An artificial intelligence image recognition method based on deep learning

Abstract

The application relates to an artificial intelligence image recognition method based on deep learning, and belongs to the technical field of image processing. In the application, an optimization algorithm is used to optimize a Back Propagation neural network. Due to the strong global convergence ability and strong robustness of the swarm intelligence optimization algorithm, the slow convergence speed and the shortcoming of falling into a local optimum of the neural network can be well compensated, so that the whole system can accurately judge the shielding condition of a photovoltaic power station caused by cloud movement compared with a traditional artificial neural network prediction method and a cloud shielding prediction method considering the influence of cloud height. Meanwhile, the optimized long short-term memory network model can meet the application requirements of high precision, high efficiency and high stability of photovoltaic ultra-short-term power prediction under full climate conditions, so that efficient and economic operation of solar photovoltaics is possible, and the prediction precision of the ultra-short-term prediction can be effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and relates to an artificial intelligence image recognition method based on deep learning. BACKGROUND

[0002] Solar photovoltaic power generation is developing rapidly worldwide, but the intermittency of solar energy hinders its large-scale grid-connected operation. Most of the super-short-term output power changes of solar energy are caused by irregular movement of clouds, and such mutations can significantly affect the output power within a time scale of several minutes. Short-term prediction can use historical meteorological data and relatively wide time resolution to reduce the influence of such mutations, but the time scale of super-short-term prediction and the shortage of historical data cannot correct such errors. Studies have shown that the size, transmittance and height of sky clouds can greatly affect the solar radiation intensity received by photovoltaic cells. If solar energy is to be fully and efficiently developed and integrated into existing power grids, future cloud changes must be understood. However, the global convergence ability and robustness of common prediction methods are not high enough when used, so that the prediction accuracy of the entire method is not high enough. SUMMARY

[0003] Therefore, the purpose of the present application is to provide an artificial intelligence image recognition method based on deep learning.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0005] An artificial intelligence image recognition method based on deep learning, comprising the following steps:

[0006] S1: Preprocessing the image, first extracting the effective range information, geometric correction does not change the size of the cloud pixels, in order to avoid waste of computing resources, deleting irrelevant areas according to the latitude and longitude coordinates, the super-short-term prediction time scale is N hours, and the movement range of the cloud cluster within the prediction time scale needs to be fully considered when determining the target area;

[0007] S2: Identifying the image features, first performing gray scale processing and threshold segmentation on the cloud image, and converting the color cloud image into a gray scale image; and using a weighted average method to perform gray scale conversion on the standardized satellite cloud image;

[0008] S3: Threshold segmentation; distinguishing the target cloud cluster and the ground through a set threshold value; if the pixel intensity in the gray scale image is greater than the set threshold value, the point is considered to be a cloud pixel point; otherwise, it is other pixel points; because the thickness of the cloud is different from the reflectivity of sunlight, the greater the threshold value, the fewer the number of cloud pixel points identified;

[0009] S4: normalization processing, the complex data information is converted to a certain fixed interval by a certain method and standard; the maximum value of the data is Q max , the minimum value of the data is Q min , the current data value is Q i , the processed value is Q i2 , the data normalization process is as follows:

[0010] Q i2 =(Q i -Q min ) / (Q max -Q min ),

[0011] S5: model training:

[0012] S51: network initialization: determine the number of neurons, give the weight, threshold, learning rate and iteration number parameters;

[0013] S52: input the characteristic value into the model for training, obtain the output value and calculate the error;

[0014] S53: update the weight and threshold according to the error, and calculate again;

[0015] S54: judge whether the error is within the allowable range, if the error is within the allowable range, the training is ended, otherwise return to S52;

[0016] S6: model evaluation, using the data of the test set to test the U-shaped hollow full convolution segmentation network model built and evaluate the model performance;

[0017] S7: using the trained BP neural network model to predict the prediction data, using the power, irradiance and temperature data of the last moment to predict the power data after several minutes, in order to better analyze the photovoltaic power ultra-short term prediction under different seasonal conditions, select one day's prediction result from the 7-day prediction data of each season for analysis, then output the prediction result and end the whole prediction process.

[0018] Optionally, in S4, in order to improve the cloud recognition efficiency, the RGB color value of the strong light area is extracted; through comparative analysis, it is found that the B component color value b of the strong light area is quite different from the color value of other channels, and the RGB color value of the pixels of the cloud layer, especially the thick cloud layer, has no obvious difference; the sky cloud picture obtained through CGI gray scale processing, wherein the lower the gray scale value, the higher the light transmittance and the higher the ground irradiance; similarly, the higher the gray scale value, the lower the light transmittance and the lower the ground irradiance.

[0019] Optionally, in the S5, when threshold segmentation is performed, an Otsu threshold-based image segmentation method is selected to binarize the image, the collected image is denoted as p(x, y), the whole image is divided into foreground and background, the foreground is the detection target, and the background is the image element except the target, the segmentation threshold is denoted as T, the pixel proportions of the foreground and the background in the whole image are denoted as w0 and w1, the average gray scales of the foreground and the background are denoted as μ0 and μ1, and the total average gray scale and the inter-class variance of the image are denoted as μ and g.

[0020] Optionally, in the S5, when threshold segmentation is performed, the image size is M*N, the number of pixels with a pixel gray scale less than the threshold T and the number of pixels with a pixel gray scale greater than the threshold T in the image are denoted as N0 and N1, and the specific calculation is shown in the following formulas (1) to (4), and finally the inter-class variance is simplified as formula (5):

[0021] w0=N0 / N0+N1; w1=N1 / N0+N1 (1)

[0022] N0+N1=M*N (2)

[0023] μ=w0μ0+w1μ1 (3)

[0024] g=w0w1(μ0-μ) 2 +w1(μ 1- μ) 2 (4)

[0025] g=w0w1(μ0-μ) 0- μ1) 2 (5);

[0026] The process of obtaining the maximum inter-class variance is as follows: firstly, an initial threshold T is given at random, the image is divided into two parts of target and background, the average gray scales of the two parts are calculated respectively, and the inter-class variance under the T value is obtained; finally, the threshold is looped from 0 to 255, the inter-class variances of the two parts under all values are calculated, the maximum value is obtained, and the threshold value at this time is the best binarization threshold.

[0027] Optionally, in the S5, Q i is a sample point; Q min is the minimum value in the sample; Q max is the maximum value in the sample; and the normalization method adopted is linear normalization of data to the interval [0, 1].

[0028] Optionally, in the S6, an improved LSTM single neuron model is established, and an association gate is added, so that the LSTM model can directly analyze a plurality of continuous images in a time period of a cloud image to obtain a dynamic abstract information; by performing matrix expansion on the plurality of images, pixel information of each time point based on the cloud image is obtained, and a motion trajectory of a next time period is analyzed to realize feature association of the cloud image.

[0029] Optionally, the image convolution decoder in the LSTM model is a convolution pooling network, batch normalization results are used in each layer, and random pooling is used after each layer to reduce spatial resolution and improve the abstract ability of the model, so that the network is easier to optimize and higher accuracy is obtained; the input CGI cloud image and CPT cloud image pass through the image convolution decoder, combined with local meteorological data, are input into the LSTM model for training to obtain an ultra-short-term prediction output result.

[0030] Optionally, in the S6, when the number of LSTM hidden layers increases, the time used for each iteration is significantly longer, and the prediction error is reduced; the prediction error of the 4-layer LSTM hidden layer structure is reduced by a very small margin compared with the 3-layer LSTM hidden layer structure, and the prediction error of the 5-layer LSTM hidden layer structure is reduced by a relatively obvious margin compared with the 4-layer LSTM hidden layer structure.

[0031] Optionally, in the S6, the model adopts a structure that the input layer is a full connection layer, the hidden layer is a 5-layer LSTM layer, and the output layer is a full connection layer, the number of nodes of each layer is gradually reduced from the first layer to the rear, and the model adopts this structure in subsequent parameter determination.

[0032] The application has the advantages that the BP neural network is optimized by using the optimization algorithm. Since the swarm intelligence optimization algorithm has strong global convergence ability and strong robustness, it can well compensate for the slow convergence speed and local optimal defect of the neural network. Therefore, the combination of the two can complement each other, both the generalization mapping ability of the neural network and the convergence speed and learning ability of the neural network can be improved, so that the whole system can accurately judge the cloud movement shielding condition of the photovoltaic power station compared with the traditional artificial neural network prediction method, and the cloud shielding prediction method considering the influence of cloud height can accurately judge the cloud movement shielding condition of the photovoltaic power station, and the optimized LSTM model can meet the application requirements of high precision, high efficiency and high stability of photovoltaic ultra-short-term power prediction under full climate conditions, so as to provide the possibility of efficient and economic operation of solar photovoltaic and effectively improve the prediction accuracy of ultra-short-term prediction.

[0033] Additional advantages, objects, and features of the application will be apparent to those skilled in the art upon examination of the following detailed description, it being understood that each of the foregoing general statements are true of the particular embodiments of the application yet the particular embodiments thereof are illustrative of the more general principles of the application. The present application will be appreciated more fully from the detailed description below. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred embodiments of the present application will be described in detail below with reference to the drawings, in which:

[0035] Figure 1 System flowchart of the present application. DETAILED DESCRIPTION

[0036] The other advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the present specification. The present application can also be implemented or applied by means of other different specific embodiments, and each detail in the present specification can be modified or changed in different views and applications without departing from the spirit of the present application. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0037] It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0038] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only for illustrative purposes, and cannot be understood as a limitation of the present application, for those skilled in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0039] Reference Figure 1 The image recognition method for the ultra-short-term prediction method based on deep learning includes the following steps:

[0040] S1: Geometric correction is performed on the photographed cloud picture. First, the latitude and longitude coordinates of each pixel point are calculated. Then, the coordinates are associated with the original satellite cloud picture to generate a cloud picture containing geographical information. Finally, the distorted cloud picture is processed in the remote sensing satellite processing platform to complete the geometric correction of the image.

[0041] S2: The image is preprocessed. First, the effective range information is extracted. The geometric correction does not change the size of the satellite cloud image pixels. To avoid wasting computing resources, the irrelevant area needs to be deleted based on the latitude and longitude coordinates. The ultra-short-term prediction time scale is Nh. When determining the target area, the moving range of the cloud cluster within the prediction time scale needs to be fully considered. The method of comparing the sine value of the solar elevation angle with the pixel is used to eliminate the influence of the solar elevation angle on the pixel size, thereby enhancing the brightness of the thick cloud layer at sunrise.

[0042] S3: The image features are identified. First, the cloud picture is grayed and threshold segmented. The color cloud picture is converted into a gray image. The standardized satellite cloud picture is converted to gray by using the weighted average method.

[0043] S4: Threshold segmentation is performed. The target cloud cluster and the ground are distinguished by setting a threshold. If the pixel intensity in the gray image is greater than the set threshold, it is considered to be a cloud pixel point; otherwise, it is other pixel points. Because the thickness of the cloud affects the reflectivity of sunlight, the larger the threshold, the fewer the number of cloud pixel points identified.

[0044] S5: Normalization is performed. The process of converting complex data information to a fixed interval by a certain method and standard is as follows: the maximum data value is Q max , the minimum data value is Q min , the current data value is Q i , and the processed data value is Q i2 . The data normalization process is as follows:

[0045] Q i2 =(Q i -Q min ) / (Q max -Q min )

[0046] S6: Model training:

[0047] Step 1: Network initialization: determine the number of neurons, give the weight, threshold, learning rate, iteration number, and other parameters;

[0048] Step 2: input the feature value into the model for training, get the output value and calculate the error;

[0049] Step 3: update the weight and threshold according to the error, and calculate again;

[0050] Step 4: determine whether the error is within the allowable range, if the error is within the allowable range, the training is ended, otherwise return to step 2.

[0051] S7: model evaluation, using the data of the test set to test the U-shaped cavity full convolution segmentation network model built and evaluate the model performance,

[0052] S8: using the trained BP neural network model to predict the prediction data, using the power, irradiance and temperature data of the last moment to predict the power data after several minutes, in order to better analyze the photovoltaic power ultra-short-term prediction under different seasonal conditions, select one day of prediction results from the 7-day prediction data of each season for analysis, and then output the prediction results, that is, the whole prediction process can be ended.

[0053] In step S3, in order to improve the cloud recognition efficiency, the RGB color value of the strong light area is extracted. Through comparative analysis, it is concluded that the B component color value b of the strong light area is quite different from the color value of other channels, while the RGB color value of each channel of the cloud layer, especially the thick cloud layer, has no obvious difference; the sky cloud picture obtained through CGI grayscale processing, wherein the lower the grayscale color value, the higher the light transmittance of the region, and the higher the ground irradiance at this time. Similarly, the higher the grayscale value, the lower the light transmittance of the region, and the thicker the cloud layer, the solar radiation is not easy to penetrate the cloud layer, and the ground irradiance is also lower.

[0054] In step S4, when threshold segmentation is performed, the Otsu threshold-based image segmentation method is selected to binarize the image. Let the collected image be p(x,y), the whole image is divided into foreground and background, the foreground is the detection target, and the background is the image element except the target, the segmentation threshold is set as T, the pixel proportion of the foreground and the background in the total image is set as w0 and w1 respectively, the average grayscale of the foreground and the background is set as μ0 and μ1 respectively, and the total average grayscale of the image and the inter-class variance are set as μ and g respectively.

[0055] In step S4, when threshold segmentation is performed, the image size is M*N, and the number of pixels with pixel grayscale less than the threshold T and the number of pixels with pixel grayscale greater than the threshold T in the image are denoted as N0 and N1 respectively, which are calculated as shown in the following formula (1) to formula (4), and finally simplified to obtain the inter-class variance as formula (5):

[0056] w0=N0 / N0+N1;w1=N1 / N0+N1(1)

[0057] N0+N1=M*N(2)

[0058] μ=w0μ0+w1μ1(3)

[0059] g=w0(μ 0- μ)2 + w1(μ 1- μ) 2 (4)

[0060] g = w0w1(μ 0- μ1) 2 (5);

[0061] The process of obtaining the maximum inter-class variance is that, first, an initial threshold T is given at random, the image is divided into target and background two parts, and then the gray mean values of the two parts are calculated to obtain the inter-class variance under the T value. Finally, the threshold is looped from 0 to 255, the inter-class variance of the two parts under all values is calculated, and the maximum value is obtained, at this time, the threshold is the best binary threshold.

[0062] In the step S5, Q i is a sample point; Q min is the minimum value in the sample; and Q max is the maximum value in the sample. The normalization method used in the present study is to linearly normalize the data to the interval [0, 1].

[0063] In the step S6, an improved LSTM single neuron model is established, and an associative gate is added. This special structure enables the LSTM model to directly analyze a plurality of continuous images in a time period of a cloud image to obtain a dynamic abstract information. Through matrix expansion of the plurality of images, pixel information at each time point based on the cloud image is obtained, so that the motion trajectory of the next time period is analyzed to realize feature association of the cloud image.

[0064] The image convolution decoder in the structure of the improved LSTM model is a convolutional pooling network, batch normalization results are used in each layer, and in addition, random pooling is used after each layer to reduce the spatial resolution and improve the abstract ability of the model, so that the network is easier to optimize and has higher accuracy. The input CGI cloud image and CPT cloud image pass through the image convolution decoder, are combined with local meteorological data, are input into the LSTM model for training, and an ultra-short-term prediction output result is obtained.

[0065] In the step S6, when the number of LSTM hidden layers increases, the time used for each iteration is significantly longer, and the prediction error is reduced. The prediction error of the 4-layer LSTM hidden layer structure is reduced by a very small margin compared with that of the 3-layer LSTM hidden layer structure, and the prediction error of the 5-layer LSTM hidden layer structure is reduced by a relatively obvious margin compared with that of the 4-layer LSTM hidden layer structure.

[0066] In the step S6, the model adopts a structure in which the input layer is a full connection layer, the hidden layer is a 5-layer LSTM layer, and the output layer is a full connection layer, the number of nodes in each layer is gradually reduced from the first layer to the rear, and the model adopts this network structure in subsequent parameter determination.

[0067] In the present application, the BP neural network is optimized by using an optimization algorithm. Because the swarm intelligence optimization algorithm has strong global convergence ability and strong robustness, it can well make up for the slow convergence speed and the shortcoming of falling into local optimum of the neural network. Therefore, the combination of the two can complement each other, both can exert the generalization mapping ability of the neural network, and can also improve the convergence speed and learning ability of the neural network, so that the whole system can accurately judge the shielding condition of the cloud movement to the photovoltaic power station compared with the traditional artificial neural network prediction method, and the cloud shielding prediction method considering the influence of cloud height. At the same time, the optimized LSTM model can meet the application requirements of high precision, high efficiency and high stability of photovoltaic ultra-short-term power prediction under all weather conditions, so as to provide the possibility for efficient and economic operation of solar photovoltaic and effectively improve the prediction accuracy of ultra-short-term prediction.

[0068] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, which should be covered by the scope of the claims of the present application.

Claims

1. A deep learning-based artificial intelligence image recognition method, characterized in that: The method comprises the following steps: S1: preprocessing the image, first extracting the effective range information, geometric correction does not change the size of the cloud image pixels, in order to avoid waste of computing resources, according to the latitude and longitude coordinates, the irrelevant area is deleted, the ultra-short-term prediction time scale is N hours, and the moving range of the cloud cluster in the prediction time scale is fully considered when determining the target area; S2: identify the image features, first, the cloud image is processed by gray scale and threshold segmentation, and the color cloud image is converted into a gray scale image; the standardized satellite cloud image is converted into a gray scale image by using the weighted average method; S3: threshold segmentation; the target cloud cluster and the ground are distinguished by setting the threshold value; if the pixel intensity in the gray scale image is greater than the set threshold value, it is considered that the point is a cloud pixel point; otherwise, it is other pixel points; because the thickness of the cloud is different from the reflectivity of the sunlight, the greater the threshold value, the fewer the number of identified cloud pixel points; S4: normalization, the process of converting complex data information to a certain fixed interval by certain methods and standards; the maximum data value is Q max , the minimum data value is Q min , the current data value is Q i , the processed data value is Q i2 , the data normalization process is as follows: Q i2 = (Q i - Q min ) / (Q max - Q min ), S5: model training: S51: network initialization: determine the number of neurons, give the weight, threshold value, learning rate and iteration number parameters; S52: input the feature value into the model for training, obtain the output value and calculate the error; S53: update the weight and threshold value according to the error, and calculate again; S54: judge whether the error is within the allowable range, if the error is within the allowable range, the training is ended, otherwise, return to S52; S6: model evaluation, the data of the test set are used to test the U-shaped hollow full convolution segmentation network model and evaluate the model performance; S7: using the trained BP neural network model to predict the prediction data, using the power, irradiance and temperature data of the last moment to predict the power data after several minutes, in order to better analyze the photovoltaic power ultra-short-term prediction under different seasonal conditions, one day of prediction results is selected from the 7-day prediction data of each season for analysis, and then the prediction results are output, and the whole prediction process is ended.

2. The artificial intelligence image recognition method based on deep learning according to claim 1, characterized in that: In S4, in order to improve the cloud recognition efficiency, the RGB color value of the strong light area is extracted; through comparative analysis, it is found that the B component color value b of the strong light area is quite different from the color values of other channels, and the RGB color values of the cloud layer, especially the thick cloud layer, are not obviously different; the sky cloud image obtained through CGI gray scale processing, wherein the lower the gray scale color value, the higher the light transmittance and the higher the ground irradiance; similarly, the higher the gray scale color value, the lower the light transmittance and the lower the ground irradiance.

3. The artificial intelligence image recognition method based on deep learning according to claim 2, characterized in that: In S5, when threshold segmentation is performed, the Otsu method based on threshold value is selected to binarize the image, the collected image is p(x, y), the whole image is divided into foreground and background, the foreground is the detection target, the background is the image element except the target, the segmentation threshold value is set as T, the pixel proportion of the foreground and the background in the total image is set as w0 and w1 respectively, the average gray scale of the foreground and the background is set as μ0 and μ1 respectively, and the total average gray scale of the image and the inter-class variance are set as μ and g respectively.

4. The artificial intelligence image recognition method based on deep learning according to claim 3, characterized in that: In the S5, when threshold segmentation is performed, the image size is M*N, and the number of pixels with a pixel grayscale less than the threshold T and the number of pixels with a pixel grayscale greater than the threshold T in the image are denoted as N0 and N1 respectively, and the specific calculation is shown in the following formulas (1) to (4), and finally the inter-class variance is obtained as formula (5): w0=N0 / N0+N1; w1=N1 / N0+N1 (1) N0+N1=M*N (2) μ=w0μ0+w1μ1 (3) g = w0(μ0 - μ) 2 + w1(μ1 - μ) 2 (4) g = w0w1(μ0- μ1) 2 (5); The process of obtaining the maximum inter-class variance is as follows: firstly, an initial threshold T is given at random, the image is divided into two parts of target and background, the grayscale mean values of the two parts are calculated respectively to obtain the inter-class variance under the T value; finally, the threshold is looped from 0 to 255, the inter-class variances of the two parts under all values are calculated, and the maximum value is obtained, and the threshold value at this time is the optimal binary threshold.

5. The artificial intelligence image recognition method based on deep learning according to claim 4, characterized in that: In S5, Q i is the sample point; Q min is the minimum value in the sample; Q max is the maximum value in the sample; the normalization method used is linear normalization of the data to the interval [0, 1].

6. The artificial intelligence image recognition method based on deep learning according to claim 5, characterized in that: In the S6, an improved LSTM single neuron model is established, and an associative gate is added, so that the LSTM model can directly analyze a plurality of continuous images in a time period of a cloud image to obtain a dynamic abstract information; the pixel information of each time point based on the cloud image is obtained by matrix expansion of the plurality of images, the motion trajectory of the next time period is analyzed, and the feature association of the cloud image is realized.

7. The artificial intelligence image recognition method based on deep learning according to claim 6, characterized in that: The image convolution decoder in the LSTM model is a convolution pooling network, batch normalization results are used in each layer, random pooling is used after each layer to reduce the spatial resolution, the abstract ability of the model is improved, the network is easier to optimize, and higher accuracy is obtained; the input CGI cloud image and CPT cloud image are input into the LSTM model for training after the image convolution decoder and in combination with local meteorological data, and an ultra-short-term prediction output result is obtained.

8. The artificial intelligence image recognition method based on deep learning according to claim 7, characterized in that: In the S6, when the number of LSTM hidden layers increases, the time used for each iteration is significantly longer, and the prediction error is reduced; the prediction error of the 4-layer LSTM hidden layer structure is reduced by a very small margin compared with that of the 3-layer LSTM hidden layer structure, and the prediction error of the 5-layer LSTM hidden layer structure is reduced by a relatively obvious margin compared with that of the 4-layer LSTM hidden layer structure. 9.The deep learning-based artificial intelligence image recognition method of claim 8, wherein: In the S6, the model adopts a structure in which the input layer is a full connection layer, the hidden layer is a 5-layer LSTM layer, and the output layer is a full connection layer, the number of nodes of each layer is gradually reduced from the first layer to the rear, and the model adopts this structure in subsequent parameter determination.

Images