Photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification

Abstract

The application discloses a photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification. First, a ground-based cloud classification dataset, a multi-modal meteorological dataset and a photovoltaic power generation power dataset are constructed; then, a ground-based cloud classification model and a photovoltaic power generation power prediction model are constructed; the ground-based cloud classification model comprises a visual main network, a first multi-modal subnetwork, a second multi-modal subnetwork, a dense fusion module and a multi-modal feature fusion network; the photovoltaic power generation power prediction model comprises a fluctuation power prediction submodel and a smooth power prediction submodel, and the sum of the fluctuation power and the smooth power prediction value is obtained to obtain a photovoltaic power generation power prediction value; finally, the ground-based cloud classification model and the photovoltaic power generation power prediction model are trained, and the trained model is used for photovoltaic power generation power prediction. The method fully considers the influence of cloud types on photovoltaic power generation on the basis of multi-modal meteorological data, and improves the prediction accuracy of photovoltaic power generation power.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power prediction, and particularly relates to a photovoltaic power prediction method based on multi-modal ground-based cloud classification. BACKGROUND

[0002] In recent years, with the transformation of global energy structure and the requirement of sustainable development, the proportion of coal-fired power generation is decreasing year by year, and the proportion of clean energy power generation such as solar energy is increasing year by year. Photovoltaic power generation has the characteristics of intermittency, randomness and volatility, which will affect the stable operation of the power system and further affect the quality of power supply when large-scale photovoltaic power generation is connected to the power grid.

[0003] Among the many variables that affect the efficiency of photovoltaic power generation, solar irradiance is one of the main influencing factors, and its uncertainty directly leads to the volatility of photovoltaic power generation. The dynamic change of clouds and their shielding effect on solar radiation make them a key factor directly affecting solar irradiance. Therefore, it is necessary to consider the influence of cloud types on photovoltaic power generation.

[0004] At present, photovoltaic power generation power is mainly predicted based on historical meteorological data, photovoltaic data and multi-physical field simulation. Most of them directly extract key features of cloud images from ground-based cloud images, including aerosol scattering coefficient, sky cloud cover, light intensity, sky color information and texture information, etc. However, the influence of cloud types on photovoltaic power generation is not considered, and the influence of meteorological factors on cloud formation is not considered, which leads to insufficient mining of cloud features. The formation of clouds is a process affected by many natural factors, including temperature, humidity, air pressure, wind speed, etc. Different types of clouds may be similar in visual images, but the meteorological parameters are significantly different. Therefore, combining meteorological parameters can comprehensively describe the cloud features and improve the accuracy of cloud classification, which is of great significance to photovoltaic power prediction. SUMMARY

[0005] In view of the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a photovoltaic power prediction method based on multi-modal ground-based cloud classification.

[0006] The technical problem of the present application is solved by adopting the following technical scheme:

[0007] A photovoltaic power prediction method based on multi-modal ground-based cloud classification, comprising the following steps:

[0008] Step 1: Construct ground-based cloud classification dataset, multi-modal meteorological dataset and photovoltaic power generation power dataset;

[0009] Second step: constructing a ground-based cloud classification model, including a visual main network, a first multi-modal sub-network, a second multi-modal sub-network, a dense fusion module and a multi-modal feature fusion network; the visual main network includes a plurality of four-channel feature extraction modules, the first and second multi-modal sub-networks each include a plurality of multi-modal modules connected in series, and the multi-modal modules correspond one-to-one to the four-channel feature extraction modules; the output features of the current four-channel feature extraction module are fused with the output features of the corresponding multi-modal modules of the first multi-modal sub-network and the second multi-modal sub-network through two dense fusion modules, and the output features of the two dense fusion modules are multiplied with the output features of the current four-channel feature extraction module according to the channel to obtain the input features of the next four-channel feature extraction module; the output features of the last four-channel feature extraction module are sequentially subjected to average pooling, layer normalization and a fully connected layer to obtain the output features of the visual main network; the multi-modal feature fusion network fuses the ground-based cloud visual features extracted by the visual main network, the temperature and humidity features extracted by the first multi-modal sub-network, and the air pressure and wind speed features extracted by the second multi-modal sub-network, and the output features of the multi-modal feature fusion network are subjected to a fully connected layer to obtain the cloud classification result;

[0010] Third step: constructing a photovoltaic power prediction model, which predicts the photovoltaic power according to multi-modal meteorological data, cloud classification results and historical photovoltaic power;

[0011] Fourth step: training the ground-based cloud classification model and the photovoltaic power prediction model, and using the trained models for photovoltaic power prediction.

[0012] Further, the photovoltaic power prediction model includes a fluctuation power prediction sub-model and a smooth power prediction sub-model; the input features of the photovoltaic power prediction model are decomposed into high-frequency components and low-frequency components through wavelet transform, the high-frequency components are subjected to the fluctuation power prediction sub-model to obtain a fluctuation power prediction value, and the low-frequency components are subjected to the smooth power prediction sub-model to obtain a smooth power prediction value; the fluctuation power prediction value and the smooth power prediction value are summed to obtain a photovoltaic power prediction value.

[0013] Further, the fluctuation power prediction sub-model and the smooth power prediction sub-model have the same architecture, and each includes a feature extraction network and a long short-term memory network; the feature extraction network includes a 3D convolution layer, a batch normalization layer, a ReLU activation function, a maximum pooling layer and a Dropout layer connected in sequence, and the output features of the feature extraction network are used as the input features of the long short-term memory network.

[0014] Further, the first channel of the four-channel feature extraction module is a residual connection, the second channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 1 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features, the third channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 2 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features, and the fourth channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 3 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features; the output features of the four channels are added and then pass through a channel attention mechanism to obtain the output features of the four-channel feature extraction module.

[0015] Further, in the dense fusion module, the output features of the four-channel feature extraction module pass through global average pooling and are connected head-to-tail with the output features of the multi-modal module to obtain spliced features; the spliced features sequentially pass through a fully connected layer, an ELU activation function, a fully connected layer and a Sigmoid activation function to obtain the output features of the dense fusion module.

[0016] Further, the multi-modal feature fusion network is divided into two branches, one branch includes a plurality of fusion modules connected in series, and a Dropout layer is embedded between adjacent fusion modules, and the other branch includes a fusion module, and the output features of the two branches are added to obtain the output features of the multi-modal feature fusion network.

[0017] Further, the fusion module includes a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence.

[0018] Further, the multi-modal module includes a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence.

[0019] Compared with the prior art, the present application has the following beneficial effects:

[0020] The cloud classification result is used for photovoltaic power prediction, the influence of cloud types on photovoltaic power generation is fully considered on the basis of multi-modal meteorological data, and the prediction accuracy is improved. In order to improve the cloud classification accuracy, on the basis of fully extracting the visual features of the ground cloud image by the visual main network, the multi-modal meteorological features are extracted by the multi-modal sub-network, and the heterogeneous fusion of the ground cloud image visual features and the multi-modal meteorological features is realized through the dense fusion module. The ground cloud image visual features are combined with the multi-modal meteorological features, the ground cloud image visual features are used as main information, and the multi-modal meteorological features are used as auxiliary information for cloud classification, which can significantly improve the model's ability to distinguish similar clouds, improve the classification accuracy of similar clouds, and further improve the prediction accuracy of photovoltaic power generation. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1is a schematic diagram of the overall process of the present application;

[0022] Figure 2 is a structural schematic diagram of the ground cloud classification model of the present application;

[0023] Figure 3 is a structural schematic diagram of the dense fusion module of the present application;

[0024] Figure 4 is a structural schematic diagram of the photovoltaic power prediction model of the present application. DETAILED DESCRIPTION

[0025] Specific embodiments will be described below with reference to the accompanying drawings, which are used only to specifically introduce the technical solutions of the present application, and do not limit the protection scope of the present application.

[0026] The present application provides a photovoltaic power prediction method based on multi-modal ground cloud classification, comprising the following steps:

[0027] Step 1: Obtain historical ground cloud, historical multi-modal meteorological data and historical photovoltaic power, and obtain ground cloud classification data set, multi-modal meteorological data set and photovoltaic power data set after preprocessing;

[0028] The preprocessing of the ground cloud includes normalization, noise reduction and removal of ghosting and direct sunlight area in the image; the preprocessing of the multi-modal meteorological data and the photovoltaic power includes data cleaning and standardization processing; the multi-modal meteorological data includes temperature, humidity, air pressure and wind speed;

[0029] Step 2: Build a ground cloud classification model; as shown in Figure 2 The ground cloud classification model includes a visual main network, a first multi-modal sub-network, a second multi-modal sub-network, a dense fusion module and a multi-modal feature fusion network; wherein the visual main network is used to extract the visual features of the ground cloud, the first multi-modal sub-network is used to extract the temperature and humidity features, the second multi-modal sub-network is used to extract the air pressure and wind speed features, the dense fusion module is used to realize the fusion of the ground cloud visual features and the multi-modal meteorological features in the middle layer, and the multi-modal feature fusion network fully fuses the ground cloud visual features, the temperature and humidity features, the air pressure and wind speed features, and the output features of the multi-modal feature fusion network are input into the full connection layer to obtain the cloud classification result.

[0030] The visual main network comprises a plurality of four-channel feature extraction modules (five in the embodiment), the output features of the current four-channel feature extraction module are respectively fused with the output features of the multi-modal modules corresponding to the first and second multi-modal sub-networks through two dense fusion modules, and the output features of the two dense fusions are multiplied with the output features of the current four-channel feature extraction module by channel to serve as the input features of the next four-channel feature extraction module; the output features of the last four-channel feature extraction module are sequentially subjected to average pooling, layer normalization and a full connection layer to obtain the ground-based cloud image visual features. The first channel of the four-channel feature extraction module adopts a short-chain branch, which aims to directly take the input of the module as the output to realize residual feature extraction and prevent the accuracy from decreasing with the deepening of the network; the second channel sequentially uses a standard deep convolution (with an expansion rate of 1) with a convolution kernel size of 3x3 and a point-by-point convolution with a convolution kernel size of 1x1 to extract features, which can not only extract small cloud boundary features and relatively abstract cloud cover features, but also keep the output and input dimensions consistent to facilitate the fusion of multi-channel ground-based cloud image features; the third channel sequentially uses a deep convolution with a convolution kernel size of 3x3 and an expansion rate of 2 and a point-by-point convolution with a convolution kernel size of 1x1 to extract features, and the fourth channel sequentially uses a deep convolution with a convolution kernel size of 3x3 and an expansion rate of 3 and a point-by-point convolution with a convolution kernel size of 1x1 to extract features; the output features of the four channels are added and then subjected to a channel attention mechanism to obtain the output features of the four-channel feature extraction module. The deep convolution with a larger expansion rate has a larger receptive field and can inspect the information in a larger neighborhood range of the input image, and is more likely to extract global features; the use of different sizes of convolution kernels in each channel enables each channel to obtain information with different receptive fields, so as to extract features of the cloud object with rich levels; at the same time, the appropriate use of deep convolution can ensure that the depth of the network is reduced when extracting features of the same level, reduce the redundant calculation existing in the residual connection, speed up the inference speed of the network and improve the image processing accuracy; the channel attention mechanism enables the network to focus on useful information in each channel.

[0031] In the ground-based cloud image classification task, temperature and humidity have more significant influence on the classification result than pressure and wind speed, so the classification result based on temperature and humidity information is given greater weight, and therefore temperature, humidity, pressure and wind speed are divided into two groups, temperature and humidity are taken as the input of the first multi-modal subnetwork, and pressure and wind speed are taken as the input of the second multi-modal subnetwork, and temperature and humidity features, pressure and wind speed features are extracted respectively. The first and second multi-modal subnetworks have the same architecture but different parameters, and both include multiple multi-modal modules connected in sequence, the number of which is the same as the number of four-channel feature extraction modules of the visual main network, and the multi-modal modules of the two multi-modal subnetworks correspond one-to-one to the four-channel feature extraction modules of the visual main network, and the multi-modal meteorological features are extracted by the multi-modal subnetworks to provide auxiliary information for the ground-based cloud image classification task, which helps to distinguish cloud types with similar visual features but different meteorological features and improve the classification accuracy of similar clouds. The multi-modal module includes a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence, and a part of neurons in the fully connected layer of each multi-modal module is randomly deleted to enhance the generalization ability of the model and prevent overfitting.

[0032] As shown in Figure 3 In the dense fusion module, the output feature V of the corresponding four-channel feature extraction module of the visual main network is subjected to global average pooling to obtain a global average pooling feature V'; the global average pooling feature V' and the output feature M of the corresponding multi-modal module are connected head to tail to obtain a splicing feature, and the splicing feature is sequentially subjected to a fully connected layer, an ELU activation function, a fully connected layer and a Sigmoid activation function to obtain an output feature of the dense fusion module.

[0033] The ground-based cloud image visual feature, temperature and humidity feature, pressure and wind speed feature are added together as the input feature of the multi-modal feature fusion network. The multi-modal feature fusion network has two branches, one branch includes a plurality of fusion modules connected in series, and a Dropout layer is embedded between adjacent fusion modules, and the other branch consists of one fusion module, and the output features of the two branches are added together to obtain the output feature of the multi-modal feature fusion network; through the two branches of the multi-modal feature fusion network, not only the information flow can be optimized, but also the regularization effect can be indirectly realized, and the stability and reliability of the model can be effectively improved. The fusion module consists of a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence.

[0034] Step 3: Construct a photovoltaic power prediction model, including a fluctuation power prediction submodel and a smooth power prediction submodel; as Figure 4As shown, the current time multi-modal weather data, cloud classification results and the photovoltaic power at each time of the last time step form the input features of the photovoltaic power prediction model, the input features are decomposed into high-frequency components and low-frequency components through wavelet transform, the high-frequency components are input into the fluctuation power prediction sub-model, which mainly depends on the cloud classification results to predict the fluctuation power caused by cloud cover; the low-frequency components are input into the smooth power prediction sub-model, which mainly depends on the multi-modal weather data and the historical change trend of photovoltaic power to predict the smooth power; the sum of the fluctuation power and the smooth power prediction values is obtained to obtain the photovoltaic power prediction value.

[0035] The fluctuation power prediction sub-model and the smooth power prediction sub-model have the same architecture but different parameters, and both include a feature extraction network and a long short-term memory network; the feature extraction network includes a 3D convolution layer, a batch normalization layer, a ReLU activation function, a max pooling layer and a Dropout layer connected in turn, and the dropout rate of the Dropout layer is set to 0.5, that is, 50% of the neurons are randomly shielded to prevent overfitting or underfitting during training; the output features of the feature extraction network are captured through the long short-term memory network to capture the time correlation, and the model output is obtained.

[0036] Fourth step: using the ground-based cloud image classification dataset, the multi-modal weather dataset and the photovoltaic power dataset, training the ground-based cloud image classification model and the photovoltaic power prediction model, and using the trained ground-based cloud image classification model and the photovoltaic power prediction model for photovoltaic power prediction of the photovoltaic power station;

[0037] During the training process, the deviation between the classification results of the ground-based cloud image classification model and the true labels is estimated by using a double-guided loss function; the expression of the double-guided loss function is:

[0038] (1)

[0039] (2)

[0040] (3)

[0041] wherein, is the number of cloud categories, is the adjustment coefficient of the cloud category is the prediction result of the ground-based cloud image, if the cloud category is the true label of the ground-based cloud image, then , otherwise 0; represents the prediction probability of the cloud category , and is the output of the neuron of the last fully connected layer of the ground-based cloud image classification model for predicting the cloud category .​ is a cloud category The classification accuracy of the cloud category with low classification accuracy is given a higher weight;

[0042] The double-guided loss function is improved from the traditional cross-entropy loss function, different weights are assigned to different cloud categories according to the difficulty of the ground-based cloud chart classification, and the model is forced to pay more attention to the difficult-to-classify cloud categories during the training process; when the ground-based cloud chart sample belongs to the easy-to-classify cloud category, the loss is almost unchanged; when the ground-based cloud chart sample belongs to the difficult-to-classify cloud category, the loss also increases due to the increase of the weight of the difficult-to-classify cloud category; at the same time, the double-guided loss function can adjust the influence of the difficult-to-classify cloud category by using the adjustment coefficient When , the double-guided loss function degenerates into the traditional cross-entropy loss function; when , the influence of the adjustment coefficient increases, so when the sample is misclassified, it will be severely punished, forcing the model to pay more attention to the difficult-to-classify cloud categories.

[0043] The unmentioned part of the application is applicable to the prior art.

Claims

1. A photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification, characterized in that, The method comprises the following steps: Step 1: constructing a ground-based cloud image classification dataset, a multi-modal meteorological dataset and a photovoltaic power dataset; Step 2: constructing a ground-based cloud image classification model, which comprises a visual main network, a first multi-modal sub-network, a second multi-modal sub-network, a dense fusion module and a multi-modal feature fusion network; the visual main network comprises a plurality of four-channel feature extraction modules; the first and second multi-modal sub-networks each comprise a plurality of multi-modal modules connected in series, and the multi-modal modules correspond one-to-one to the four-channel feature extraction modules; the output features of a current four-channel feature extraction module are fused with the output features of the multi-modal modules corresponding to the first multi-modal sub-network and the second multi-modal sub-network through two dense fusion modules, respectively; the output features of the two dense fusion modules are multiplied with the output features of the current four-channel feature extraction module according to channels to obtain the input features of a next four-channel feature extraction module; the output features of the last four-channel feature extraction module are sequentially subjected to average pooling, layer normalization and a fully connected layer to obtain the output features of the visual main network; the multi-modal feature fusion network fuses the ground-based cloud image visual features extracted by the visual main network, the temperature and humidity features extracted by the first multi-modal sub-network and the air pressure and wind speed features extracted by the second multi-modal sub-network; the output features of the multi-modal feature fusion network are subjected to a fully connected layer to obtain a cloud classification result; Step 3: constructing a photovoltaic power prediction model, which predicts the photovoltaic power according to the multi-modal meteorological data, the cloud classification result and historical photovoltaic power; The photovoltaic power prediction model comprises a fluctuation power prediction sub-model and a smooth power prediction sub-model; the input features of the photovoltaic power prediction model are decomposed into high-frequency components and low-frequency components through wavelet transform; the high-frequency components are subjected to the fluctuation power prediction sub-model to obtain a fluctuation power prediction value, and the low-frequency components are subjected to the smooth power prediction sub-model to obtain a smooth power prediction value; the fluctuation power prediction value and the smooth power prediction value are summed to obtain a photovoltaic power prediction value; Step 4: training the ground-based cloud image classification model and the photovoltaic power prediction model, and using the trained models for photovoltaic power prediction.

2. The photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification according to claim 1, characterized in that, The fluctuation power prediction sub-model and the smooth power prediction sub-model have the same architecture, and each comprises a feature extraction network and a long short-term memory network; the feature extraction network comprises a 3D convolution layer, a batch normalization layer, a ReLU activation function, a maximum pooling layer and a Dropout layer connected in sequence; the output features of the feature extraction network are used as the input features of the long short-term memory network.

3. The photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification according to claim 1, characterized in that, The first channel of the four-channel feature extraction module is a residual connection, the second channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 1 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features, the third channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 2 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features, and the fourth channel sequentially uses a deep convolution with a convolution kernel size of 3*3 and an expansion rate of 3 and a point-by-point convolution with a convolution kernel size of 1*1 to extract features; the output features of the four channels are added and then pass through a channel attention mechanism to obtain the output features of the four-channel feature extraction module.

4. The photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification according to claim 1, characterized in that, In the dense fusion module, the output features of the four-channel feature extraction module pass through global average pooling and are connected head-to-tail with the output features of the multi-modal module to obtain spliced features; the features obtained by splicing sequentially pass through a fully connected layer, an ELU activation function, a fully connected layer and a Sigmoid activation function to obtain the output features of the dense fusion module.

5. The photovoltaic power production forecasting method based on multi-modal ground-based cloud classification according to claim 1, 3 or 4, characterized in that, The multi-modal feature fusion network is divided into two branches, one branch includes a plurality of fusion modules connected in series, and a Dropout layer is embedded between adjacent fusion modules, and the other branch includes one fusion module; the output features of the two branches are added to obtain the output features of the multi-modal feature fusion network.

6. The photovoltaic power generation power prediction method based on multi-modal ground-based cloud classification according to claim 5, characterized in that, The fusion module includes a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence.

7. The photovoltaic power production forecasting method based on multi-modal ground-based cloud classification according to claim 1, characterized in that, The multi-modal module includes a fully connected layer, a batch normalization layer and a GELU activation function connected in sequence.

Images