A microgrid cloud-edge collaborative control method based on deep learning
Through deep learning technology, cloud-edge coordinated control is realized in the microgrid, and the linkage calculations between the edge and the cloud are solved, and the problems of complexity and low efficiency in microgrid regulation are improved, and the accuracy of computing efficiency and strategies are improved.
Patent Information
- Application Number
- CN202111036998.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-06
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-09-06
AI Technical Summary
The existing microgrid control methods have problems such as redundant and complex solution process, disaster in dimensionality, low optimization calculation efficiency, and easy to fall into local optimal solutions in the complex regulation process, which is difficult to meet the increasingly complex real-time requirements of microgrid operation. .
Adopting a microgrid cloud-edge collaborative control method based on deep learning, the edge-end controller collects operating status data to generate characteristic images, and links with the cloud server to build a neural network model to identify and regulate strategies, so as to achieve simple calculation of the microgrid regulation strategy and precision.
The calculation efficiency of the microgrid regulation strategy is improved, the local optimal solution is avoided, and the real-time requirements of complex microgrid operation are met.
Smart Images

Figure CN113872322B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of microgrid collaborative control, and in particular to a microgrid cloud-edge collaborative control method based on deep learning. Background Art
[0002] With the continuous improvement of the penetration rate of renewable energy, the microgrid system has a large-scale and wide-area trend in terms of construction time and geographical space. In terms of microgrid control issues, in terms of control system construction, existing microgrid engineering projects usually adopt a two-layer control structure of local unit layer and monitoring layer. The monitoring layer is the decision-making layer for monitoring and controlling the microgrid system and safe and economic operation. It makes decisions based on the operation information uploaded by each unit in the local layer and sends them to the corresponding physical components in the local layer for execution, so as to realize the dispatching and control of each power generation, energy storage and load unit in the microgrid; in terms of microgrid control strategy formulation, existing research is mostly based on physical mechanism analysis, through the establishment of optimization control model, and the use of optimization algorithm to formulate control strategy. The above traditional microgrid control and regulation methods have problems such as redundant and complex solution process, dimensionality disaster, low optimization calculation efficiency, and easy to fall into local optimal solution under the current development trend of microgrids with multiple points and wide range and increasingly complex control methods. It is difficult to meet the real-time requirements of the increasingly complex microgrid operation. Summary of the invention
[0003] In order to solve the above problems, the present invention provides a microgrid cloud-edge collaborative control method based on deep learning, in which the edge controller collects operating status data to generate a characteristic image and uploads it to a cloud server, and directly calls or fine-tunes the source-load-storage control strategy issued by the cloud server according to the parameter identification of the cloud server as the microgrid control strategy. The cloud server generates a source-load-storage control strategy based on the operating status data uploaded by the edge controller and sends it to the edge controller, and constructs a neural network model to obtain the model parameters and the confirmation judgment threshold of each label category according to the characteristic image uploaded by the edge controller and sends it to the edge controller. Through the joint calculation of the edge controller and the cloud server, the calculation of the microgrid control strategy is simplified and precise, the optimization calculation efficiency is improved, and it is prevented from falling into the local optimal solution, so as to meet the current increasingly complex real-time operation requirements of microgrids.
[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows:
[0005] A microgrid cloud-edge collaborative control method based on deep learning includes the following steps of an edge controller:
[0006] S1: Collect operation status data;
[0007] S2: Generate a characteristic image based on the operating status data;
[0008] S3: receiving the model parameters and label category confirmation threshold from the cloud server, and using the model to identify the label category and its probability value to which the characteristic image belongs;
[0009] S4: Compare the highest probability value among the probability values with the confirmation decision threshold of the corresponding label category;
[0010] If the highest probability value is greater than the confirmation judgment threshold, the source-load-storage control strategy of the corresponding tag category issued by the cloud server is called, and the characteristic image and operation status data are uploaded to the cloud server;
[0011] If the highest probability value is less than the confirmation judgment threshold, the source-load-storage control strategy of the corresponding tag category is fine-tuned to form a new source-load-storage control strategy, and the characteristic image and operation status data are uploaded to the cloud server;
[0012] S5: Decompose the source-load-storage control strategy and send it to the corresponding distributed sensing unit for execution.
[0013] Furthermore, the operating status data includes source-load-storage operating status data of the microgrid wind power, photovoltaic, energy storage and load units.
[0014] Furthermore, the characteristic image is a comprehensive characteristic image of the source-load-storage operation status.
[0015] Furthermore, the model adopts a CNN convolutional neural network model.
[0016] A microgrid cloud-edge collaborative control method based on deep learning includes the following steps of a cloud server:
[0017] A1: Build an optimization model and input the operating status data uploaded by the edge controller, generate source-load-storage control strategies under each tag category and send them to the edge controller;
[0018] A2: Build a neural network model and input the characteristic image uploaded by the edge controller for training and testing, obtain the neural network model parameters and the confirmation judgment threshold of each label category and send them to the edge controller.
[0019] Further, the construction of the optimization model and inputting the operation status data uploaded by the edge controller, generating the source-load-storage regulation strategy under each tag category and issuing it to the edge controller includes the following steps:
[0020] A11: Receive the operation status data uploaded by the edge controller and set the corresponding label category;
[0021] A12: Generate corresponding typical operation data for the operation status data of each tag category;
[0022] A13: Build an optimization model and input typical operation data to obtain the corresponding source-load-storage regulation strategy;
[0023] A14: Send the source-load-storage regulation strategy and the corresponding label category to the edge controller.
[0024] Furthermore, the construction of the neural network model and inputting the characteristic image uploaded by the edge controller for training and testing, obtaining the neural network model parameters and the confirmation judgment threshold of each label category and sending them to the edge controller, includes the following steps:
[0025] A21: Receive the characteristic image uploaded by the edge controller and set the corresponding label category;
[0026] A22: The feature image data under each label category is divided into a training set and a test set, where the proportion of the training set is larger than that of the test set;
[0027] A23: Build a neural network model and train the neural network model using the training set;
[0028] A24: Use the test set to test the trained neural network model;
[0029] A25: Send the successfully tested neural network model parameters and the confirmation judgment thresholds of each label category to the edge controller.
[0030] Furthermore, the training of the neural network model using the training set includes two parts: a forward propagation process and a back propagation process. The forward propagation process includes the following steps:
[0031] A231: Initialize the neural network model;
[0032] A232: Input the training set feature images into the neural network model;
[0033] A233: The neural network model calculates the label category and its probability value;
[0034] A234: Calculate the deviation between the target value and the probability value, and determine whether it is within the allowable range; if it is within the allowable range, fix the neural network model weight parameters and label category confirmation judgment threshold; if it is not within the allowable range, enter the back propagation process to optimize the weight parameters until the deviation is within the allowable range;
[0035] The back-propagation process consists of the following steps:
[0036] A235: End training and save the training model;
[0037] A236: Calculate the correction amount of the connection weights between the input layer and the hidden layer and the correction amount of the connection weights between the hidden layer and the output layer;
[0038] A237: Update the connection weights of the input layer, hidden layer and output layer respectively, and execute step A233.
[0039] Furthermore, the use of the test set to test the trained neural network model includes the following steps:
[0040] A241: Input the test set feature images into the neural network model one by one to calculate the probability value of each label category;
[0041] A242: Determine whether the label category corresponding to the highest probability value is the label category to which the characteristic image belongs. If not, return to the training process to retrain the neural network model. If yes, use the highest probability value as the probability value of the label category to which the characteristic image belongs.
[0042] A243: After obtaining the probability values of the label categories of all images in the test set, determine whether all probability values are greater than a preset threshold; if so, fix the neural network parameters and send them to the edge controller; if not, return to the training process and retrain the neural network model.
[0043] Furthermore, the operating status data includes source-load-storage operating status data of microgrid wind power, photovoltaic, energy storage and load units; the characteristic image is a comprehensive characteristic image of source-load-storage operating status; and the neural network model adopts a CNN convolutional neural network model.
[0044] The present invention provides a microgrid cloud-edge collaborative control method based on deep learning, wherein the edge controller collects operating status data to generate a characteristic image, identifies the label category to which the characteristic image belongs after receiving the model parameters and confirmation judgment threshold of the cloud server, directly calls or fine-tunes the source-load-storage control strategy issued by the cloud server as the microgrid control strategy according to the label category, and uploads the operating status data and the characteristic image to the cloud server. The cloud server generates source-load-storage control strategies of each label category according to the operating status data uploaded by the edge controller and sends it to the edge controller, constructs a neural network model for training and testing according to the characteristic image uploaded by the edge controller, obtains the model parameters and the confirmation judgment threshold of each label category and sends it to the edge controller. Through the linkage calculation of the edge controller and the cloud server, the microgrid control strategy calculation is simplified and precise, the optimization calculation efficiency is improved, and it is prevented from falling into the local optimal solution, so as to meet the current increasingly complex real-time operation requirements of the microgrid. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1This is a flow chart of the edge controller of a microgrid cloud-edge collaborative control method based on deep learning;
[0046] Figure 2 This is a cloud server flow chart of a microgrid cloud-edge collaborative control method based on deep learning;
[0047] Figure 3 This is a flowchart for training and testing CNN convolutional neural network models;
[0048] Figure 4 This is a schematic diagram of the CNN convolutional neural network model structure;
[0049] Figure 5 Schematic diagram of neuron structure. DETAILED DESCRIPTION
[0050] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.
[0051] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure 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 disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.
[0052] Embodiment 1
[0053] like Figure 1 The figure shows a schematic diagram of an embodiment of a microgrid cloud-edge collaborative control method based on deep learning. This embodiment is described in detail with the edge controller as the main body, including the following steps:
[0054] S1: Collect the operating status data of wind power, photovoltaic, energy storage and load units in the microgrid;
[0055] S2: Generate a comprehensive characteristic image of the source-load-storage operation state of the microgrid in a new round of regulation cycle according to the operation state data;
[0056] S3: Receive the CNN convolutional neural network model parameters and the confirmation threshold corresponding to each label category sent by the cloud server; use the CNN convolutional neural network model to identify the label category to which the comprehensive characteristic image of the microgrid source-load-storage operation status of the regulation cycle belongs, and calculate the probability value of belonging to each label category;
[0057] S4: Compare the highest probability value among the probability values with the confirmation threshold of the label category corresponding to the highest probability value;
[0058] If the highest probability value is greater than the judgment threshold, it means that the tag category is confirmed successfully, and the source-load-storage control strategy under the tag category corresponding to the highest probability value is called in the microgrid control strategy library as the microgrid control strategy for this control cycle. At the same time, the corresponding microgrid source-load-storage operation state comprehensive characteristic image is transmitted to the cloud server and stored in the image sub-library of the corresponding tag category, and the corresponding source-load-storage operation data is stored in the source-load-storage operation data sub-library of the corresponding tag category of the cloud server;
[0059] If the highest probability value is less than the judgment threshold, it means that the label category has not been confirmed successfully. Based on the source-load-storage control strategy under the label category corresponding to the highest probability value, fine-tuning is performed according to expert experience to form a new source-load-storage control strategy as the microgrid control strategy for this control cycle, and it is stored in the microgrid control strategy library. At the same time, unsuccessful classification and identification means that the microgrid operation state of this control cycle belongs to a new category of operation state. The cloud server is notified to add a new category label and its corresponding image sub-library and source-load-storage operation data sub-library, and the corresponding microgrid source-load-storage operation state comprehensive characteristic image is stored in the image sub-library corresponding to the new category label of the cloud server, and the corresponding source-load-storage operation state data is stored in the source-load-storage operation data sub-library corresponding to the new category label of the cloud server.
[0060] S5: Decompose the microgrid control strategy of the control cycle and send it to the corresponding distributed sensing unit for execution.
[0061] In the specific implementation, the above process is executed by the edge controller online cycle according to the preset control time scale to realize online decision-making and control of microgrid operation.
[0062] In a specific implementation, the microgrid control strategy library stores various tag categories and their corresponding source-load-storage control strategies issued by the cloud server.
[0063] Embodiment 2
[0064] like Figure 2 The figure shows another embodiment of a microgrid cloud-edge collaborative control method based on deep learning. This embodiment is described in detail with a cloud server as the main body, and includes the following steps:
[0065] A1: Construct a microgrid optimization model and input the source-load-storage operation data under each tag category of the microgrid, and use a genetic algorithm to solve and generate the source-load-storage control strategy under each tag category of the microgrid;
[0066] A2: Construct a CNN convolutional neural network and input the comprehensive characteristic images of the source-load-storage operation status under each label category of the microgrid for training and testing to obtain the CNN convolutional neural network model parameters and the confirmation judgment thresholds of each label category.
[0067] In specific implementation, Figure 2 As shown, the construction of the microgrid optimization model and input of the source-load-storage operation data under each tag category of the microgrid, and the use of a genetic algorithm to solve and generate the source-load-storage regulation strategy under each tag category of the microgrid include the following steps:
[0068] A11: Establish a source-load-storage operation data sub-library to store the source-load-storage operation data reflecting the microgrid operation status uploaded by the edge controller, and set the corresponding label category; each microgrid operation status has a corresponding source-load-storage operation data sub-library to store the corresponding source-load-storage operation data, and has a corresponding label category;
[0069] A12: For the source-load-storage operation data under each tag category, a weighted average method is used to generate the corresponding microgrid typical operation data under the tag category;
[0070] A13: Construct a mathematical model for microgrid operation optimization, input the typical operation data of the microgrid under each tag category into the mathematical model for microgrid operation optimization, and use classical optimization algorithms such as genetic algorithms to solve the mathematical model for microgrid operation optimization, and obtain the optimal source-load-storage control strategy for the microgrid operation status under the corresponding tag category;
[0071] A14: The optimal source-load-storage control strategy is stored in the microgrid strategy sub-library and set with a corresponding label category, and then sent to the edge controller.
[0072] In specific implementation, Figure 2 As shown, the construction of the CNN convolutional neural network and inputting the comprehensive characteristic images of the source-load-storage operation status under each label category of the microgrid for training and testing, and obtaining the CNN convolutional neural network model parameters and the confirmation judgment threshold of each label category include the following steps:
[0073] A21: Establish a source-load-storage operation state comprehensive characteristic image sub-library to store the source-load-storage operation state comprehensive characteristic image uploaded by the edge controller, and set the corresponding label category; each microgrid operation state has a corresponding source-load-storage operation state comprehensive characteristic image sub-library to store the corresponding source-load-storage operation state comprehensive characteristic image, and has a corresponding label category;
[0074] A22: randomly divide the image data in the source-load-storage operation status comprehensive characteristic image sub-library under each label category into a training set and a test set, where the proportion of the training set is larger than that of the test set;
[0075] A23: Build a CNN convolutional neural network model and use the training set to train the CNN convolutional neural network model;
[0076] A24: Use the test set to test the trained CNN convolutional neural network model;
[0077] A25: Send the successfully tested CNN convolutional neural network model parameters and the confirmation thresholds of each label category to the edge controller.
[0078] In specific implementation, Figure 3 As shown, the training of the CNN convolutional neural network model using the training set includes a forward propagation process and a back propagation process, and the forward propagation process includes the following steps:
[0079] A231: Initialize the CNN convolutional neural network model;
[0080] A232: Input the comprehensive characteristic image of the source-load-storage operation status of the training set into the CNN convolutional neural network model;
[0081] A233: CNN convolutional neural network model calculates the most similar label category and probability value;
[0082] A234: Calculate the target value and probability value to obtain the deviation e, and determine whether e is within the allowable range; if the deviation e is within the allowable range, the CNN convolutional neural network model meets the requirements, the training is terminated, the weight parameters are fixed, and the judgment threshold is confirmed; if the deviation e is not within the allowable range, the CNN convolutional neural network model does not meet the requirements, and enters the back propagation process to optimize the weight parameters until the actual output deviation e is within the allowable range;
[0083] The back-propagation process consists of the following steps:
[0084] A235: End training and save the training model;
[0085] A236: Use the BP (Back Propagation) algorithm to calculate the correction amount of the weights of the input layer and the hidden layer and the correction amount of the weights of the hidden layer and the output layer;
[0086] A237: Update the connection weights of the input layer, hidden layer and output layer respectively, and proceed to step A233.
[0087] In a specific implementation, the use of the BP (Back Propagation) algorithm in step A236 to calculate the correction amount of the weights of the input layer and the hidden layer and the correction amount of the weights of the hidden layer and the output layer includes the following steps:
[0088] Output A from the input layer to the hidden layer j for:
[0089]
[0090] Where W ih is the weight of the input layer and the hidden layer, W ih The value range is (-1,1), and the weights are assigned in the form of random numbers; b ih is the bias constant of the input layer and the hidden layer, A i is the input source-load-storage operation state comprehensive characteristic image training set, k is A i , f(x) is the activation function.
[0091] The output after the activation function is:
[0092] H i =f(A i )
[0093] The output of the hidden layer to the output layer H j for:
[0094]
[0095] Where W ho is the weight of the hidden layer and the output layer, b ho is the bias constant of the hidden layer and the output layer, W ho The value range is (-1,1).
[0096] The output Y after the activation function j for:
[0097] Y j =f(H j )
[0098] The error function E is:
[0099]
[0100] Where D j is the expected output.
[0101] The correction amount for each input weight is:
[0102]
[0103] Where l is the learning rate, ΔW ih is the correction amount of the weights of the input layer and the hidden layer, ΔW ho is the correction amount for the weights of the hidden layer and the output layer.
[0104] In a specific implementation, the formula for updating the connection weight in step A237 is as follows:
[0105]
[0106] In the formula, and They are the weights of the input layer and the hidden layer and the weights of the hidden layer and the output layer after the Nth learning; and They are the weights of the input layer and the hidden layer and the weights of the hidden layer and the output layer after the N+1th learning.
[0107] In specific implementation, Figure 4 The CNN convolutional neural network model shown in FIG. 1 includes an input layer, a hidden layer and an output layer, wherein the hidden layer is connected to the input layer and the output layer respectively, and the input layer, the hidden layer and the output layer are all composed of learnable weights and bias variables such as Figure 5 Several neuron structures are shown.
[0108] In a specific implementation, the neuron is described by a mathematical expression as follows:
[0109] t=f(w 1 ×a 1 +w 2 ×a 2 +...w n ×a n +b)
[0110] t is the output, where a 1 ~a n represents the input vector, b represents the bias constant, and w 1 ~w n represents the weight, f(x) is the activation function; when the input is a 1 ~a n When the vector arrives at the neuron, each input vector has a different weight w n , the neuron outputs the result t after the activation function considering the input of different weights and the bias constant.
[0111] In a specific implementation, the input layer is used to receive the comprehensive characteristic image of the source-load-storage operation status. Usually, computers divide the color depth into 256 color levels. In the computer, the essence of the image is actually an array composed of the color information of each pixel. Each pixel of the grayscale image only has light and dark information, while each pixel of the color image has three color channels of R / G / B. When the comprehensive characteristic image of the source-load-storage operation status is input, data of a certain area size is actually input into the computer, that is, the corresponding three-dimensional matrix. The length and width of the matrix represent the size of the image.
[0112] In the specific implementation, the hidden layer includes a convolution layer, a pooling layer, a fully connected layer and a Softmax regression layer. It mainly extracts the features of the comprehensive characteristic image of the source-load-storage operation status through multiple rounds of alternating "convolution-pooling" structures to obtain a feature matrix, and then expands the extracted feature matrix into a one-dimensional matrix through the fully connected layer. Finally, the Softmax regression layer outputs the corresponding most similar label category and probability value.
[0113] In a specific implementation, the output layer outputs the most similar label category probability value obtained by the Softmax regression layer.
[0114] In the specific implementation, the convolution layer is the most important layer in the convolutional neural network. After the pixel data of the source-load-storage operation state comprehensive characteristic curve image is converted into a matrix through the input layer, the convolution kernel in the convolution layer uses the "local connection, weight sharing" method to extract the features of the image data. The mathematical expression describing the convolution process is:
[0115]
[0116] Where G represents the comprehensive characteristic image matrix of the microgrid source-load-storage operation state after feature extraction, m, n are rows and columns, f represents the input microgrid source-load-storage operation state comprehensive characteristic image matrix; w represents the convolution kernel; s and t define the size of the convolution kernel.
[0117] After the convolution operation, the size of the comprehensive characteristic image of the source-load-storage operating status becomes smaller, but the original image features are still retained. In addition, the parameters involved in the convolution layer include the convolution step size and edge padding. The convolution step size refers to the step size of each movement of the convolution kernel when performing convolution on the image elements. The output feature map is different for different step sizes. The smaller the step size, the finer the extracted features and the richer the features. The larger the step size, the coarser the extracted features and the simpler the features. Edge padding is to add a certain number of rows and columns on each side of the input feature map, so that the output and input feature maps have the same size, which improves the utilization of edge information to a certain extent. The calculation formula for the convolution result is:
[0118]
[0119]
[0120] In the formula, H 1 , W 1 Indicates the length and width of the input, H 2 , W 2 Represents the length and width of the output feature map, f u 、f w Represents the length and width of the convolution kernel, s represents the step size of the sliding window, and p represents the boundary padding.
[0121] In the specific implementation, the dimensionality reduction and feature extraction of the comprehensive characteristic image of the source-load-storage operating status are completed through the convolution layer, but the dimension of the feature image is still very high at this time. Too high a dimension not only increases the amount of calculation, but also easily leads to overfitting, which reduces the accuracy of the actual prediction. Therefore, the features extracted by the convolution layer are divided into different small areas through the pooling layer, and then the maximum value or average value of the area is used to represent the convolution features after pooling and dimensionality reduction, which plays the role of secondary feature extraction. In this way, the difference in image position becomes smaller, which greatly and efficiently reduces the number of parameters and the amount of calculation, and controls the overfitting problem to a certain extent.
[0122] In a specific implementation, the pooling layer includes maximum pooling and mean pooling; the maximum pooling is to take the maximum value of the feature points in the region, which is better for texture extraction; the mean pooling is to find the average value of the feature points in the region, which is better for background retention; according to its own requirements, the present invention selects maximum pooling.
[0123] In the specific implementation, the fully connected layer is referred to as the FC layer. The neurons in the fully connected layer are connected to the neurons in the adjacent layers before and after the CNN convolutional neural network, and can extract and integrate useful information from the upper layer and pass it to the next connected layer. The main purpose of the fully connected layer is dimensionality transformation. The high-dimensional distributed feature representation extracted by the convolution layer is converted into a low-dimensional sample label through a convolution kernel of size 1×1, so as to calculate the classification evaluation value of the features extracted by the convolution and pooling layers. In order to improve the performance of the CNN convolutional neural network model, the activation function of each neuron in the fully connected layer adopts the ReLU (Rectified Linear Unit) function. The full name of ReLU is the linear rectification function, and its mathematical expression is:
[0124] f(x)=max(0,x)
[0125] The ReLU function can remove redundant data, retain and map data features, eliminate the influence of other complex activation functions, and simplify the calculation process.
[0126] In specific implementation, the Softmax regression layer is also called logistic regression analysis, and the Softmax regression algorithm can be used for multi-classification. The principle of the Softmax algorithm is: weighted average the pixels of the source-load-storage operation status comprehensive characteristic image training data containing the label category to obtain the weight C of each label category at each pixel point; if C>0, it belongs to the label category; if C<0, it does not belong to the label category. The Softmax regression layer outputs the previous FC layer and performs a linear regression. The classification evaluation value obtained by the FC layer is regularized by calculating the specific data to obtain the probability of each label category. If a source-load-storage operation status comprehensive characteristic image training set {(x (i) ,y(i) ; i∈1,...,N,y (i) ∈0,...,K-1}, where x (i) is the i-th input image patch; y (i) The calculation formula for its class label is:
[0127]
[0128] In the formula, represents the output probability that the i-th input belongs to the j-th class; represents the output probability that the i-th input belongs to the j-th class.
[0129] In a specific implementation, the confirmation threshold is the minimum value of the probability values obtained by testing a test set under a certain label category through a CNN convolutional neural network model, which is used as the determination threshold for online identification and confirmation of the label category.
[0130] In the specific implementation, it is necessary to train and test the CNN convolutional neural network model regularly every day so that as the volume of microgrid operation data increases and the microgrid control strategy library is updated, the corresponding microgrid operation status identification will be more accurate, the formulated control strategy will also be more accurate, and continuous deep learning will be achieved for more perfect control.
[0131] In a specific implementation, when the error function E is small enough or the number of learning times reaches the set upper limit N, the model stops training and saves parameters such as weights used for identification and classification.
[0132] In specific implementation, Figure 3 As shown, the trained CNN convolutional neural network model is tested using the test set as described in step A24, including the following steps:
[0133] A241: Input the images in the source-load-storage operation status comprehensive characteristic image test set one by one, and calculate the probability corresponding to each label category by calling the trained CNN convolutional neural network model;
[0134] A242: Determine whether the label category corresponding to the highest probability value is the label category to which it belongs. If it is different from the label category to which it belongs, it means that the CNN convolutional neural network model test has failed and it is necessary to return to the training process to retrain the CNN convolutional neural network model; if the label category corresponding to the highest probability value is the same as the label category to which it belongs, then output the probability value of the corresponding label category;
[0135] A243: After all images in the test set are tested and the corresponding label category probabilities are calculated through the above steps, if the label category probability values corresponding to all images in the test set are greater than the preset threshold, the model test is successful, indicating that the CNN convolutional neural network model is trained successfully. The CNN convolutional neural network parameters are fixed to form a model and sent to the edge controller; if the label category probability corresponding to any image in the test set is less than the preset threshold, it means that the CNN convolutional neural network model training has failed, and it is necessary to return to the training process to retrain the convolutional neural network model.
[0136] The present invention provides a microgrid cloud-edge collaborative control method based on deep learning, wherein the method collects operating status data at the edge controller, generates a characteristic image according to the operating status data, identifies the label category to which the characteristic image belongs after receiving the model parameters and label category confirmation judgment threshold of the cloud server, directly calls or fine-tunes the source-load-storage control strategy issued by the cloud server according to the label category as the microgrid control strategy, and uploads the operating status data and the characteristic image to the cloud server. The cloud server generates source-load-storage control strategies of each label category according to the operating status data uploaded by the edge controller and sends it to the edge controller, constructs a neural network model and trains and tests it according to the characteristic image uploaded by the edge controller, obtains the model parameters and the confirmation judgment threshold of each label category and sends it to the edge controller. By calculating and linking the edge controller and the cloud server respectively, the microgrid control strategy calculation is simplified and precise, the optimization calculation efficiency is improved, and it is prevented from falling into the local optimal solution, so as to meet the current increasingly complex real-time requirements of microgrid operation.
[0137] The above is only for explaining the implementation mode of the present invention and is not intended to limit the present invention. For those skilled in the art, any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention without creative work should be included in the protection scope of the present invention.
Claims
1. A microgrid cloud-edge collaborative control method based on deep learning, It is characterized in that The following steps are included in the edge controller: S1: Collect operation status data; S2: Generate a characteristic image based on the operating status data; S3: receiving the CNN convolutional neural network model parameters and the label category confirmation threshold from the cloud server, and using the CNN convolutional neural network model to identify the label category and its probability value to which the characteristic image belongs; S4: Compare the highest probability value among the probability values with the confirmation decision threshold of the corresponding label category; If the highest probability value is greater than the confirmation threshold, the source-load-storage control strategy of the corresponding tag category issued by the cloud server is called, and the characteristic image and operation status data are uploaded to the cloud server; If the highest probability value is less than the confirmation judgment threshold, the source-load-storage control strategy of the corresponding tag category is fine-tuned to form a new source-load-storage control strategy, and the characteristic image and operation status data are uploaded to the cloud server; S5: Decompose the source-load-storage control strategy and send it to the corresponding distributed sensing unit for execution.
2. According to the deep learning-based microgrid cloud-edge collaborative control method of claim 1, It is characterized in that The operating status data includes source-load-storage operating status data of the microgrid wind power, photovoltaic, energy storage and load units.
3. According to the deep learning-based microgrid cloud-edge collaborative control method of claim 1, It is characterized in that The characteristic image is a comprehensive characteristic image of the source-load-storage operation status.
4. A microgrid cloud-edge collaborative control method based on deep learning, It is characterized in that The following cloud server steps are included: A1: Build an optimization model and input the operating status data uploaded by the edge controller, generate the source-load-storage control strategy under each tag category and send it to the edge controller; A2: Build a neural network model and input the characteristic image uploaded by the edge controller for training and testing, obtain the neural network model parameters and the confirmation judgment threshold of each label category and send them to the edge controller; The step of constructing a neural network model and inputting the characteristic image uploaded by the edge controller for training and testing, obtaining the neural network model parameters and the confirmation judgment threshold of each label category and sending them to the edge controller includes the following steps: A21: Receive the characteristic image uploaded by the edge controller and set the corresponding label category; A22: The feature image data under each label category is divided into a training set and a test set, where the proportion of the training set is larger than that of the test set; A23: Build a neural network model and train the neural network model using the training set; A24: Use the test set to test the trained neural network model; A25: Send the successfully tested neural network model parameters and the confirmation judgment thresholds of each label category to the edge controller.
5. According to the deep learning-based microgrid cloud-edge collaborative control method of claim 4, It is characterized in that The step of constructing an optimization model and inputting the operation status data uploaded by the edge controller, generating a source-load-storage control strategy under each tag category and issuing the strategy to the edge controller includes the following steps: A11: Receive the operation status data uploaded by the edge controller and set the corresponding label category; A12: Generate corresponding typical operation data for the operation status data of each tag category; A13: Build an optimization model and input typical operation data to obtain the corresponding source-load-storage regulation strategy; A14: Send the source-load-storage regulation strategy and the corresponding label category to the edge controller.
6. According to the deep learning-based microgrid cloud-edge collaborative control method of claim 4, It is characterized in that The training of the neural network model using the training set includes two parts: a forward propagation process and a back propagation process. The forward propagation process includes the following steps: A231: Initialize the neural network model; A232: Input the training set feature images into the neural network model; A233: The neural network model calculates the label category and its probability value; A234: Calculate the deviation between the target value and the probability value, and determine whether it is within the allowable range; if it is within the allowable range, fix the neural network model weight parameters and label category confirmation judgment threshold; if it is not within the allowable range, enter the back propagation process to optimize the weight parameters until the deviation is within the allowable range; The back-propagation process consists of the following steps: A235: End training and save the training model; A236: Calculate the correction amount of the connection weights between the input layer and the hidden layer and the correction amount of the connection weights between the hidden layer and the output layer; A237: Update the connection weights of the input layer, hidden layer and output layer respectively, and execute step A233.
7. According to the deep learning-based microgrid cloud-edge collaborative control method of claim 4, It is characterized in that The method of using the test set to test the trained neural network model includes the following steps: A241: Input the test set feature images into the neural network model one by one to calculate the probability value of each label category; A242: Determine whether the label category corresponding to the highest probability value is the label category to which the characteristic image belongs. If not, return to the training process to retrain the neural network model. If yes, use the highest probability value as the probability value of the label category to which the characteristic image belongs. A243: After obtaining the probability values of the label categories of all images in the test set, determine whether all probability values are greater than a preset threshold; if so, fix the neural network model parameters and send them to the edge controller; if not, return to the training process and retrain the neural network model.
8. The microgrid cloud-edge collaborative control method based on deep learning according to any one of claims 4 to 7, It is characterized in that The operating status data includes the source-load-storage operating status data of the microgrid wind power, photovoltaic, energy storage and load units; the characteristic image is a comprehensive characteristic image of the source-load-storage operating status; and the neural network model adopts a CNN convolutional neural network model.
Citation Information
Patent Citations
ABC-algorithm-based short-term micro-grid load forecasting method
CN108053084A
Micro-power-grid energy storage scheduling method and device based on deep Q-value network (DQN) reinforcement learning
CN109347149A
Electric vehicle fast charging station energy storage system based on source-network-load-storage cooperative service and method thereof
CN110649641A
Model optimization method of edge cloud collaborative support vector machine based on online learning
CN112287990A
Power load prediction method based on modular recurrent neural network
CN112700057A
