CN114444767A - A regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model - Google Patents

Abstract

The invention relates to a regional wind speed prediction method based on a data fusion and convolution length and time sequence analysis network model, which comprises the steps of firstly processing NCEP wind speed grid data and wind speed data of a meteorological station observation point by using bilinear interpolation and inverse distance weight interpolation, then performing data fusion on the two data by using a data fusion model based on Gaussian process regression to obtain high-resolution and high-accuracy wind speed data, and finally putting the wind speed data into a wind speed prediction model to complete the prediction of wind speed; compared with the prior art, the whole scheme design has stronger noise immunity and accuracy, improves the accuracy of the wind field wind speed prediction of the target area, and can realize the long-time effective prediction of the windy weather.

Description

A regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model

technical field

The invention relates to a regional wind speed prediction method based on data fusion and convolution long and short time series analysis network models, and belongs to the technical field of wind speed prediction.

Background technique

Strong winds will have a great impact on citrus production. When the wind speed reaches a certain level, the production of citrus will be reduced. In order to enable fruit farmers to take measures to deal with strong winds in advance, it is necessary to predict strong winds.

Existing wind speed forecasts often involve NCEP and meteorological station data. NCEP is meteorological data jointly launched by the US Center for Environmental Prediction and the US Center for Atmospheric Research. Its data is freely available to the world, including wind speed, humidity and other data. Factors such as timely and free data update are widely used in meteorological forecasting.

However, in the existing wind speed prediction methods, there is no complete and efficient system measurement method, and in the actual wind speed prediction process, the accuracy of the prediction results is not high.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model, which adopts a new model structure and control strategy, and can efficiently and accurately realize wind speed prediction.

In order to solve the above technical problems, the present invention adopts the following technical solutions: the present invention designs a regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model, for realizing the wind speed prediction of the target area, including the following steps:

Step A. Based on the current collection duration range of the preset number of each time from the current time to the historical time direction, obtain the NCEP wind speed grid data corresponding to each time in the current collection duration range in the target area, and obtain each weather station in the target area. The observation points correspond to the measured wind speed data at each time in the current collection time range, and then enter step B;

Step B. According to the target area corresponding to the NCEP wind speed grid data of each time in the current collection time range, in combination with the high grid resolution preset greater than the grid resolution in the NCEP wind speed grid data, obtain the target area high grid Under the resolution, each grid corresponds to the first wind speed data at each time in the current collection time range;

At the same time, according to the measured wind speed data of each meteorological station in the target area corresponding to each time in the current collection time range, the second wind speed of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area is obtained. data, and then enter step C;

Step C. According to the first wind speed data and the second wind speed data of each grid corresponding to each time in the current collection duration range according to the high grid resolution of the target area, and the longitude and latitude information of each grid under the high grid resolution of the target area , height information, slope information, slope aspect information, the application has been trained to input the first wind speed data, second wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the grid, and the grid corresponds to The fused wind speed data is the output data fusion model, and the fused wind speed data of each grid corresponding to each time in the current collection duration range under the high grid resolution of the target area are obtained, and then enter step D;

Step D. According to the fusion wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area, the application has trained the fusion wind speed data of each time in the current collection time range corresponding to the grid as: Input, the grid corresponds to the wind speed prediction data of the next time in the future relative to the current collection time range is the output wind speed prediction model, and each grid under the high grid resolution of the target area corresponds to the next time in the future relative to the current collection time range. The wind speed prediction data, that is, the wind speed prediction of the target area corresponding to the next time in the future relative to the current collection time range is realized.

As a preferred technical solution of the present invention: based on the fused wind speed data at each time in the current collection time range, and the predicted wind speed data at the next time in the future relative to the current collection time range, perform step D to roll the prediction K times, that is, to achieve the goal The regions correspond to wind speed predictions for K future times from the next time of the current time, respectively.

As a preferred technical solution of the present invention: in the step B, according to the target area, the NCEP wind speed grid data corresponding to each time in the current collection duration range respectively, combined with the preset grid resolution greater than that in the NCEP wind speed grid data The high grid resolution of the target area is applied, and the bilinear interpolation method is applied to obtain the first wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area;

At the same time, according to the measured wind speed data of each meteorological station in the target area corresponding to each time in the current collection time range, the inverse distance weight interpolation method is used to obtain the current collection time range corresponding to each grid under the high grid resolution of the target area. The second wind speed data at each time in .

As a preferred technical solution of the present invention: in the step B, for each grid under the high grid resolution of the target area, according to the observation points of n meteorological stations closest to the grid, corresponding to the current collection time range For the measured wind speed data at each time, the inverse distance weighted interpolation method is used to obtain the interpolated wind speed data corresponding to each time in the current collection time range, as the grid respectively corresponding to the second wind speed data of each time in the current collection time range; and then obtain Under the high grid resolution of the target area, each grid corresponds to the second wind speed data at each time in the current collection time range; where 1<n≤N, N represents the number of observation points of meteorological stations in the target area.

As a preferred technical solution of the present invention: the data fusion model in the step C is obtained by training according to the following steps C1 to C4;

Step C1. Obtain the NCEP wind speed grid data corresponding to each preset historical time in the target area, and obtain the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, and then enter step C2;

Step C2. Perform bilinear interpolation according to the NCEP wind speed grid data corresponding to each preset historical time in the target area, and obtain the first interpolated wind speed data corresponding to each preset historical time at each meteorological station observation point in the target area;

At the same time, according to the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, the inverse distance weight interpolation method is applied to obtain the second interpolation value of each meteorological station observation point in the target area corresponding to each preset historical time. wind speed data;

Then enter step C3;

Step C3. According to the actual measured wind speed data, the first interpolated wind speed data, the second interpolated wind speed data of each weather station observation point corresponding to preset historical times in the target area, and the latitude and longitude information and elevation information corresponding to the observation points of each weather station respectively , slope information, slope aspect information, take the first interpolated wind speed data, second interpolated wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation point of the meteorological station as the input, and the corresponding observation point of the meteorological station The measured wind speed data is the output, and the data fusion model to be trained based on Gaussian process regression is trained to obtain the data fusion model, and then enter step C4;

Step C4. For the obtained data fusion model, define the first interpolation wind speed data, second interpolation wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation points of the meteorological station in each input, corresponding to The first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid in the target area, as well as the measured wind speed data corresponding to the observation points of the meteorological station in the definition output, corresponding to the target The fused wind speed data corresponding to the grid in the area, that is, the updated and obtained first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid are input. The fusion wind speed data is the output data fusion model.

As a preferred technical solution of the present invention: in the step C2, for each meteorological station observation point in the target area, respectively, according to the n meteorological station observation points closest to the meteorological station observation point corresponding to the preset historical time For the measured wind speed data, the inverse distance weighted interpolation method is applied to obtain the interpolated wind speed data corresponding to each preset historical time, as the second interpolated wind speed data corresponding to each preset historical time for the observation point of the meteorological station; The observation points of the meteorological station correspond to the second interpolated wind speed data of each preset historical time; wherein, 1<n≤N, N represents the number of observation points of the meteorological station in the target area.

As a preferred technical solution of the present invention: in the step C3, based on the Gaussian process regression-based data fusion model to be trained constructed by the constant kernel function and the Gaussian kernel function, combined with the constant kernel function parameter constant_value=1.0, the Gaussian kernel function The initialization of the parameter length_scale=1.0, by updating the constant kernel function parameter constant_value and the Gaussian kernel function parameter length_scale in the training process, trains the data fusion model to be trained, and obtains the data fusion model.

As a preferred technical solution of the present invention: in the step D, according to the high grid resolution of the target area, each grid corresponds to the radial wind data and the latitudinal wind data in the fused wind speed data at each time in the current collection duration range. Wind data, the application has been trained to take the radial wind data at each time in the current collection time range corresponding to the grid as input, and the grid corresponding to the radial wind forecast data at the next time in the future relative to the current collection time range as the output diameter. The wind speed prediction model is used to obtain the radial wind speed prediction data of the next time in the future relative to the current collection time range corresponding to each grid under the high grid resolution of the target area;

At the same time, apply the zonal wind speed prediction model that has been trained using the zonal wind data at each time in the current collection time range corresponding to the grid as input, and the zonal wind prediction data corresponding to the grid at the next time in the future as the output to obtain the target. Each grid under high regional grid resolution corresponds to the zonal wind speed prediction data in the next time relative to the current collection time range, which constitutes the target area under high grid resolution. Each grid corresponds to the current collection time range. The wind speed prediction data of the next time in the future can be used to realize the wind speed prediction of the target area in the next time in the future.

As a preferred technical solution of the present invention: the radial wind speed prediction model and the zonal wind speed prediction model in the step D are obtained by training according to the following steps D1 to D3;

Step D1. Convolution time series analysis network model based on attention mechanism, combined with the preset size of each convolution kernel and the preset optimizer of the model, build a wind speed prediction model to be trained, and then enter step D2;

Step D2. Based on the NCEP wind speed grid data of each time in the target area corresponding to the preset historical time range and the measured wind speed data of each time in the preset historical time range corresponding to each meteorological station observation point in the target area, press steps A to C method, obtain the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area, and then enter step D3;

Step D3. Take the corresponding direction wind speed data in the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area as the input, and take each grid under the high grid resolution of the target area as the input. The corresponding direction wind speed data in the fusion wind speed data of the next time relative to the preset historical duration range is output, and combined with the preset loss function, the wind speed prediction model to be trained is trained to obtain the corresponding direction wind speed prediction model.

As a preferred technical solution of the present invention: the preset loss function in the step D3 is as follows:

表示目标区域高网格分辨率下网格i的对应向预测风速数据，y_i表示目标区域高网格分辨率下网格i的对应向实测风速数据；在针对待训练风速预测模型进行训练的过程中，各次迭代中，判断所获

是否大于预设阈值，是则更新待训练风速预测模型中的卷积核数目、以及参数λ和参数σ，并进入下次迭代；否则保留待训练风速预测模型中的卷积核数目、以及参数λ和参数σ，即获得对应向风速预测模型。in,

Represents the corresponding predicted wind speed data of grid i in the high grid resolution of the target area, and y _i represents the corresponding measured wind speed data of grid i in the high grid resolution of the target area; In the process, in each iteration, the judgment obtained

Whether it is greater than the preset threshold, if yes, update the number of convolution kernels, parameters λ and σ in the wind speed prediction model to be trained, and enter the next iteration; otherwise, keep the number of convolution kernels and parameters in the wind speed prediction model to be trained λ and parameter σ, that is, to obtain the corresponding wind speed prediction model.

The method for forecasting regional wind speed based on data fusion and convolution long and short time series analysis network model according to the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

(1) The regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed by the present invention is based on the NCEP wind speed grid data and the wind speed data of each meteorological station observation point in the target area, compared with the traditional wind speed prediction data The processing is different. Inverse distance weight interpolation is performed on the wind speed data of the observation points of the meteorological station, bilinear interpolation is performed on the NCEP wind speed grid data, and the data fusion model based on Gaussian process regression is used to fuse the two data to improve the wind speed. Field data resolution and accuracy;

(2) The regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed by the present invention, by introducing the convolution long and short time series analysis neural network, and using the wind speed prediction model based on the convolution long and short time series analysis network, it is realized to The long-term effective prediction of wind fields with medium and high wind speeds in the target area has stronger noise immunity than traditional short-term wind speed prediction methods;

(3) The regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed by the present invention improves the wind speed prediction based on convolution long and short time series analysis network by introducing an improved attention mechanism loss function based on MSE loss function. The model's attention to the high wind speed wind field improves the accuracy of the high wind speed wind field prediction without losing the prediction accuracy of the model.

Description of drawings

Fig. 1 is the flow chart of the regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed by the present invention;

Fig. 2 is the NCEP wind speed grid data of the target area in the application of the present invention and the schematic diagram of the wind speed data of the observation point of the meteorological station;

Fig. 3 is the application flow schematic diagram of step A to step C of the present invention;

Figure 4a is a schematic diagram of the wind speed before the interpolation of each meteorological station observation point in the target area in step B of the present invention;

4b is a schematic diagram of the wind speed after interpolation of observation points of each meteorological station in the target area in step B of the present invention;

Fig. 5a is the schematic diagram of NCEP wind speed grid data in the target area in step B of the present invention;

Fig. 5b is the schematic diagram after interpolation of NCEP wind speed grid data in the target area in step B of the present invention;

Fig. 6 is the training flow chart of the data fusion model based on Gaussian process regression in step C of the present invention;

Fig. 7 is the wind speed prediction model training flow chart of analyzing the neural network based on convolution length and short sequence analysis in step D of the present invention;

FIG. 8 is a comparison between the wind field predicted in the next 1, 5 and 10 days and the real wind field in the implementation and application of the present invention.

Detailed ways

The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

The present invention designs a regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model, which is used to realize the wind speed prediction of the target area. In practical application, as shown in Figure 1, the specific steps are as follows.

Step A. Based on the current collection duration range of the preset number of each time from the current time to the historical time direction, obtain the NCEP wind speed grid data corresponding to each time in the current collection duration range in the target area, and obtain each weather station in the target area. The observation points correspond to the measured wind speed data at each time in the current collection time range, as shown in Figure 3, and then go to step B. As shown in Figure 2, it shows the visualization of the NCEP wind speed grid data involved in the present invention and the wind speed data of the observation point of the meteorological station in the target area, wherein the black triangle represents the corresponding position of the meteorological station in the target area.

Step B. As shown in Figure 3, utilize the NetCDF4 module in python to read the data obtained in step A, respectively correspond to the NCEP wind speed grid data of each time in the current collection duration range according to the target area, in conjunction with the preset greater than this NCEP wind speed grid The high grid resolution of the grid resolution in the data, the bilinear interpolation method is applied to obtain the first wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area, as shown in Figure 5a as shown in Figure 5b.

Bilinear interpolation, also known as bilinear interpolation, in mathematics, bilinear interpolation is a linear interpolation extension of an interpolation function with two variables. The core idea is to perform linear interpolation in two directions respectively.

At the same time, according to the measured wind speed data of each meteorological station in the target area corresponding to each time in the current collection time range, the inverse distance weight interpolation method is applied to obtain the corresponding current collection of each grid under the high grid resolution of the target area. The second wind speed data at each time in the duration range is shown in FIG. 4 a and FIG. 4 b , and then step C is entered.

Inverse distance weighted interpolation method (Inverse Distance Weighted, IDW) uses the distance between the interpolation point and the sample point as the weight to perform a weighted average, and the sample point that is closer to the interpolation point is given a greater weight.

Among them, regarding the acquisition of the second wind speed data corresponding to each grid at each time in the current collection time range under the high grid resolution of the target area, for each grid under the high grid resolution of the target area, according to the distance between the grids The nearest n meteorological station observation points correspond to the measured wind speed data at each time in the current collection time range, and the inverse distance weight interpolation method is used to obtain the interpolated wind speed data corresponding to each time in the current collection time range, as the grid corresponding to each time. The second wind speed data at each time in the current collection time range; and then obtain the second wind speed data corresponding to each grid at each time in the current collection time range under the high grid resolution of the target area; where, 1<n≤N, N Indicates the number of weather station observation points in the target area.

Step C. As shown in Figure 3, using the numpy module in python, according to the high grid resolution of the target area, each grid corresponds to the first wind speed data, the second wind speed data at each time in the current collection duration range, and the target area. The longitude and latitude information, height information, slope information, and slope aspect information of each grid under high grid resolution, and the first wind speed data, second wind speed data, longitude and latitude information, elevation information, and slope corresponding to the grid have been trained. Information and aspect information are input, and the fused wind speed data corresponding to the grid is the output data fusion model, and the fused wind speed data corresponding to each grid at each time in the current collection time range under the high grid resolution of the target area are obtained, and then Go to step D.

In practical applications, the data fusion model in the above step C, as shown in FIG. 6 , is specifically obtained by training according to the following steps C1 to C4.

Step C1. Obtain the NCEP wind speed grid data corresponding to each preset historical time in the target area, and obtain the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, and then proceed to step C2.

Step C2. Perform bilinear interpolation according to the NCEP wind speed grid data corresponding to each preset historical time in the target area to obtain first interpolated wind speed data corresponding to each preset historical time at each meteorological station observation point in the target area.

At the same time, according to the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, the inverse distance weight interpolation method is applied to obtain the second interpolation value of each meteorological station observation point in the target area corresponding to each preset historical time. Wind speed data; then go to step C3.

In the application, regarding the acquisition of the second interpolated wind speed data corresponding to each preset historical time for each observation point of the meteorological station in the target area, the specific design is for each observation point of the meteorological station in the target area, according to the nearest observation point of the meteorological station. The observation points of n meteorological stations correspond to the measured wind speed data of each preset historical time respectively, and the inverse distance weight interpolation method is applied to obtain the interpolated wind speed data corresponding to each preset historical time respectively, as the observation points of the meteorological station corresponding to each preset historical time respectively. The second interpolated wind speed data of time; and then obtain the second interpolated wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time; wherein, 1<n≤N, N represents the meteorological station observation point in the target area. quantity.

In practical applications, for example, the measured wind speed data at the observation point of each meteorological station is not used for interpolation at the observation point of the meteorological station, and only the three points within the interpolation radius or the nearest points are used as the weights of the interpolation, and the weights are determined by the distance. Dynamic radius: radius R>=100, weight point P>=3, that is, the interpolation radius is 100 kilometers. If the range of 100 kilometers does not meet 3 points, the interpolation radius is expanded until the three points within the range are satisfied.

Step C3. According to the actual measured wind speed data, the first interpolated wind speed data, the second interpolated wind speed data of each weather station observation point corresponding to preset historical times in the target area, and the latitude and longitude information and elevation information corresponding to the observation points of each weather station respectively , slope information, slope aspect information, take the first interpolated wind speed data, second interpolated wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation point of the meteorological station as the input, and the corresponding observation point of the meteorological station The measured wind speed data is the output, and the data fusion model to be trained based on the Gaussian process regression is trained to obtain the data fusion model, and then step C4 is entered.

Gaussian Process Regression (GPR) is a nonparametric model that uses Gaussian Process (GP) priors to perform regression analysis on data. GPR is widely used in the field of machine learning and has a strict theoretical basis for statistical learning. Based on the Bayesian linear regression method, it replaces the linear kernel of the Bayesian linear regression with a kernel function, so the Gaussian process regression has a good adaptability to dealing with complex problems such as high dimensionality, small samples, and nonlinearity. It has the characteristics of strong generalization ability.

Therefore, in practical applications, based on the Gaussian process regression-based data fusion model to be trained constructed by the constant kernel function and the Gaussian kernel function, combined with the initialization of the constant kernel function parameter constant_value=1.0 and the Gaussian kernel function parameter length_scale=1.0, through the training process In the update of the constant kernel function parameter constant_value and the Gaussian kernel function parameter length_scale, the data fusion model to be trained is trained to obtain the data fusion model.

In the above training process of the data fusion model, it is judged whether the RMSE of the training result is greater than the preset threshold. If so, update the constant kernel function parameter constant_value and the Gaussian kernel function parameter length_scale respectively, and return to continue training; otherwise, obtain the current constant kernel function parameters constant_value and Gaussian kernel function parameter length_scale, output the trained data fusion model based on Gaussian process regression.

Step C4. For the obtained data fusion model, define the first interpolation wind speed data, second interpolation wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation points of the meteorological station in each input, corresponding to The first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid in the target area, as well as the measured wind speed data corresponding to the observation points of the meteorological station in the definition output, corresponding to the target The fused wind speed data corresponding to the grid in the area, that is, the updated and obtained first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid are input. The fusion wind speed data is the output data fusion model.

Step D. According to the fusion wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area, the application has trained the fusion wind speed data of each time in the current collection time range corresponding to the grid as: Input, the grid corresponds to the wind speed prediction data of the next time in the future relative to the current collection time range is the output wind speed prediction model, and each grid under the high grid resolution of the target area corresponds to the next time in the future relative to the current collection time range. The wind speed prediction data, that is, the wind speed prediction of the target area corresponding to the next time in the future relative to the current collection time range is realized.

Based on the division of wind in relation to radial wind and zonal wind, in the specific practical application of the above step D, according to the high grid resolution of the target area, each grid corresponds to the fused wind speed data of each time in the current collection time range. Radial wind data and zonal wind data, the application has been trained to use the radial wind data at each time in the current collection time range corresponding to the grid as input, and the grid corresponds to the radial wind data at the next time in the future relative to the current collection time range. The wind prediction data is the output radial wind speed prediction model, and the radial wind speed prediction data corresponding to each grid at the next time in the future relative to the current collection time range under the high grid resolution of the target area are obtained.

At the same time, apply the zonal wind speed prediction model that has been trained using the zonal wind data at each time in the current collection time range corresponding to the grid as input, and the zonal wind prediction data corresponding to the grid at the next time in the future as the output to obtain the target. Each grid under high regional grid resolution corresponds to the zonal wind speed prediction data in the next time relative to the current collection time range, which constitutes the target area under high grid resolution. Each grid corresponds to the current collection time range. The wind speed prediction data of the next time in the future can be used to realize the wind speed prediction of the target area in the next time in the future.

From this further analysis, the radial wind speed prediction model and the zonal wind speed prediction model in the above step D, as shown in FIG. 7 , are obtained by training according to the following steps D1 to D3.

Step D1. Convolution time series analysis network model based on attention mechanism, combined with the preset size of each convolution kernel, such as 19×19, and the preset optimizer of the model, such as the Nadam optimizer, to construct the wind speed prediction model to be trained, Then go to step D2.

And in practical applications, the hidden layers of the radial wind speed prediction model are set to three layers, and the hidden layers of the zonal wind speed prediction model are set to two layers, and each hidden layer of the radial wind speed prediction model and the zonal wind speed prediction model rolls The number of cores is 100.

Step D2. Based on the NCEP wind speed grid data of each time in the target area corresponding to the preset historical time range and the measured wind speed data of each time in the preset historical time range corresponding to each meteorological station observation point in the target area, press steps A to C method to obtain the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area, and then proceed to step D3.

Step D3. Take the corresponding direction wind speed data in the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area as the input, and take each grid under the high grid resolution of the target area as the input. The corresponding direction wind speed data in the fusion wind speed data of the next time relative to the preset historical duration range is output, and combined with the preset loss function, the wind speed prediction model to be trained is trained to obtain the corresponding direction wind speed prediction model.

And in practical applications, the preset loss function in the above step D3 is as follows:

表示目标区域高网格分辨率下网格i的对应向预测风速数据，y_i表示目标区域高网格分辨率下网格i的对应向实测风速数据；在针对待训练风速预测模型进行训练的过程中，各次迭代中，判断所获

是否大于预设阈值，是则更新待训练风速预测模型中的卷积核数目、以及参数λ和参数σ，并进入下次迭代；否则保留待训练风速预测模型中的卷积核数目、以及参数λ和参数σ，即获得对应向风速预测模型。并且实际应用当中，参数λ和参数σ分别基于0.1和2初始化进行训练。in,

Represents the corresponding predicted wind speed data of grid i in the high grid resolution of the target area, and y _i represents the corresponding measured wind speed data of grid i in the high grid resolution of the target area; In the process, in each iteration, the judgment obtained

Whether it is greater than the preset threshold, if yes, update the number of convolution kernels, parameters λ and σ in the wind speed prediction model to be trained, and enter the next iteration; otherwise, keep the number of convolution kernels and parameters in the wind speed prediction model to be trained λ and parameter σ, that is, to obtain the corresponding wind speed prediction model. And in practical applications, the parameter λ and the parameter σ are trained based on 0.1 and 2 initialization, respectively.

In practical applications, based on the fused wind speed data at each time in the current collection time range, and the predicted wind speed data at the next time in the future relative to the current collection time range, perform step D to roll the prediction K times, that is, to achieve the target area corresponding to the current time. The wind speed prediction for K times in the future from the next time, that is, to realize the wind speed prediction for the target area for a period of time.

Applying the regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed by the present invention to practice, about rolling prediction K times for the above step D, in the application, such as first using the predicted wind speed data of the previous two days to obtain The predicted wind speed data of the third day, then delete the predicted wind speed data of the first day in the predicted wind speed data of the previous two days, add the predicted wind speed data of the third day obtained after the prediction, and then put it into the model for prediction, Repeat the above process until the wind forecast wind speed data on the twelfth day is predicted; the overall prediction effect of the model is shown in Figure 8, showing the data fusion model based on Gaussian process regression and the convolution length based on the attention mechanism loss function respectively The time series analysis neural network wind speed prediction model predicts the comparison effect of the target area wind field and the real wind field in the next 1, 5 and 10 days.

To sum up, the regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model designed in the present invention is based on the NCEP wind speed grid data and the wind speed data of observation points of each meteorological station in the target area, compared with the traditional wind speed. The prediction data processing is different. Inverse distance weight interpolation is performed on the wind speed data of the observation points of the meteorological station, bilinear interpolation is performed on the NCEP wind speed grid data, and the data fusion model based on Gaussian process regression is used to fuse the two data. Improve the resolution and accuracy of wind field data; and by introducing a convolutional long and short time series analysis neural network, and using the wind speed prediction model based on the convolution long and short time series analysis network, the long-term effective prediction of the medium and high wind speed wind field in the target area is realized. Compared with the traditional short-term wind speed prediction method, it has stronger anti-noise; and by introducing the improved attention mechanism loss function based on the MSE loss function, the wind speed prediction model based on the convolutional long and short time series analysis network is improved for high wind speed wind. Field attention improves the accuracy of wind field predictions for high wind speeds without losing model prediction accuracy.

The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned embodiments, and can also be made within the scope of knowledge possessed by those of ordinary skill in the art without departing from the purpose of the present invention. Various changes.

Claims (10)

1. a regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model, for realizing the wind speed prediction to target area, it is characterized in that, comprise the steps: Step A. Based on the current collection duration range of the preset number of each time from the current time to the historical time direction, obtain the NCEP wind speed grid data corresponding to each time in the current collection duration range in the target area, and obtain each weather station in the target area. The observation points correspond to the measured wind speed data at each time in the current collection time range, and then enter step B; Step B. According to the target area corresponding to the NCEP wind speed grid data of each time in the current collection time range, in combination with the high grid resolution preset greater than the grid resolution in the NCEP wind speed grid data, obtain the target area high grid Under the resolution, each grid corresponds to the first wind speed data at each time in the current collection time range; At the same time, according to the measured wind speed data of each meteorological station in the target area corresponding to each time in the current collection time range, the second wind speed of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area is obtained. data, and then enter step C; Step C. According to the first wind speed data and the second wind speed data of each grid corresponding to each time in the current collection duration range according to the high grid resolution of the target area, and the longitude and latitude information of each grid under the high grid resolution of the target area , height information, slope information, slope aspect information, the application has been trained to input the first wind speed data, second wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the grid, and the grid corresponds to The fused wind speed data is the output data fusion model, and the fused wind speed data of each grid corresponding to each time in the current collection duration range under the high grid resolution of the target area are obtained, and then enter step D; Step D. According to the fusion wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area, the application has trained the fusion wind speed data of each time in the current collection time range corresponding to the grid as: Input, the grid corresponds to the wind speed prediction data of the next time in the future relative to the current collection time range is the output wind speed prediction model, and each grid under the high grid resolution of the target area corresponds to the next time in the future relative to the current collection time range. The wind speed prediction data, that is, the wind speed prediction of the target area corresponding to the next time in the future relative to the current collection time range is realized. 2. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 1, it is characterized in that: based on the fusion wind speed data of each time in the current collection duration range, and relative to the current collection duration range The predicted wind speed data at the next time in the future, perform step D rolling prediction K times, that is, to achieve wind speed prediction corresponding to the target area at the next K times in the future from the current time. 3. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 1, is characterized in that: in described step B, according to the target area corresponding to each time in the current collection duration range respectively. The NCEP wind speed grid data, combined with the preset high grid resolution larger than the grid resolution in the NCEP wind speed grid data, applies the bilinear interpolation method to obtain the target area under the high grid resolution. Collect the first wind speed data at each time in the duration range; At the same time, according to the measured wind speed data of each meteorological station in the target area corresponding to each time in the current collection time range, the inverse distance weight interpolation method is used to obtain the current collection time range corresponding to each grid under the high grid resolution of the target area. The second wind speed data at each time in . 4. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 3, is characterized in that: in described step B, respectively for each network under the high grid resolution of target area. According to the measured wind speed data at each time in the current collection time range corresponding to the n meteorological station observation points closest to the grid, the inverse distance weight interpolation method is applied to obtain the interpolated wind speed data corresponding to each time in the current collection time range, As the grid respectively corresponding to the second wind speed data of each time in the current collection time range; and then obtain the second wind speed data of each grid corresponding to each time in the current collection time range under the high grid resolution of the target area; wherein, 1 <n≤N, N represents the number of observation points of weather stations in the target area. 5. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 1, is characterized in that: the data fusion model in described step C, is trained according to following steps C1 to step C4 get; Step C1. Obtain the NCEP wind speed grid data corresponding to each preset historical time in the target area, and obtain the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, and then enter step C2; Step C2. Perform bilinear interpolation according to the NCEP wind speed grid data corresponding to each preset historical time in the target area, and obtain the first interpolated wind speed data corresponding to each preset historical time at each meteorological station observation point in the target area; At the same time, according to the measured wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time, the inverse distance weight interpolation method is applied to obtain the second interpolation value of each meteorological station observation point in the target area corresponding to each preset historical time. wind speed data; Then enter step C3; Step C3. According to the actual measured wind speed data, the first interpolated wind speed data, the second interpolated wind speed data of each weather station observation point corresponding to preset historical times in the target area, and the latitude and longitude information and elevation information corresponding to the observation points of each weather station respectively , slope information, slope aspect information, take the first interpolated wind speed data, second interpolated wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation point of the meteorological station as the input, and the corresponding observation point of the meteorological station The measured wind speed data is the output, and the data fusion model to be trained based on Gaussian process regression is trained to obtain the data fusion model, and then enter step C4; Step C4. For the obtained data fusion model, define the first interpolation wind speed data, second interpolation wind speed data, longitude and latitude information, elevation information, slope information, and slope aspect information corresponding to the observation points of the meteorological station in each input, corresponding to The first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid in the target area, as well as the measured wind speed data corresponding to the observation points of the meteorological station in the definition output, corresponding to the target The fused wind speed data corresponding to the grid in the area, that is, the updated and obtained first wind speed data, second wind speed data, latitude and longitude information, elevation information, slope information, and slope aspect information corresponding to the grid are input. The fusion wind speed data is the output data fusion model. 6. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 5, is characterized in that: in described step C2, for each weather station observation point in the target area respectively, according to The observation points of the meteorological station closest to the observation points of the meteorological station correspond to the actual measured wind speed data of each preset historical time, and the inverse distance weight interpolation method is applied to obtain the interpolated wind speed data corresponding to the preset historical times respectively, as the observation of the meteorological station. The points correspond to the second interpolated wind speed data of each preset historical time respectively; and then obtain the second interpolated wind speed data of each meteorological station observation point in the target area corresponding to each preset historical time; wherein, 1<n≤N, N represents the target The number of weather station observations in the area. 7. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 5, is characterized in that: in described step C3, based on being constructed by constant kernel function and Gaussian kernel function based on Gaussian The data fusion model to be trained for the process regression, combined with the initialization of the constant kernel function parameter constant_value=1.0 and the Gaussian kernel function parameter length_scale=1.0, through the updating of the constant kernel function parameter constant_value and the Gaussian kernel function parameter length_scale during the training process, for the to-be-trained The data fusion model is trained to obtain a data fusion model. 8. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 1, is characterized in that: in described step D, according to each grid under the high grid resolution of target area, respectively. Corresponding to the radial wind data and zonal wind data in the fusion wind speed data at each time in the current collection time range, the application has been trained to use the radial wind data at each time in the current collection time range corresponding to the grid as input, and the grid The radial wind forecast data corresponding to the next time in the future relative to the current collection time range is the output radial wind speed prediction model, and each grid under the high grid resolution of the target area corresponds to the next time in the future relative to the current collection time range. The radial wind speed forecast data of ; At the same time, apply the zonal wind speed prediction model that has been trained using the zonal wind data at each time in the current collection time range corresponding to the grid as input, and the zonal wind prediction data corresponding to the grid at the next time in the future as the output to obtain the target. Each grid under high regional grid resolution corresponds to the zonal wind speed prediction data in the next time relative to the current collection time range, which constitutes the target area under high grid resolution. Each grid corresponds to the current collection time range. The wind speed prediction data of the next time in the future can be used to realize the wind speed prediction of the target area in the next time in the future. 9. a kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 8, is characterized in that: the radial wind speed prediction model and the zonal wind speed prediction model in the described step D, both Perform training to obtain according to the following steps D1 to D3; Step D1. Convolution time series analysis network model based on attention mechanism, combined with the preset size of each convolution kernel and the preset optimizer of the model, build a wind speed prediction model to be trained, and then enter step D2; Step D2. Based on the NCEP wind speed grid data of each time in the target area corresponding to the preset historical time range and the measured wind speed data of each time in the preset historical time range corresponding to each meteorological station observation point in the target area, press steps A to C method, obtain the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area, and then enter step D3; Step D3. Take the corresponding direction wind speed data in the fusion wind speed data of each grid corresponding to each time in the preset historical duration range under the high grid resolution of the target area as the input, and take each grid under the high grid resolution of the target area as the input. The corresponding direction wind speed data in the fusion wind speed data of the next time relative to the preset historical duration range is output, and combined with the preset loss function, the wind speed prediction model to be trained is trained to obtain the corresponding direction wind speed prediction model. 10. A kind of regional wind speed prediction method based on data fusion and convolution long and short time series analysis network model according to claim 9, is characterized in that: the preset loss function in described step D3 is as follows:

in,

Represents the corresponding predicted wind speed data of grid i in the high grid resolution of the target area, and y _i represents the corresponding measured wind speed data of grid i in the high grid resolution of the target area; In the process, in each iteration, the judgment obtained

Whether it is greater than the preset threshold, if yes, update the number of convolution kernels, parameters λ and σ in the wind speed prediction model to be trained, and enter the next iteration; otherwise, keep the number of convolution kernels and parameters in the wind speed prediction model to be trained λ and parameter σ, that is, to obtain the corresponding wind speed prediction model.

Images