A wind speed prediction method and device fusing ground-based cloud images

Abstract

The present application belongs to the technical field of wind power generation, and specifically relates to a wind speed prediction method and device fusing ground-based cloud images. The method comprises the following steps: S1, collecting measured wind speed of a wind farm and corresponding time-based ground-based cloud image time series data, and dividing a training set and a test set; S2, extracting cloud movement information from continuous ground-based cloud images; S3, identifying sky and cloud pixels in the ground-based cloud images, and calculating cloud coverage; S4, constructing a cloud movement vector prediction model, taking historical cloud movement information extracted from the ground-based cloud images as input, taking cloud movement vectors in a future period of time as a target variable, and training a regression model; and S5, building a wind speed prediction model, taking historical wind speed data and estimated future cloud speed as input, taking wind speed data in a future period of time as output, training the wind speed prediction model, and calculating an evaluation index of the predicted wind speed. The present application overcomes the shortcomings of single input data and low information quantity of a wind speed prediction model, and avoids the problem of limited model prediction performance caused by NWP uncertainty and non-locality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind power generation, and particularly relates to a wind speed prediction method and device fusing ground-based cloud images. BACKGROUND

[0002] Wind energy is a renewable, clean and widely distributed energy source, and wind power generation is the third largest electric energy source after water power and thermal power, which has large-scale development technical conditions. However, the volatility and intermittence of wind energy result in the volatility of wind power, and large-amplitude wind power fluctuation has an adverse effect on power balance and frequency regulation of the power grid. Therefore, the power generation prediction of the wind farm is crucial to the normal operation of the power system.

[0003] Wind power generation depends on weather factors to a great extent, and therefore, the NWP model becomes the basic model for wind power generation prediction. However, the NWP data has uncertainty, and the inherent error of the input data limits the learning ability of the model; at the same time, the NWP data involves a wide area and cannot reflect the local information of the weather value of the wind farm. Therefore, the wind power prediction model with NWP as the input will produce greater prediction error.

[0004] Since the wind power is positively correlated with the cube of the wind speed, researchers indirectly realize wind power prediction by constructing a wind speed prediction model. The related technology for wind speed prediction relying only on historical wind speed data is constantly improving. However, the prediction model constructed by this kind of method has single input data and low information quantity, which limits the accuracy improvement of the wind speed prediction model.

[0005] The continuous ground-based cloud image collected above the wind farm contains the cloud movement information above the wind farm, which has strong correlation with the wind speed data of the wind farm and can better reflect the local characteristics of the wind farm. Therefore, the ground-based cloud image is used to assist the historical wind speed sequence to predict the wind speed, which not only increases the input information of the prediction model, but also avoids the problem that the model prediction performance is limited by the uncertainty and non-locality of the NWP. SUMMARY

[0006] The application provides a wind speed prediction device and method of a wind farm fusing ground-based cloud images, extracts the cloud movement information above the wind farm from the collected continuous ground-based cloud image, combines the historical wind speed sequence to predict the wind speed, and further realizes more accurate wind power generation prediction.

[0007] The application adopts the following technical scheme: a wind speed prediction method fusing ground-based cloud images, comprising the following steps:

[0008] S1: collecting the measured wind speed of the wind farm and the ground-based cloud image time series data at the corresponding moment, and dividing a training set and a test set;

[0009] S2: extracting cloud movement information from the continuous ground-based cloud image;

[0010] S3: Identify the sky and cloud pixels in the ground-based cloud image, and calculate the cloud coverage rate;

[0011] S4: Build a cloud motion vector prediction model, use the historical cloud motion information extracted from the ground-based cloud image as input, and the cloud motion vector in the future period of time as the target variable to train the regression model;

[0012] S5: Build a wind speed prediction model, use the historical wind speed data and the estimated future cloud speed as input, and the wind speed data in the future period of time as output, train the wind speed prediction model, and calculate the evaluation index of the predicted wind speed.

[0013] In step S1, the ground-based cloud image time series data is collected by the all-sky imager, and the collected image is stored in the upper computer.

[0014] Step S2 includes:

[0015] S21: Read the original ground-based cloud image, set the center of the image as the origin, set the radius to retain the sky and cloud area in the image, remove the noise information around the shooting environment, replace the noise area with black pixel points, and perform mask processing on the image;

[0016] S22: Image enhancement;

[0017] S23: Extract feature descriptors from the ground-based cloud image after image enhancement in step S22;

[0018] S24: Match the feature descriptors extracted from adjacent ground-based cloud images, record the Euclidean distance of the optimal match and suboptimal match points, and discard the feature points when the optimal distance is greater than the threshold of the suboptimal distance;

[0019] S25: Filter the cloud image matching feature points to obtain a further corrected feature point set;

[0020] S26: Calculate the cloud motion vector according to the position coordinates and time information of the matching feature point set.

[0021] Step S3 includes:

[0022] S31: According to the RGB color features of the ground-based cloud image, construct two new features , The operation value after multiplying the square root of the two is used to distinguish cloud and sky pixels;

[0023] S32: Get the cloud coverage rate according to the ratio of cloud pixels to total pixel number, that is:

[0024]

[0025] In the formula, representing cloud coverage, is the sum of the number of cloud pixels and clear sky pixels, representing the number of cloud pixels, the number of cloud pixels and clear sky pixels The calculation is specifically expressed as:

[0026]

[0027] wherein, is a counter, is a pixel point, The value of is 0 or 1.

[0028] The construction of two new features in step S31 includes:

[0029] a) The difference between the R color feature value and the minimum value of the G and B color feature values, expressed as:

[0030]

[0031] b) The reciprocal of the red-to-blue ratio R / B, expressed as:

[0032]

[0033] wherein r, g, b respectively represent the RGB color space values of the ground-based cloud image.

[0034] Step S4 includes:

[0035] S41: Construct a cloud motion information sliding window, set the sequence length of a single window, divide the cloud motion information of each day into windows, and divide the window input and output with time t as the boundary;

[0036] S42: The data after time t in the cloud motion information sliding window is taken as the model output, and the data before time t is taken as the model input to train an LSTM regression model about the future time cloud motion vector.

[0037] Step S5 includes:

[0038] S51: Construct a wind speed data sliding window, and divide the window input and output with time t as the boundary;

[0039] S52: Build an LSTM model for wind speed prediction in clear sky; a hybrid prediction model with MLP and LSTM architecture is used for wind speed prediction in non-clear sky, first, input the input data into the MLP network to realize feature fusion of cloud information and historical wind speed information, and connect the output of the MLP network to the LSTM network to realize time series regression prediction;

[0040] S53: according to the calculated cloud coverage determine whether it is a clear sky;

[0041] S54: when it is not a clear sky, first, the cloud speed at the future time is predicted according to the cloud motion vector prediction model, and then, the wind speed at the future time is predicted according to the wind speed prediction optimization model constructed in advance by the strong correlation between the wind speed at the corresponding time and the cloud speed at the future time. t k

[0042]

[0043] In the formula, is a prediction model containing input X and parameter , is wind speed data containing historical records, the real wind speed at the future d time and the predicted cloud speed k , time information , wind speed is obtained by the cloud motion vector prediction model with historical cloud motion information as input; T When it is a clear sky, an LSTM model is constructed, historical wind speed data is taken as input, and future wind speed is taken as output, and a wind speed prediction optimization model in advance k steps is constructed:

[0044]

[0045] .

[0046] In step S53, when the cloud coverage is greater than the set threshold 0.1, it is considered that the sky is not clear, and when the cloud coverage is less than the set threshold 0.1, it is considered that the sky is clear.

[0047] In step S5, the evaluation index includes root mean square error and mean absolute error, and the calculation formulas are respectively:

[0048]

[0049]

[0050] In the above formula, m represents the number of test sets, represents the i-th actual wind speed value in the test set, represents the i-th wind speed prediction value of the model.

[0051] A wind speed prediction device fusing ground-based cloud images, comprising:

[0052] ​​​​​​Temperature and humidity sensor, used to detect the temperature and humidity in the device;

[0053] PI heating film, used to prevent the remaining devices from being damaged or stopped working when the temperature in the device is too low;

[0054] Raspberry Pi, used to control the data acquisition of the device and the allowed temperature and humidity range in the device;

[0055] Camera module and fisheye lens, used to collect ground cloud data;

[0056] Relay, used to control the operation or stop of the PI heating film according to the Raspberry Pi signal;

[0057] Switching power supply, outputting 24V and 5V DC voltage, 24V voltage is used for power supply of the relay and the PI heating film, and 5V voltage is used for power supply of the Raspberry Pi;

[0058] Host computer, used to store and process data, and the host computer uses the cloud motion information and wind speed prediction method of fusing ground cloud as claimed in claims 1-9 to extract the cloud motion information and predict the wind speed.

[0059] Compared with the prior art, the present application provides a wind speed prediction device and method fusing ground cloud, which collects wind speed data and ground cloud time series data set of a wind farm, takes the measured wind speed as the target variable, takes the extracted ground cloud information and historical wind speed as the input, and trains a hybrid wind speed prediction model based on MLP and LSTM. The present application extracts cloud motion information above the wind farm from the collected continuous ground cloud, fully utilizes the high correlation between wind speed and cloud motion information, and predicts the wind speed value of the wind farm to realize more accurate wind power prediction. The present application not only overcomes the shortcomings of single input data and low information quantity of the wind speed prediction model, but also avoids the problem that the prediction performance of the model is limited due to the uncertainty and non-locality of NWP. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A wind speed prediction device fusing ground cloud provided for an embodiment of the present application;

[0061] Figure 2 A whole flowchart of a wind speed prediction method fusing ground cloud provided for an embodiment of the present application;

[0062] Figure 3 A flowchart of extracting cloud motion information above a wind farm provided for an embodiment of the present application;

[0063] Figure 4 A ground cloud grayscale image provided for an embodiment of the present application;

[0064] Figure 5 A ground cloud image after image enhancement and mask impurity information provided for an embodiment of the present application;

[0065] Figure 6 The ground cloud feature descriptor diagram at the moment provided by the embodiment of the application is as shown in the following figure: t-1 The ground cloud feature descriptor diagram at the moment provided by the embodiment of the application is as shown in the following figure:

[0066] Figure 7 The preliminary screening feature matching effect diagram provided by the embodiment of the application is as shown in the following figure: t The ground cloud feature descriptor diagram at the moment provided by the embodiment of the application is as shown in the following figure:

[0067] Figure 8 The preliminary screening feature matching effect diagram provided by the embodiment of the application is as shown in the following figure:

[0068] Figure 9 The feature fine matching effect diagram provided by the embodiment of the application is as shown in the following figure:

[0069] Figure 10 The wind speed sequence noise distribution diagram provided by the embodiment of the application is as shown in the following figure:

[0070] Figure 11 The MLP+LSTM hybrid prediction model architecture diagram provided by the embodiment of the application is as shown in the following figure:

[0071] Figure 12 The wind speed prediction model flowchart provided by the embodiment of the application is as shown in the following figure:

[0072] Figure 13 The test set wind speed prediction result visualization diagram provided by the embodiment of the application is as shown in the following figure. DETAILED DESCRIPTION

[0073] The application provides a wind speed prediction device and method fusing ground cloud, which is further described in detail below in combination with the drawings and specific embodiments.

[0074] Figure 1 The wind speed prediction device fusing ground cloud provided by the embodiment of the application is as shown in the following figure: Figure 1 As shown in the figure, the device comprises an Arducam IMX477 camera module, an Arducam 220° fisheye lens, a Raspberry Pi, a DHT22 temperature and humidity sensor, a PI heating film, a double-way relay, a switching power supply, a hemispherical optical lens, a waterproof shell and the like. The Raspberry Pi is a core device, which is used to control data acquisition of the device and a temperature and humidity range allowed in the device; the Arducam IMX477 camera module is matched with the Arducam 220° fisheye lens to collect ground cloud data; the relay is used to control work or stop of the PI heating film according to a signal of the Raspberry Pi; the DHT22 temperature and humidity sensor is used to detect temperature and humidity in the device; the switching power supply outputs 24V and 5V direct current voltage; the hemispherical optical lens and the waterproof shell are respectively used to protect the fisheye lens and the internal device; and the upper computer is used to store and process data, extract cloud motion information, predict wind speed and the like.

[0075] Figure 2 The wind speed prediction method provided by the embodiment of the present application is shown in the whole flowchart of the method, which comprises the following processes:

[0076] S1: Establish a database;

[0077] S1: Establish a database;

[0078] S2 and S3: Extract the feature information of the ground-based cloud image;

[0079] The cloud movement information and the cloud coverage rate are extracted by using the image processing technology. For example, Figure 3 The flowchart for extracting the cloud movement information above the wind farm provided by the embodiment of the present application comprises the following processes:

[0080] a) Image preprocessing: read the original ground-based cloud image, set a certain radius (560) around the center of the image to retain the sky and cloud area in the image, remove the noise information such as obstacles around the shooting environment, replace the noise area with black pixel points, and perform mask processing on the image.

[0081] b) Image enhancement: the local histogram equalization algorithm is used to realize image enhancement, the histogram distribution of the image is changed to remove noise, and the overall characteristics of the image are improved. For example, Figure 4 and Figure 5 The comparison chart before and after image enhancement provided by the embodiment of the present application, Figure 4 is the ground-based cloud image gray chart, Figure 5 is the image after image enhancement and mask impurity information.

[0082] c) The SURF algorithm is used to extract the feature descriptor of the adjacent ground-based cloud image after image enhancement. For example, Figure 6 and Figure 7 The feature descriptor of the ground-based cloud image at time 1 and t- 1 provided by the embodiment of the present application is shown in the schematic diagram. t

[0083] d) The feature descriptors extracted from the adjacent ground-based cloud images are matched based on the FLANN algorithm, and the Euclidean distance of the optimal matching and suboptimal matching points is recorded. When the optimal distance is greater than 0.6 times the suboptimal distance, the feature point is discarded to reduce the calculation complexity in the RANSAC algorithm precise matching process. For example, Figure 8 is the feature matching effect chart after preliminary screening provided by the embodiment of the present application.

[0084] ​e) According to the RANSAC algorithm, the cloud map matching feature points are filtered to obtain a further corrected feature point set. Figure 9 The feature matching effect diagram after the RANSAC algorithm optimization provided by the embodiment of the application.

[0085] f) According to the position coordinates and time information of the matching feature point set, the cloud motion vector is calculated.

[0086] In the embodiment of the application, according to the RGB color features of the ground-based cloud map, two new features 、 are constructed, and the operation value after the product of the two is square rooted is used to identify the cloud pixels in the ground-based cloud map.

[0087]

[0088]

[0089] wherein r, g, b respectively represent the ground-based cloud map RGB color space value, and the operation value after the product of the two is square rooted is used to distinguish the cloud and sky pixels. Secondly, the cloud coverage is obtained according to the ratio of the cloud pixels to the total number of pixels, that is:

[0090]

[0091] In the formula, represents the cloud coverage, is the sum of the cloud pixel and the number of clear sky pixels, represents the number of cloud pixels, and the calculation of the number of cloud pixels and the number of clear sky pixels is specifically expressed as:

[0092]

[0093] wherein, is a counter, is a pixel point, the value of which is 0 or 1.

[0094] S4: Constructing a cloud motion vector prediction model;

[0095] Taking the historical cloud motion information extracted from the ground-based cloud map as input and the cloud motion vector at the future time as output, the LSTM regression model is trained, and in the embodiment, a double-layer LSTM prediction model with 128 neurons is built, and the activation function is tanh. The steps of constructing the cloud motion vector prediction model are:

[0096] a) According to the sliding window technology, a cloud motion information sliding window is constructed, the sequence length of a single window is set, the cloud motion information of each day is divided in the form of a window, and the time point t is taken as the boundary to divide the window input and output.

[0097] b) An LSTM regression model is built, the data after the time point in the cloud motion information sliding window is taken as the model output, t the data before the time point is taken as the model input, and the LSTM regression model about the future time point cloud motion vector is trained. t

[0098] A wind speed prediction model is built and trained.

[0099] In the wind speed data collection process, various noises will be introduced due to the influence of working environment, sensor material properties and the like. In the embodiment of the present application, the wavelet threshold denoising technology is used to remove the noise, a fixed threshold , N is taken as the sample quantity, and the wind speed sequence is divided into two levels using the wavelet basis db2. Figure 10 is a noise distribution diagram existing in the actual wind speed sequence. When the sky is clear, the wind speed prediction model is an LSTM model, and when the sky is not clear, the wind speed prediction model is a hybrid prediction model with MLP and LSTM as the architecture. As Figure 11 is the hybrid prediction model architecture provided by the embodiment of the present application, the MLP network is used to fuse features, and the LSTM realizes the regression prediction of the time sequence. As Figure 12 is the wind speed prediction flowchart of the model provided by the embodiment of the present application, whether the sky is clear or not, i.e. whether there is cloud or not, is determined through the cloud coverage rate at time t. When the cloud coverage rate is greater than the set threshold value 0.1, it is considered that the sky is not clear, and the future wind speed is predicted by combining the future time cloud speed and the historical wind speed sequence. When the cloud coverage rate is less than the set threshold value 0.1, it is considered that the sky is clear, i.e. there is no cloud time, and the cloud motion information with high correlation with the wind speed cannot be obtained, at this time, only the LSTM model is trained by relying on the historical wind speed to realize the wind speed prediction. The specific steps of building the wind speed prediction model are:

[0100] a) According to the sliding window technology, a wind speed data sliding window is constructed, the wind speed data of each day is divided in the form of a window, and the time point t is taken as the boundary to divide the window input and output. In the embodiment of the present application, the input sequence length of the wind speed prediction model is 20, which means that the current time input is composed of the historical information of the past 20 minutes, and the output sequence length is 1, which means that the future 1 minute wind speed value is predicted.

[0101] ​b) Construct an LSTM model for wind speed prediction under clear skies; use a hybrid prediction model based on MLP and LSTM for wind speed prediction under non-clear skies. The MLP model is a two-layer architecture with 64 and 128 neurons respectively and a linear activation function; the LSTM model has two stacked layers with 128 neurons each and a tanh activation function.

[0102] c) Determine whether the sky is clear based on the calculated cloud cover rate. If the cloud cover rate is greater than a set threshold... At that time, the sky is considered not clear when the cloud cover is less than a set threshold. At that time, they believed the sky was clear.

[0103] d) Maintaining consistency with the prediction range of the cloud motion vector prediction model, when the sky is not clear, first predict the cloud velocity at future times based on the cloud motion vector prediction model, and then... Corresponding t Wind speed at all times Based on the strong correlation, construct an optimized wind speed prediction model k steps in advance:

[0104]

[0105] In the formula, For input X and parameters Predictive models, For inclusion d Historical wind speed data, future k The actual wind speed at that moment With predicted cloud speed Time information T Wind speed The cloud motion vector prediction model F is obtained by taking historical cloud motion information as input.

[0106] In clear skies, cloud motion information highly correlated with wind speed is unavailable. Therefore, cloud motion information extracted from ground-based cloud images is discarded, and an LSTM model is constructed, using historical wind speed data as input and future wind speed as output. A wind speed prediction optimization model is built, predicting wind speed k steps in advance.

[0107]

[0108] In this embodiment, the model loss is the mean squared error (MSE) loss function, the optimizer is the Adam optimizer, the initial learning rate is 0.01, and the learning rate decreases continuously as the number of iterations increases. The batch size is set to 64, and the iterations are 50.

[0109] The test wind speed prediction model is used to determine whether it is sunny according to cloud coverage, the test set is input into the trained accurate LSTM prediction model when it is sunny, the test set is input into the trained accurate MLP+LSTM hybrid prediction model when it is not sunny, and then the wind speed prediction value is obtained, and the multiple evaluation indexes of the predicted wind speed are calculated. In the embodiment of the application, the test set data is also divided into window forms by using the sliding window technology, input into the trained accurate wind speed prediction model, and the wind speed prediction value at the future time is obtained, Figure 13 The test set wind speed prediction result visualization diagram provided by the embodiment of the application. Finally, the multiple evaluation indexes of the predicted wind speed are calculated by using the measured wind speed of the wind farm, including: RMSE, MAE, and the calculation formulas are respectively:

[0110]

[0111]

[0112] In the above formula, m represents the number of test sets, represents the i th actual wind speed value in the test set, represents the i th wind speed prediction value of the model. In the embodiment, the MAE evaluation index of the standardized test set is 0.013, and the RMSE evaluation index is 0.026.

Claims

1. A method of fusing a cloud base chart for wind speed prediction, characterized by, The method comprises the following steps: S1: collecting the measured wind speed of the wind farm and the corresponding time ground cloud time series data, and dividing the training set and the test set; S2: extracting cloud motion information from continuous ground cloud images; Step S2 comprises: S21: reading the original ground cloud image, taking the center of the image as the origin, setting the radius to retain the sky and cloud area in the image, removing the noise information around the shooting environment, replacing the noise area with black pixel points, and performing image mask processing; S22: image enhancement; S23: extracting feature descriptors from the ground cloud image after image enhancement in step S22; S24: matching the feature descriptors extracted from adjacent ground cloud images, recording the Euclidean distance of the optimal matching and suboptimal matching points, and discarding the feature descriptors when the optimal distance is greater than the threshold value of the suboptimal distance; S25: filtering the cloud image matching feature points to obtain a further corrected feature point set; S26: calculating the cloud motion vector according to the position coordinates and time information of the matching feature point set; S3: identifying the sky and cloud pixels in the ground cloud image, and calculating the cloud coverage rate; S4: constructing a cloud motion vector prediction model, taking the historical cloud motion information extracted from the ground cloud image as the input, and taking the cloud motion vector in the future period of time as the target variable, training the regression model; S5: building a wind speed prediction model, taking the historical wind speed data and the estimated future cloud speed as the input, and taking the wind speed data in the future period of time as the output, training the wind speed prediction model, and calculating the evaluation index of the predicted wind speed.

2. The method of claim 1, wherein the wind speed prediction is a fusion of ground-based cloud maps. In the step S1, the ground cloud time series data is collected by a full-sky imager, and the collected image is stored in an upper computer.

3. The method of claim 1, wherein the wind speed prediction is a fusion of ground-based cloud maps. The step S3 comprises: S31: Construct two new features according to the RGB color features of the ground cloud image , , the operation value after the product is square rooted for distinguishing cloud and sky pixels; S32: obtaining the cloud coverage rate according to the ratio of the number of cloud pixels to the total number of pixels, that is: In the formula, represents the cloud coverage, is the sum of the number of cloud pixels and the number of clear sky pixels, represents the number of cloud pixels, the number of cloud pixels and the number of clear sky pixels The calculation is specifically expressed as: wherein, is a counter, is a pixel point, has a value of 0 or 1.

4. The method of claim 1, wherein the wind speed prediction is a fusion foundation cloud map. The step S31 comprises: a) the difference between the R color feature value and the minimum value of the G and B color feature values, expressed as: b) the reciprocal of the red-to-blue ratio R / B, expressed as: wherein r, g, b represent the RGB color space values of the ground cloud image.

5. The method of claim 1, wherein the wind speed prediction is a fusion foundation cloud map. The step S4 comprises: S41: Construct a cloud motion information sliding window, set the sequence length of a single window, divide the cloud motion information of each day in the form of a window, and divide the window input and output with time as the boundary. t as the boundary. S42: sliding window of cloud motion information t data after the time instance as model output, t data before the time instance as model input, train the LSTM regression model on cloud motion vector at future time instance.

6. The method of claim 1, wherein the wind speed prediction is a fusion foundation cloud map. The step S5 comprises: S51: Construct a wind speed data sliding window to time point t Divide the window input and output as boundaries; S52: constructing an LSTM model for wind speed prediction in clear sky; a hybrid prediction model based on MLP and LSTM is used for wind speed prediction in non-clear sky, first, the input data is input into the MLP network to realize feature fusion of cloud information and historical wind speed information, the output of the MLP network is connected with the LSTM network to realize regression prediction of time series; S53: determining the cloud coverage based on the calculation determining whether the sky is clear; S54: When the sky is not clear, firstly, the cloud speed at future times is predicted based on the cloud motion vector prediction model; secondly, based on the cloud motion speed... Corresponding t Wind speed at all times Strong correlation, construct in advance k Step-by-step wind speed prediction optimization model: wherein, is a prediction model comprising input X and parameters , is a wind speed data comprising d a number of historical records, real wind speed k at future time instants and predicted cloud velocity , time information T , wind speed is obtained by a cloud motion vector prediction model with historical cloud motion information as input; In clear sky, an LSTM model is constructed, historical wind speed data is taken as input, and future wind speed is taken as output, and a wind speed prediction optimization model for k steps in advance is constructed: 。 7. The method of claim 5, wherein the wind speed prediction is based on a fusion of ground-based cloud images. In the step S53, when the cloud coverage rate is greater than the set threshold value 0.1, the sky is considered to be non-clear sky, and when the cloud coverage rate is less than the set threshold value 0.1, the sky is considered to be clear sky.

8. The method of claim 5, wherein the wind speed prediction is a fusion foundation cloud map. In the step S5, the evaluation index includes: root mean square error and mean absolute error, and the calculation formulas are respectively: In the above formula, m denotes the number of test sets, denotes the i-th actual wind speed value in the test set, denotes the i-th wind speed prediction value of the model.

9. A device for predicting wind speed from a fusion of base cloud maps, characterized in that, It comprises: A temperature and humidity sensor is used to detect the temperature and humidity in the device; A PI heating film is used to prevent other components from being damaged or stopped working when the temperature in the device is too low; A Raspberry Pi is used to control the data acquisition of the device and the allowable temperature and humidity range in the device; Camera module and fisheye lens for collecting ground cloud data; Relay for controlling the work or stop of the PI heating film according to the Raspberry Pi signal; Switching power supply, outputting 24V and 5V DC voltage, the 24V voltage is used for power supply of the relay and the PI heating film, and the 5V voltage is used for power supply of the Raspberry Pi; Host computer for storing and processing data, the host computer uses the cloud motion information extraction and wind speed prediction method of the fused ground cloud as claimed in claims 1-8 to predict the wind speed.

Images