CN113011645A - Power grid strong wind disaster early warning method and device based on deep learning - Google Patents

Abstract

A method and device for early warning of power grid wind disaster based on deep learning, collecting radar data as input data, inputting it into a pre-trained wind speed prediction model, and outputting wind speed information; the wind speed information is used as the input element of the pre-trained wind disaster early warning model The wind disaster early warning model combines the data of the automatic meteorological observation station in the same period to correct the wind speed information, and outputs the wind speed revision information after excluding the errors caused by the micro-topographic information. Location warning solves the problem of fine-grained forecasting of wind damage in the power grid. Strong convective winds form a spatial resolution of 1km × 1km and forecast results in 0-120 minutes; a business operation software system is formed, which further improves the early warning and forecast accuracy of wind disasters, and improves the power grid and The pertinence of equipment for disaster prevention and mitigation and the efficiency of emergency rescue have improved the safe operation level of the power grid and the reliability of social electricity consumption.

Description

A method and device for early warning of power grid gale disaster based on deep learning

technical field

The invention relates to the technical field of weather forecasting, and more particularly, to a method and device for early warning of a power grid gale disaster based on deep learning.

Background technique

The normal operation of the power grid is closely related to the meteorological conditions. The meteorological changes have a significant impact on the peak occurrence time and peak value of the power grid load, and meteorological disasters also threaten the safe operation of the power grid. Therefore, it is of great significance to effectively and advance forecast and early warning of the meteorological environment and meteorological disasters of power grid operation, which is an effective way to achieve disaster prevention and mitigation.

Among meteorological disasters, strong winds have the most serious impact on the transmission lines in the power grid system. Usually, strong wind disasters include: causing towers to collapse, interrupting power transmission, blowing foreign objects to cause line short-circuits, and causing linkage faults to cause larger power outages, wind power The abnormality causes the line to dance, induces wind deflection and flashover, and reduces the success rate of reclosing.

In the prior art, a Chinese patent (CN109902885B) proposes a "typhoon prediction method based on a deep learning hybrid CNN-LSTM model" to predict whether a typhoon will form and its path and intensity after it is formed; the Chinese patent application (CN109492756A) proposes " Multi-element conductor galloping early warning method and related devices based on deep learning" to obtain effective early warning information of transmission line galloping; Chinese patent application (CN109447315A) proposes "a numerical forecasting method for power meteorological elements" based on the fusion of power meteorological data and medium Based on the short-term forecast results and radar monitoring data, the deep convolutional neural network method is used to identify the characteristic information of the radar echo puzzle. Finally, based on the characteristic information, the deep learning method is used to analyze the power and meteorological elements. Carry out short-term forecasting and obtain short-term forecasting results; Chinese patent (CN106126896B) proposes "a hybrid model wind speed forecasting method and system based on empirical mode decomposition and deep learning", obtains the original wind speed time series, and decomposes it according to the empirical mode The original wind speed time series is decomposed to obtain multiple eigenmode functions, and the corresponding wind speed prediction sub-models are used for prediction to obtain the predicted output value of each wind speed prediction sub-model, and then combined and superimposed to obtain the final overall prediction output. The Chinese patent (CN103413174B) proposes a "short-term wind speed multi-step prediction method based on deep learning method", which uses a deep neural network regression model with a multi-input and multi-output structure to perform multi-step prediction of wind speed.

In the prior art, the characteristics of strong winds generated by different weather systems are very different, and their early warning and forecasting methods are also different. It is difficult to automatically identify different types of strong winds by using the three-dimensional echo intensity data observed by radar, and it is necessary to identify and track the convective system. , extraction of the main features of the squall line, identification of the squall line, etc. Then, according to the judgment of the development stage of the squall line, the judgment of the shape of the squall line, etc., analyze the time, location and level of the strong wind. The impact of different types of strong winds on power production and equipment is not only related to wind speed, but also to wind direction, as well as to the duration of the wind, the instantaneous change of wind speed and wind direction, and so on. Existing gale forecasting and early warning technologies are mostly aimed at public services, but there are few professional gale disaster warning services for power grid production, and there is no classification for wind damage in different production stages of the power grid, and there is no systematic gale warning and forecasting method and graded forecasting products. .

SUMMARY OF THE INVENTION

In order to solve the deficiencies in the prior art, the purpose of the present invention is to provide a power grid gale disaster early warning method and device based on deep learning, directly using Doppler radar data as data input, and gale related information as output, through According to the actual operation of the power grid and the influence of micro-topography, the wind speed is corrected, and the corrected wind speed is used as the wind speed for wind disaster warning.

The present invention adopts the following technical solutions.

The steps of a deep learning-based grid wind disaster early warning method are as follows:

Step 1, collect radar data to form input data;

Step 2, input the input data into the pre-trained wind speed prediction model; wherein the wind speed prediction model is obtained after training the sample data by using a deep learning algorithm, and the wind speed prediction model can output wind speed information according to the input sample data;

Step 3: Obtain the wind speed information output by the wind speed prediction model, as an input element of the pre-trained wind disaster early warning model, and input it into the wind disaster early warning model; wherein the wind disaster early warning model combines the data of the automatic meteorological observation station in the same period to fit the wind speed information Corrected, the wind disaster early warning model can output the revised wind speed information after excluding the errors caused by the micro-topographic information;

Step 4, obtain wind speed revision information, and give early warning of high wind disasters.

Preferably, in step 1, the radar data includes the maximum echo intensity obtained from the current body scan echo, the height corresponding to the maximum echo intensity, the echo top height of the convective cell, the time variation of the echo intensity, the radial velocity, Geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude.

Preferably,

In step 2, the training method of the wind speed prediction model includes:

Step 2.1: Collect historical radar data as the input sample data for the wind speed prediction model. The historical radar data includes the maximum echo intensity obtained from the historical volume sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective cell, and the echo Time-varying wave intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude;

Step 2.2, obtaining wind speed information for characterizing wind speed and wind direction according to the input sample data, and using the wind speed information as the output sample data of the wind speed prediction model;

Step 2.3, combine the input sample data and the output sample data into sample data, the sample data includes the maximum echo intensity obtained from the historical volume sweep echo, the corresponding height of the maximum echo intensity, the echo top height of the convective monomer, the echo Time-varying intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude, and historical volume sweep echo correspondence wind speed and direction;

Step 2.4, combine multiple sample data into a sample data set, use all the input sample data in the sample data set as the input data of the wind speed prediction model, and use all the output sample data in the sample data set as the output data of the wind speed prediction model, from the bottom layer. To the top layer, unsupervised learning is carried out layer by layer, and the weight parameters of each layer are trained;

Step 2.5, based on the weight parameters of each layer obtained by training, apply the wake-sleep algorithm, carry out supervised learning layer by layer from the top layer to the bottom layer, and adjust the weight parameters of each layer;

Step 2.6: Input all the input sample data in the sample data set into the trained and adjusted wind speed prediction model, and obtain the test data output by the trained and adjusted wind speed prediction model; when the error between the test data and the output sample data satisfies For the preset requirements, the trained and adjusted model is used as the final available wind speed prediction model; when the error between the test data and the output sample data cannot meet the preset requirements, return to step 2.4 to train and adjust the model.

The wind speed prediction model is a deep learning neural network model, including an input layer, 3 hidden layers, and an output layer; the input layer includes 32 network nodes, the output layer includes 3 network nodes, and the hidden layer goes from the input layer to the output layer. The number of network nodes is 16, 8 and 4 in sequence; the learning rate is set to 0.001.

The hidden layer adopts a network structure composed of a stack of 3 individually trained restricted Boltzmann machines.

The activation function of the hidden layer adopts the sigmoid function, and the activation function of the output layer adopts the softmax function.

Preferably, the sample data needs to be normalized before being input into the model.

Preferably,

In step 3, the training method of the wind disaster early warning model includes:

Step 3.1, collect the data of automatic meteorological observation stations in the same period in the wind speed prediction area and the micro-topographic information of the wind speed prediction area;

Step 3.2, take the difference between the wind speed information and the data of the automatic meteorological observation station in the same period as the objective function;

Step 3.3, constructing a micro-terrain correction coefficient according to the micro-terrain information;

Step 3.4, take the optimal value of the objective function as the training target, and adjust the micro-terrain correction coefficient;

In step 3.5, the final usable wind disaster warning model is constructed by using the adjusted micro-terrain correction coefficient pair.

The micro-terrain correction coefficient satisfies the following relationship:

S=1+K ₁ K ₂ K ₃

In the formula,

K ₁ represents the terrain category parameter, which is 2.2H/L and 1.4H/L for mountain peaks and slopes respectively, where H is the height of the mountain, L is the width of the windward side of the mountain,

K ₂ represents the horizontal coordinate parameter, which satisfies K ₂ =1-|x|/L, where x is the horizontal coordinate from the top of the mountain,

K ₃ represents the height coordinate parameter, which satisfies K ₃ =1-z/2.5H, where z is the height from the mountain surface,

When z>2.5H, let S=1;

When H/L>0.3, take H/L=0.3.

A power grid wind disaster early warning device based on deep learning includes: a collection module, a wind speed prediction module, a wind disaster early warning module and an output module;

The acquisition module is used to collect radar data as input data; the radar data includes the maximum echo intensity obtained from the current body sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective monomer, and the time-varying echo intensity. , radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to the automatic station, mesocyclone, reflectivity, latitude and longitude;

The wind speed prediction module is used to input the input data into the pre-trained wind speed prediction model; the wind speed prediction model is obtained after training the sample data by using the deep learning algorithm, and the wind speed prediction model can output the wind speed information according to the input sample data. ;

The wind disaster early warning module is used to obtain the wind speed information output by the wind speed prediction model, which is used as the input element of the pre-trained wind disaster early warning model and input into the wind disaster early warning model. By fitting and correcting, the wind disaster warning model can output the wind speed revision information after excluding the errors caused by the micro-topographic information;

The output module is used to obtain wind speed revision information and send high wind disaster graded warnings to users.

Compared with the prior art, the invention has the beneficial effects that, compared with the prior art, it realizes the graded early warning and forecast of gale disasters, provides real-time graded gale position early warning, and solves the problem of fine forecasting of wind damage in the power grid. 1km spatial resolution, 0-120 minutes forecast results; a business operation software system is formed, which further improves the early warning and forecast accuracy of wind disasters, improves the pertinence of power grid and equipment disaster prevention and mitigation, and emergency rescue efficiency, and improves the level of safe operation of the power grid and social security. Electricity reliability.

Description of drawings

1 is a flowchart of a deep learning-based grid wind disaster early warning method of the present invention;

2 is a schematic diagram of a framework of a wind speed prediction model and a wind disaster early warning model in a deep learning-based grid wind disaster early warning method of the present invention;

FIG. 3 is a schematic structural diagram of a power grid gale disaster early warning device based on deep learning of the present invention.

Detailed ways

The present application will be further described below with reference to the accompanying drawings. The following examples are only used to more clearly illustrate the technical solutions of the present invention, and cannot be used to limit the protection scope of the present application.

As shown in Figure 1, the steps of a deep learning-based grid wind disaster early warning method are as follows:

Step 1: Collect radar data to form input data.

Specifically, in step 1, the radar data includes, but is not limited to, the maximum echo intensity obtained from the current body scan echo, the height corresponding to the maximum echo intensity, the echo top height of the convective cell, the time variation of the echo intensity, Radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude.

In the wind speed prediction stage, the deep learning deep neural network input parameters have a total of 31 characteristic parameters, and those skilled in the art select different types and different quantities of characteristic parameters as the input parameters of the deep neural network according to actual needs. Input parameters are a non-limiting preference.

Step 2, as shown in Figure 2, input the input data into the pre-trained wind speed prediction model; the wind speed prediction model is obtained after training the sample data by using a deep learning algorithm, and the wind speed prediction model can be based on the input sample data. Output wind speed information.

specifically,

In step 2, the training method of the wind speed prediction model includes:

Step 2.1: Collect historical radar data as the input sample data for the wind speed prediction model. The historical radar data includes the maximum echo intensity obtained from the historical volume sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective cell, and the echo Time-varying wave intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude;

Step 2.2, obtaining wind speed information for characterizing wind speed and wind direction according to the input sample data, and using the wind speed information as the output sample data of the wind speed prediction model;

Step 2.3, combine the input sample data and the output sample data into sample data, the sample data includes the maximum echo intensity obtained from the historical volume sweep echo, the corresponding height of the maximum echo intensity, the echo top height of the convective monomer, the echo Time-varying intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude, and historical volume sweep echo correspondence wind speed and direction;

Step 2.4, combine multiple sample data into a sample data set, use all the input sample data in the sample data set as the input data of the wind speed prediction model, and use all the output sample data in the sample data set as the output data of the wind speed prediction model, from the bottom layer. To the top layer, unsupervised learning is carried out layer by layer, and the weight parameters of each layer are trained;

In this preferred embodiment, the uncalibrated data is used to train the parameters of each layer hierarchically, which is an unsupervised training process. Restrictions and sparsity constraints, so that the resulting model can learn the structure of the data itself, so as to obtain features that are more representative than the input; after learning to obtain the n-1th layer, the output of the n-1 layer is used as the nth layer The input of the nth layer is trained, and the parameters of each layer are obtained respectively.

Step 2.5, based on the weight parameters of each layer obtained by training, apply the wake-sleep algorithm, carry out supervised learning layer by layer from the top layer to the bottom layer, and adjust the weight parameters of each layer;

In this preferred embodiment, the labeled data is used for training, the error is transmitted from top to bottom, and the network is fine-tuned. Supervised training process; the first step is similar to the random initialization initial value process of neural network. Since the first step of deep learning is not random initialization, but obtained by learning the structure of the input data, this initial value is closer to the global optimum, Thereby a better effect can be achieved.

Step 2.6: Input all the input sample data in the sample data set into the trained and adjusted wind speed prediction model, and obtain the test data output by the trained and adjusted wind speed prediction model; when the error between the test data and the output sample data satisfies For the preset requirements, the trained and adjusted model is used as the final available wind speed prediction model; when the error between the test data and the output sample data cannot meet the preset requirements, return to step 2.4 to train and adjust the model.

In actual training, if all layers are trained at the same time, the time complexity will be too high; if one layer is trained each time, the deviation will be transmitted layer by layer, because there are too many neurons and parameters in the deep neural network, it will cause serious problems overfitting. Therefore, the present preferred embodiment proposes the above-mentioned training method for the wind speed prediction model.

This training method makes the weights between the other layers except the top one bidirectional, so that the top layer is still a single-layer neural network, and the other layers become graph models. The upward weight is used for cognition, and the downward weight is used for generation. The Wake-Sleep algorithm is used to adjust all the weights so that cognition and generation are consistent, that is, to ensure that the generated top-level representation can restore the bottom layer as accurately as possible. 's node.

The Wake-Sleep algorithm is divided into two parts: wake and sleep. in:

(1) Wake stage: cognitive process, through external features and upward weights, namely cognitive weights, an abstract representation of each layer is generated, expressed as a node state, and gradient descent is used to modify the downward weights between layers, that is Generate weights.

(2) Sleep stage: The generation process, through the top-level representation, that is, the concepts learned in the wake stage, and the downward weight, generate the state of the bottom layer, and at the same time modify the upward weight between layers.

Specifically, the wind speed prediction model is a deep learning neural network model, including an input layer, three hidden layers, and an output layer; wherein, the input layer includes 32 network nodes, the output layer includes 3 network nodes, and the hidden layer runs from the input layer to the output layer. The number of network nodes in the direction of the output layer is 16, 8 and 4; the learning rate is set to 0.001.

The hidden layer adopts a network structure composed of three individually trained Restricted Boltzmann Machines (RBM) stacked. Bottom-up unsupervised pre-training between deep neural networks is implemented using RBM, whose learning objective is to maximize the likelihood.

The activation function of the hidden layer adopts the sigmoid function, and the activation function of the output layer adopts the softmax function.

The sample data needs to be normalized before entering the model. Since the units and value ranges of different sample data are not uniform, it is necessary to normalize the sample data.

In this preferred embodiment, the normalization method adopted is a Gaussian normalization algorithm with 0 mean and 1 variance; the value range of the normalized sample data is limited to a small range, and when training the model The normalized radar data is used as the input sample data, and the normalized observation data of the automatic weather station is used as the output sample data.

Step 3: Obtain the wind speed information output by the wind speed prediction model, as an input element of the pre-trained wind disaster early warning model, and input it into the wind disaster early warning model; wherein the wind disaster early warning model combines the data of the automatic meteorological observation station in the same period to fit the wind speed information Corrected, the wind disaster early warning model can output the revised wind speed information after excluding the errors caused by the micro-topographic information;

Preferably,

In step 3, the training method of the wind disaster early warning model includes:

Step 3.1, collect the data of automatic meteorological observation stations in the same period in the wind speed prediction area and the micro-topographic information of the wind speed prediction area;

Step 3.2, take the difference between the wind speed information and the data of the automatic meteorological observation station in the same period as the objective function;

Step 3.3, constructing a micro-terrain correction coefficient according to the micro-terrain information;

Step 3.4, take the optimal value of the objective function as the training target, and adjust the micro-terrain correction coefficient;

In step 3.5, the final usable wind disaster warning model is constructed by using the adjusted micro-terrain correction coefficient pair.

The micro-terrain correction coefficient satisfies the following relationship:

S=1+K ₁ K ₂ K ₃

In the formula,

K ₁ represents the terrain category parameter, which is 2.2H/L and 1.4H/L for mountain peaks and slopes respectively, where H is the height of the mountain, L is the width of the windward side of the mountain,

K ₂ represents the horizontal coordinate parameter, which satisfies K ₂ =1-|x|/L, where x is the horizontal coordinate from the top of the mountain,

K ₃ represents the height coordinate parameter, which satisfies K ₃ =1-z/2.5H, where z is the height from the mountain surface,

When z>2.5H, let S=1;

When H/L>0.3, take H/L=0.3.

In this preferred embodiment, the micro-terrain correction coefficient is constructed according to the micro-terrain information to obtain the corrected wind speed information. In the actual application scenario, the power grid operating conditions, line status information, etc. will all cause errors in the actual wind speed, and the parameter correction method needs to be used to eliminate them. the above errors. It is worth noting that the inventive concept of modifying the wind speed information by those skilled in the art using other information to construct a correction coefficient also falls within the protection scope of the present invention.

Step 4, obtain wind speed revision information, and give early warning of high wind disasters.

In this preferred embodiment, wind disasters are divided into four levels, see Table 1 for details.

Table 1 Early warning level of gale disaster

Warning wind speed threshold Warning level V≤13m/s Class I 13m/s<V≤20m/s Class II 20m/s<V≤29m/s Class III V>29m/s Level IV

As shown in Figure 3, a power grid wind disaster early warning device based on deep learning includes: a collection module, a wind speed prediction module, a wind disaster early warning module and an output module.

The acquisition module is used to collect radar data as input data; the radar data includes the maximum echo intensity obtained from the current body sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective monomer, and the time-varying echo intensity. , radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to the automatic station, mesocyclone, reflectivity, latitude and longitude;

The wind speed prediction module is used to input the input data into the pre-trained wind speed prediction model; the wind speed prediction model is obtained after training the sample data by using the deep learning algorithm, and the wind speed prediction model can output the wind speed information according to the input sample data. ;

The wind disaster early warning module is used to obtain the wind speed information output by the wind speed prediction model, which is used as the input element of the pre-trained wind disaster early warning model and input into the wind disaster early warning model. By fitting and correcting, the wind disaster warning model can output the wind speed revision information after excluding the errors caused by the micro-topographic information;

The output module is used to obtain wind speed revision information and send high wind disaster graded warnings to users.

The beneficial effect of the present invention is that, compared with the prior art, the present invention realizes the graded early warning and forecast of gale disasters, provides graded gale position early warning in real time, and solves the problem of refined forecasting of wind damage in the power grid. ×1km spatial resolution, forecast results in 0-120 minutes; form a business operation software system, further improve the early warning and forecast accuracy of gale disasters, improve the pertinence of power grid and equipment disaster prevention and mitigation, and emergency rescue efficiency, and improve the level of safe operation of the power grid and Social electricity reliability.

The applicant of the present invention has described and described the embodiments of the present invention in detail with reference to the accompanying drawings, but those skilled in the art should understand that the above embodiments are only preferred embodiments of the present invention, and the detailed description is only to help readers better It should be understood that the spirit of the present invention is not limited to the protection scope of the present invention. On the contrary, any improvement or modification made based on the spirit of the present invention should fall within the protection scope of the present invention.

Claims (10)

1. a power grid gale disaster early warning method based on deep learning, is characterized in that, The steps of the method are as follows: Step 1, collect radar data to form input data; Step 2, input the input data into the pre-trained wind speed prediction model; wherein the wind speed prediction model is obtained by using a deep learning algorithm to train the sample data, and the wind speed prediction model can output the wind speed according to the input sample data. information; Step 3: Obtain the wind speed information output by the wind speed prediction model, as an input element of the pre-trained wind disaster early warning model, and input it into the wind disaster early warning model; wherein the wind disaster early warning model combines the data of the automatic meteorological observation station in the same period to carry out the wind speed information. Fitting correction, the wind disaster early warning model can output wind speed revision information after excluding errors caused by micro-topographic information; Step 4, obtain wind speed revision information, and give early warning of high wind disasters. 2. a kind of power grid gale disaster early warning method based on deep learning according to claim 1, is characterized in that, In step 1, the radar data includes the maximum echo intensity obtained from the current body sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective cell, the time-varying echo intensity, the radial velocity, and the geometric center position. , cloud water content, cloud shape, wind shear, near-ground humidity, wind speed at the corresponding automatic station, mesocyclone, reflectivity, latitude and longitude. 3. a kind of power grid gale disaster early warning method based on deep learning according to claim 2, is characterized in that, In step 2, the training method of the wind speed prediction model includes: Step 2.1, collect historical radar data as input sample data for the wind speed prediction model, the radar historical data includes the maximum echo intensity obtained from the historical volume sweep echo, the height corresponding to the maximum echo intensity, and the echo top height of the convective cell , time-varying echo intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude; Step 2.2, obtaining wind speed information for characterizing wind speed and wind direction according to the input sample data, and using the wind speed information as the output sample data of the wind speed prediction model; Step 2.3, combine the input sample data and the output sample data into sample data, the sample data includes the maximum echo intensity obtained from the historical volume sweep echo, the height corresponding to the maximum echo intensity, the echo top height of the convective monomer, Time-variation of echo intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude, and historical volume sweepback The wind speed and direction corresponding to the wave; Step 2.4, combine multiple sample data into a sample data set, use all the input sample data in the sample data set as the input data of the wind speed prediction model, and use all the output sample data in the sample data set as the output data of the wind speed prediction model, from the bottom layer. To the top layer, unsupervised learning is carried out layer by layer, and the weight parameters of each layer are trained; Step 2.5, based on the weight parameters of each layer obtained by training, apply the wake-sleep algorithm, carry out supervised learning layer by layer from the top layer to the bottom layer, and adjust the weight parameters of each layer; Step 2.6, input all the input sample data in the sample data set into the trained and adjusted wind speed prediction model, and obtain the test data output by the trained and adjusted wind speed prediction model; when the difference between the test data and the output sample data is If the error meets the preset requirements, the trained and adjusted model is used as the final available wind speed prediction model; when the error between the test data and the output sample data cannot meet the preset requirements, return to step 2.4 to train the model and adjustment. 4. a kind of grid gale disaster early warning method based on deep learning according to claim 3, is characterized in that, The wind speed prediction model is a deep learning neural network model, including an input layer, three hidden layers, and an output layer; wherein, The input layer includes 32 network nodes, the output layer includes 3 network nodes, and the number of network nodes in the hidden layer from the input layer to the output layer is 16, 8 and 4 in sequence; The learning rate is set to 0.001. 5. A kind of grid gale disaster early warning method based on deep learning according to claim 4, is characterized in that, The hidden layer adopts a network structure composed of three individually trained restricted Boltzmann machines stacked. 6. A kind of grid gale disaster early warning method based on deep learning according to claim 4, is characterized in that, The activation function of the hidden layer adopts the sigmoid function, and the activation function of the output layer adopts the softmax function. 7. A kind of grid gale disaster early warning method based on deep learning according to claim 1 or 3, is characterized in that, The sample data needs to be normalized before being input into the model. 8. The method for early warning of power grid gale disaster based on deep learning according to claim 1, wherein, In step 3, the training method of the wind disaster early warning model includes: Step 3.1, collect the data of automatic meteorological observation stations in the same period in the wind speed prediction area and the micro-topographic information of the wind speed prediction area; Step 3.2, take the difference between the wind speed information and the data of the automatic meteorological observation station in the same period as the objective function; Step 3.3, constructing a micro-terrain correction coefficient according to the micro-terrain information; Step 3.4, take the optimal value of the objective function as the training target, and adjust the micro-terrain correction coefficient; In step 3.5, the final usable wind disaster warning model is constructed by using the adjusted micro-terrain correction coefficient pair. 9 . The method for early warning of power grid gale disaster based on deep learning according to claim 8 , wherein, The micro-topography correction coefficient satisfies the following relationship: S=1+K ₁ K ₂ K ₃ In the formula, K ₁ represents the terrain category parameter, which is 2.2H/L and 1.4H/L for mountain peaks and slopes respectively, where H is the height of the mountain, L is the width of the windward side of the mountain, K ₂ represents the horizontal coordinate parameter, which satisfies K ₂ =1-|x|/L, where x is the horizontal coordinate from the top of the mountain, K ₃ represents the height coordinate parameter, which satisfies K ₃ =1-z/2.5H, where z is the height from the mountain surface, When z>2.5H, let S=1; When H/L>0.3, take H/L=0.3. 10. A power grid gale disaster early warning device based on deep learning, characterized in that, The early warning device includes: a collection module, a wind speed prediction module, a wind disaster early warning module and an output module; The acquisition module is used to collect radar data as input data; the radar data includes the maximum echo intensity obtained from the current body scan echo, the height corresponding to the maximum echo intensity, the echo top height of the convective monomer, the echo Time variation of intensity, radial velocity, geometric center position, cloud water content, cloud shape, wind shear, near-ground humidity, wind speed corresponding to automatic stations, mesocyclone, reflectivity, latitude and longitude; The wind speed prediction module is used to input the input data into the pre-trained wind speed prediction model; wherein the wind speed prediction model is obtained after training the sample data by using a deep learning algorithm, and the wind speed prediction model can be based on the input data. The sample data output wind speed information; The wind disaster early warning module is used to obtain the wind speed information output by the wind speed prediction model, and input it into the wind disaster early warning model as an input element of the pre-trained wind disaster early warning model; wherein the wind disaster early warning model is combined with the data of the automatic meteorological observation station in the same period. , the wind speed information is fitted and corrected, and the wind disaster early warning model can output the wind speed revision information after excluding the error caused by the micro-topographic information; The output module is used for obtaining wind speed revision information, and sending high wind disaster graded warning to users.

Images