuzzy rules through reward signals (such as power tracking error). The exploration and utilization mechanism of reinforcement learning continuously optimizes the parameters of fuzzy rules (such as the width of the membership function, the weight of the fuzzy rules, etc.), achieving more efficient and accurate maximum power point tracking.

[0033] The parameters of the membership function of the present invention (such as shape, width and center position) are jointly optimized by the convolution layer and the fully connected layer of the convolutional neural network, and adjusted by the back propagation algorithm, so that the membership function can better adapt to the characteristic changes of the input parameters. The membership function is dynamically updated during the neural network training process and is consistent with the system input data, thereby improving the accuracy of the fuzzy rule in the maximum power point prediction process.

[0034] In the process of back-propagation optimization of fuzzy control rules and feature extraction, the error of convolutional neural network is transformed into The difference signal is back-propagated to the fuzzy control layer, so that the weight of the fuzzy control rule is gradually optimized according to the error reduction during the training process, ensuring the applicability of the fuzzy control rule to different input parameters. The feature extraction process extracts multi-scale environmental features (such as light intensity and temperature) through the convolution layer, and works in conjunction with the fuzzy membership function to achieve dynamic prediction and adjustment of the system state.

[0035] The deep reinforcement learning of the present invention dynamically adjusts fuzzy rules and neural network weights by constructing an optimization mechanism with reward signals as the core. Specifically, the reward signal guides the neural network to explore and utilize the optimal strategy based on power tracking error, system efficiency and response speed. During the training process, the reinforcement learning algorithm updates the parameters of the fuzzy rules (such as the shape of the membership function, the weight of the fuzzy rules, etc.) in real time according to changes in the system environment, and further improves the adaptability and tracking performance of the system in a dynamic environment through Q learning or policy gradient method.

[0036] The present invention improves the response speed and accuracy of MPPT tracking by reducing the power oscillation of the traditional disturbance observation method, and is suitable for photovoltaic power generation systems in dynamic environments.

[0037] In a specific embodiment of the present invention, features such as light intensity, temperature, total radiation irradiance, direct irradiance and scattered irradiance are collected, and the input data is convolved using a convolution kernel to extract the features of the input data. The time, light intensity and temperature in the photovoltaic device are extracted to form a two-dimensional matrix vector, and the maximum power point voltage of the photovoltaic array is used as the output of the convolution layer, and the fuzzy rules and defuzzification decisions are adjusted in real time.

[0038] In the specific embodiment, the collected data needs to be preprocessed. The main process of preprocessing is shown in Figure 3 Specifically, the collected data is normalized, and the normalization is shown in the following formula: , where is the normalized standard data; is the original data before normalization; , are the maximum and minimum values ​​in the original data.

[0039] The preprocessed data is clustered and analyzed to set two types of clusters. The mean, standard deviation, maximum and minimum values ​​of each feature are calculated, and the distance d(x, y) is calculated using the Euclidean metric. The distance calculation formula is as follows: , where: is the value of the mth feature of the day to be tested; is the same feature value of each possible similar day; T is the total number of features.

[0040] In this embodiment, the daily data is clustered into two types: sunny days (irradiation) and cloudy days (low irradiation). The data of the last 3 days is set as the test set, and the data belonging to the same cluster as the day to be tested is selected from the past 29 days. At the same time, the three key features that directly affect the active power output of photovoltaics—total radiation irradiance, direct irradiance, and scattered irradiance—are selected as input features for calculating the Euclidean distance. By comparing the Euclidean distance, the cluster with the highest similarity is selected as the similar day training data of the day to be tested, and 3 groups of similar days are selected for each test set (one day to be tested). The similar day data is used as the designed CNN training data, and is input into the model in 3 batches to complete the training.

[0041] The convolution layer is used to extract features from the input data. It usually uses multiple convolution kernels (filters) to perform convolution operations on the input data to extract spatial correlation features. Each input parameter (such as light intensity, voltage, etc.) will undergo a convolution operation to generate a different feature map. Assuming the input data is X and the convolution kernel is W, the convolution operation is: ,in, represents the convolution operation, Y is the output feature map.

[0042] In order to increase the nonlinearity of the model, the output of the convolutional layer is processed by an activation function. Common activation functions include ReLU, Sigmoid or Tanh. In a specific embodiment of the present invention, the ReLU activation function is used: In the fuzzy control layer, the control output is mainly derived through fuzzy rules, and the control signal is the final duty cycle (D). The input of the fuzzy control layer is the error E and the error change rate ΔE. Based on these inputs, the controller generates outputs through a set of predefined fuzzy rules: if E is positive and ΔE is positive, increase the duty cycle; if E is negative and ΔE is negative, reduce the duty cycle, etc. The output of each rule is a weighted result based on the membership function.

[0043] In order to quantify fuzzy variables (such as error E and error change rate ΔE), it is necessary to define a membership function. For each input variable, the membership function is usually a triangular or Gaussian function: Among them, the triangle membership function is: ,in, a,b,c are the three nodes of the triangle.

[0044] Gaussian membership function: ,in, c is the center of the Gaussian function, σ is the standard deviation.

[0045] The fuzzy control output is obtained by weighted average. The output of each rule is weighted by the membership function to obtain a final control signal. The weighted average formula is: ,in, is the value of the membership function, is the output of the rule, n is the number of fuzzy rules, D is the final control signal (duty cycle). The defuzzification layer generates the final duty cycle control signal based on the weighted average of the fuzzy rule outputs, which is directly used to control the maximum power point tracking (MPPT) of the photovoltaic system.

[0046] The back propagation mechanism of the fuzzy control layer in the convolutional-fuzzy neural network is to optimize the membership function and fuzzy rules. The core of back propagation is to optimize the fuzzy control parameters by calculating the error gradient, so that the network's adaptability to the dynamic environment is enhanced and more efficient global optimization is achieved.

[0047] First, forward propagation is performed to calculate the membership value ,Then fuzzy rules are applied to generate inference output, and finally defuzzification is performed to generate the final output.

[0048] The membership function is represented by a Gaussian function, and the loss function L is used to calculate the error between the target value and the network output. The formula is: ,in, is the network prediction output, is the target value; N is the total number of samples used to calculate the loss; According to the back propagation formula, calculate the parameters of the membership function , and the gradient of the fuzzy rule weight ; The calculation formula of its membership function parameters is: , , the gradient of its fuzzy rule weight The calculation formula is: ,in is the membership value corresponding to the i-th fuzzy rule, n is the number of fuzzy rules, and y is the final output value after defuzzification; the output layer defuzzification adopts weighted average, and its formula is: ,in and are the jth rule weight and fuzzy membership value respectively; n is the number of fuzzy rules; the gradient is calculated as follows: , and finally apply the gradient descent method or its variant to update the parameters, the formula is: ,in is the updated value of the model parameters at time t+1, is the current value of the model parameter at time t, is the learning rate.

[0049] Deep reinforcement learning (DRL) is an adaptive optimization process used to adjust the parameters of CNN and fuzzy controllers based on the system's behavioral feedback. The core of DRL is to feedback and optimize the strategy through reward signals, ultimately allowing the system to find the optimal control strategy in a dynamic environment. The state space consists of the key parameters and errors of the photovoltaic system, and the action space consists of the duty cycle D, which is the output variable of the network, which is used to adjust the maximum power point of the photovoltaic system. The reward function is defined as the performance of the system at each time step, usually measured by the proximity to the maximum power point: , where R t is the reward signal at time t, is the current power, is the maximum power point power.

[0050] Use Q-learning or policy gradient algorithms to optimize the control strategy. Deep reinforcement learning adjusts the fuzzy rules and CNN weights based on the reward signal so that the control system can achieve optimal power tracking under different environmental conditions.

[0051] The present invention combines deep reinforcement learning with fuzzy back propagation. In the deep reinforcement learning framework, the reward signal can be used to dynamically adjust the fuzzy rules and membership function parameters. Combined with the power output stability and dynamic response speed, the reward signal R is designed to maximize the cumulative reward. , r is the discount factor. At the same time, the fuzzy control parameters are optimized through back propagation. Through this feedback mechanism, the system gradually learns how to adjust the output according to the current state to optimize the maximum power point tracking.

[0052] After each training cycle, based on the difference between actual power and maximum power point, convolutional-fuzzy neural network and deep reinforcement learning will adjust the weights of feature extraction network and the parameters of fuzzy control rules. As training progresses, the system can more efficiently cope with environmental changes such as light and temperature, and improve MPPT efficiency.

[0053] In summary, the MPPT control method based on lightweight convolution-fuzzy neural network dynamic optimization realizes the accurate tracking of the maximum power point of the photovoltaic array and forms a complete solution suitable for improving the efficiency and stability of the photovoltaic system.

[0054] The structural schematic diagram of the intelligent control method of the present invention is as follows Figure 4The system of the present invention adjusts the logic in real time: the convolution-fuzzy neural network retrains and updates the maximum power point voltage every hour, the fuzzy control layer dynamically adjusts the duty cycle according to the latest maximum power point voltage and error to ensure that the system runs at the optimal power point, and the deep reinforcement learning layer dynamically adjusts the weight and membership function of the fuzzy control rules.

[0055] In other specific embodiments of the present invention, the Simulink simulation platform can be used to construct simulation models of a neural network module, a fuzzy controller module, and a DC-DC converter module to achieve maximum power tracking of a photovoltaic system. The neural network module can be trained and weight parameters can be loaded using MATLAB, and the fuzzy control module can be implemented using the fuzzy logic controller tool in Simulink.

[0056] The present invention improves the response speed and accuracy of MPPT tracking by reducing the power oscillation of the traditional disturbance observation method, and is suitable for photovoltaic power generation systems in dynamic environments.

[0057] The control method of the present invention can be applied to a DC-DC converter with a BOOST topology to achieve a boost function and stably transfer the maximum power of the photovoltaic cell to the load. The method can also be extended to a grid-connected system to achieve stable grid connection with the power grid and optimized power dispatching.

[0058] The preferred embodiments listed above further illustrate the objectives, technical solutions and advantages of the present invention in detail. It should be understood that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An MPPT control method based on lightweight convolutional neural network dynamic optimization, characterized in that: The following steps are involved: S0: constructing a convolutional-fuzzy neural network model, wherein the convolutional-fuzzy neural network model at least includes an input layer, a convolution layer, a fuzzy control layer, a deep reinforcement learning layer, and an output layer; S1: Real-time collection of key parameters of photovoltaic equipment as input data of the input layer, the key parameters include light intensity, temperature, total radiation irradiance, direct irradiance, and diffuse irradiance; S2: Use the convolution layer to perform convolution operation on the input data, extract the feature vector, and pass the feature vector to the fuzzy control layer for real-time adjustment of the weight of the fuzzy rule; S3: Combine the fuzzy control layer with the deep reinforcement learning layer to dynamically adjust the optimization process of the fuzzy rules through reward signals; S4: The fuzzy control layer generates a control signal based on the current maximum power point voltage prediction value and the system status, and the control signal is transmitted to the DC / DC converter of the photovoltaic system through the output layer.

2. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 1, characterized in that: The convolution-fuzzy neural network model also includes a fully connected layer, which establishes a mapping relationship between the feature vector output by the convolution layer and the shape, width and center position of the membership function; the learning and updating of the membership function are optimized through the error back propagation of the convolution layer.

3. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 2 is characterized in that: The output of the fuzzy control layer is obtained through weighted averaging. The output of each rule is weighted by the membership function to obtain a final control signal. The weighted average formula is: ,in, is the value of the membership function, n is the number of fuzzy rules, is the output of the rule, D is the duty cycle of the DC / DC converter.

4. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 2, characterized in that: The membership function is a Gaussian membership function: ,in, c is the center of the Gaussian function, σ is the standard deviation.

5. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 4 is characterized in that: Use the loss function L to calculate the error between the target value and the network output. The formula is: ,in, is the network prediction output, is the target value; N is the total number of samples used to calculate the loss; According to the back propagation formula, calculate the parameters of the membership function , and the gradient of the fuzzy rule weight ; The calculation formula of its membership function parameters is: , , the gradient of its fuzzy rule weight The calculation formula is: ,in is the membership value corresponding to the i-th fuzzy rule, n is the number of fuzzy rules, and y is the final output value after defuzzification; the output layer defuzzification adopts weighted average, and its formula is: ,in and are the jth rule weight and fuzzy membership value respectively; n is the number of fuzzy rules; the gradient is calculated as follows: , and finally apply the gradient descent method or its variant to update the parameters, the formula is: ,in is the updated value of the model parameters at time t+1, is the current value of the model parameter at time t, is the learning rate.

6. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 1, characterized in that: The state space of the deep reinforcement learning layer consists of the key parameters and errors of the photovoltaic system, and the action space consists of the duty cycle D, which is used to adjust the maximum power point of the photovoltaic system.

7. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 6, characterized in that: The reward signal R is defined as how well the system performs at each time step, measured by how close it is to the maximum power point: , where R t is the reward signal at time t, is the current power, is the maximum power point power; the deep reinforcement learning layer maximizes the cumulative reward , r is the discount factor.

8. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 1, characterized in that: The key parameters collected in S2 are normalized and preprocessed, and the preprocessed data is clustered and analyzed to set two types of clusters. The mean, standard deviation, maximum and minimum values ​​of each feature are calculated, and the cluster distance d(x,y) is calculated using the Euclidean metric. The distance calculation formula is as follows: , where: is the value of the mth feature of the day to be tested; is the same feature value of each possible similar day; T is the total number of features; the cluster with the highest similarity is selected as the similar day training data of the day to be tested.

9. The MPPT control method based on lightweight convolutional neural network dynamic optimization according to claim 1, characterized in that: The convolution layer is used to extract features from the input data. It uses multiple convolution kernels to perform convolution operations on the input data to extract spatial correlation features. Each input parameter will undergo a convolution operation to generate a different feature map. Assuming that the input data is X and the convolution kernel is W, the convolution operation is: in, represents the convolution operation, Y is the output feature map; The output of the convolutional layer is processed through a ReLU activation function.

10. A grid-connected system, characterized in that: The grid-connected system applies the MPPT control method based on dynamic optimization of a lightweight convolutional neural network as described in any one of claims 1 to 9 to perform grid connection and power scheduling.

Images