Wind speed prediction method and system fused with remote correlation factors
By fusion of telecorrelation factors, the feature maps of the wind speed prediction model are extracted using multiple models, which solves the problem of unsatisfactory effects of the existing wind speed prediction model, and improves the accuracy of wind speed prediction and the adaptability of the model.
Patent Information
- Application Number
- CN202410933521.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-12-10
AI Technical Summary
The existing wind speed prediction model has not been ideal, and it is difficult to accurately predict wind speed, which increases the difficulty of wind energy development.
Using the method of fusion telecorrelation factors, the historical meteorological data and digital elevation data of the target area and the subtropical high-pressure area in the northwest Pacific were obtained, and the MSConvLSTM, SCW, 3D-CBAM-ConvLSTM and Res-FPN models were extracted. Correlate feature maps and topographic feature maps, and combine channel stitching to predict wind speed.
It significantly improves the accuracy of wind speed prediction, improves the generalization ability and robustness of the model, effectively utilizes telecorrelation factors and terrain data, and reduces prediction errors.
Smart Images
Figure CN119106753A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind speed prediction, and in particular to a wind speed prediction method and system integrating teleconnection factors. Background Art
[0002] The energy crisis has prompted humans to look for a new alternative energy source. Wind energy is favored because of its green and renewable nature, low investment and high returns. In the application of wind energy development, wind speed is a key parameter, but it is very difficult to accurately predict wind speed, which increases the difficulty of developing wind energy. Summary of the invention
[0003] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.
[0004] In view of the above existing problems, the present invention is proposed.
[0005] Therefore, the present invention provides a wind speed prediction method and system integrating teleconnection factors to solve the problem that the existing wind speed prediction model has unsatisfactory effect.
[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0007] In a first aspect, the present invention provides a method for wind speed prediction integrating teleconnection factors, comprising:
[0008] Acquire historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, preprocess the data, and standardize the preprocessed data, wherein the preprocessing includes calculating the historical meteorological data of the target area to obtain daily average meteorological data, calculating the pentad mean of the subtropical high pressure data of the northwest Pacific Ocean through the historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, and resampling the digital elevation data to obtain altitude data and aspect data at the same resolution;
[0009] Based on the standardized pentad value of the Northwest Pacific subtropical high pressure data, the correlation feature map between the Northwest Pacific subtropical high pressure and the target area is obtained through the pre-trained MSConvLSTM model and SCW model;
[0010] Based on the standardized daily average meteorological data, the characteristic graphs of each meteorological variable are obtained through the 3D-CBAM-ConvLSTM model;
[0011] Based on the standardized elevation data and slope aspect data, the terrain feature map is obtained through the Res-FPN model;
[0012] Channel splicing is performed on the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map;
[0013] The channel spliced data is input into the prediction model to predict the average daily wind speed in the future.
[0014] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0015] The MSConvLSTM model includes the following steps:
[0016] The pentad mean of the standardized Northwest Pacific subtropical high pressure data for the previous 18 pentads was used to construct the data set as the input of the model;
[0017] The MSConvLSTM model is pre-trained using the dataset to learn and optimize model parameters;
[0018] The pre-trained MSConvLSTM model adopts a rolling prediction strategy to form a nine-phase potential height prediction sequence;
[0019] The rolling prediction strategy includes the following steps:
[0020] Using the data of the previous 18 Hou as input, the potential height G of the 19th Hou is predicted. T , G T The output is divided into two branches. The first output is to directly output the preliminary prediction results. The second output is to introduce a deep learning module consisting of two layers of 3×3 convolutional layers and two layers of upsampling layers. T Perform deep feature extraction and reconstruction to obtain the optimized prediction result G′ T ,This module is embedded in the connection layer of each layer;
[0021] The potential height G of the 19th Hou T Combined with the actual geopotential height data of the previous 17 Hou, it is used as the prediction of G at the next time point. T+1 Input
[0022] The above process is repeated until a complete prediction sequence for the next 9 phases is obtained.
[0023] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0024] The SCW model includes the following steps:
[0025] The geopotential height data of the first three phases of the subtropical high and the next eight phases, as well as the longitude and latitude data of the subtropical high area and the target area were input in parallel. The geopotential height data was used to obtain the corresponding feature map F through convolution-pooling operations;
[0026] The subtropical high latitude and longitude data are processed through two convolutional layers to obtain the feature map F1 of this stage;
[0027] The longitude and latitude data of the target area are upsampled to expand the data size, and the feature map F2 is obtained through the convolution layer;
[0028] F1 and F2 enter the shared convolutional layer to obtain their respective spatial feature weights W1 and W2, and add them together to obtain the spatial correlation feature weight W;
[0029] Multiply F and W and weight them to obtain the correlation characteristic map between the subtropical high and the wind speed in the target area.
[0030] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0031] The 3D-CBAM-ConvLSTM, wherein the 3D-CBAM is composed of a 3D channel attention module and a 3D spatiotemporal attention module, the 3D-CBAM, comprises the following steps:
[0032] The input data passes through the 3D channel attention module to obtain the feature map F C ;
[0033] Feature map F C Enter the 3D spatiotemporal attention module and obtain the feature map F CS ;
[0034] The 3D channel attention module comprises the following steps:
[0035] Use 3D maximum pooling and average pooling to aggregate the spatial information of the input data;
[0036] Perform element-based addition operations on the feature maps that have passed through MLP respectively;
[0037] The summed result is processed by the sigmoid activation function to obtain the attention weight of each channel;
[0038] The attention weight of each channel is multiplied element-by-element with the corresponding channel in the original input data, and the output feature map F is C ;
[0039] The 3D spatiotemporal attention module comprises the following steps:
[0040] Perform channel-based maximum pooling and average pooling operations on the output feature map of the channel attention module;
[0041] Merge channels of the extracted features;
[0042] Through a convolution operation with a convolution kernel of 7×7×7 and reduce the dimension to 1 channel;
[0043] Generate the spatiotemporal attention weights for each position through the sigmoid activation function;
[0044] Combine the spatiotemporal attention weights with F C Perform element-by-element multiplication to generate the spatiotemporal attention feature map F CS .
[0045] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0046] The calculation formula of the 3D channel attention module is expressed as:
[0047] M c (F)=σ(MLP(AvgPool3D(F))+MLP(MaxPool3D(F)))
[0048] Among them, MLP represents multi-layer perceptron operation, σ represents sigmoid activation function operation;
[0049] The calculation formula of the 3D spatiotemporal attention module is expressed as:
[0050] M S (F) = σ(f 7×7×7 ([AvgPool3D(F);MaxPool3D(F)]))
[0051] Among them, f 7×7×7 represents the convolution operation with a convolution kernel of 7×7×7, and σ represents the sigmoid activation function.
[0052] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0053] The ConvLSTM operation process in the 3D-CBAM-ConvLSTM model is expressed as:
[0054] I t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i )
[0055] F t =σ(Wxf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f )
[0056] O t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t-1 +b o )
[0057] C t =F t ⊙C t-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c )
[0058] H t =O t ⊙tanh(C t )
[0059] Among them, X t represents the input data at time t, I t 、F t and O t is converted into a gating coefficient between 0 and 1 through the sigmoid operation, C t is a memory unit, H t represents the hidden state of the ConvLSTM cell output at time t; W and b represent the weight and bias of the gating mechanism; ⊙, *, and σ represent Hadamard, Matmul, and Sigmoid operations, respectively.
[0060] As a preferred solution of the wind speed prediction method integrating teleconnection factors described in the present invention, wherein:
[0061] The Res-FPN model includes 4 layers of residual blocks. The number of channels of each layer of output data is expanded to twice the original number and the size is halved to obtain four feature maps of different resolutions: C1, C2, C3, and C4.
[0062] The C4 feature map is processed by the residual connection module to obtain the feature map D4;
[0063] The C4 feature map is upsampled to the same size as C3. At the same time, the C3 feature map adjusts the number of channels through 1x1 convolution. The C3 feature map with adjusted channel number is added point-to-point with the upsampled C4 feature map, and the feature map D3 is obtained through the residual connection module;
[0064] The D3 feature map is upsampled to the same size as C2. The C2 feature map also adjusts the number of channels through 1x1 convolution. The C2 feature map with adjusted channel number is added point-to-point with the upsampled D3 feature map.
[0065] The feature map D2 is obtained through the residual connection module;
[0066] The D2 feature map is upsampled to the same size as C1, and the C1 feature map adjusts the number of channels through 1x1 convolution. The C1 feature map with adjusted channel number is added point-to-point with the upsampled D2 feature map, and then the feature map D1 is obtained through the residual connection module;
[0067] All generated feature maps D1, D2, D3, and D4 are fused to obtain a terrain feature map.
[0068] In a second aspect, the present invention provides a wind speed prediction system integrating teleconnection factors, comprising:
[0069] A preprocessing module is used to obtain historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, preprocess the data, and standardize the preprocessed data. The preprocessing includes calculating the historical meteorological data of the target area to obtain daily average meteorological data, calculating the pentad mean of the subtropical high pressure data of the northwest Pacific Ocean through the historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, and resampling the digital elevation data to obtain altitude data and aspect data at the same resolution;
[0070] The subtropical high characteristic factor fusion module is used to obtain the correlation characteristic map between the subtropical high pressure in the northwest Pacific and the target area through the pre-trained MSConvLSTM model and SCW model based on the pentad value of the standardized subtropical high pressure data in the northwest Pacific;
[0071] The historical real-time feature extraction module is used to obtain the feature map of each meteorological variable through the 3D-CBAM-ConvLSTM model based on the standardized daily average meteorological data;
[0072] The terrain feature fusion module is used to obtain the terrain feature map through the Res-FPN model based on the standardized altitude data and slope aspect data;
[0073] The splicing module is used to splice the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map;
[0074] The prediction module is used to input the channel spliced data into the prediction model to predict the future daily average wind speed.
[0075] In a third aspect, the present invention provides a computing device, comprising:
[0076] Memory, used to store programs;
[0077] A processor is used to execute the computer executable instructions, which, when executed by the processor, implement the steps of the wind speed prediction method integrating teleconnection factors.
[0078] In a fourth aspect, the present invention provides a computer-readable storage medium, comprising: when the program is executed by a processor, the steps of implementing the wind speed prediction method integrating telecorrelation factors are implemented.
[0079] Beneficial effects of the present invention:
[0080] (1) In terms of data, the method of the present invention takes into account the teleconnection factor that limits the medium- and long-term wind speed forecast, namely the influence of the subtropical high, and combines terrain data with historical meteorological data. It uses multiple modules to extract temporal and spatial characteristics, and splices the characteristics of terrain, meteorological elements and subtropical high factors as the input of the prediction model. The RMSE index of the model is significantly improved.
[0081] (2) The method of the present invention adopts the channel and spatiotemporal attention mechanism, strengthens the weights of key meteorological factors, time and space, and ensures the effective use of historical meteorological data to the greatest extent;
[0082] (3) The method of the present invention improves the feature pyramid model, and the convolutional layer adopts a more stable residual connection method, which improves the generalization ability of the model and enhances the robustness and adaptability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0083] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:
[0084] Figure 1 A basic flow chart of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0085] Figure 2A Res-FPN model structure diagram of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0086] Figure 3 A schematic diagram of the structure of a ConvLSTM model of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0087] Figure 4 A schematic diagram of the structure of a 3D-CBAM model of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0088] Figure 5 A schematic diagram of the structure of a 3D channel attention model of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0089] Figure 6 A schematic diagram of the structure of a spatiotemporal attention model of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0090] Figure 7 A flowchart of a pre-trained MSConvLSTM for predicting subtropical highs in a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0091] Figure 8 An encoder-decoder structure diagram in MSConvLSTM of a wind speed prediction method integrating teleconnection factors provided by one embodiment of the present invention;
[0092] Fig. 9 A model structure diagram in a SCW module of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention;
[0093] Fig.10 A RMS graph comparing different experimental models in a test set of a wind speed prediction method integrating teleconnection factors provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0094] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0095] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0096] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.
[0097] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.
[0098] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0099] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0100] Example 1
[0101] Reference Figure 1-8 , is an embodiment of the present invention, and provides a wind speed prediction method integrating teleconnection factors, such as Figure 1 As shown, the following steps are included:
[0102] S1: Obtain historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area in the northwest Pacific, preprocess the data, and standardize the preprocessed data. The preprocessing includes calculating the historical meteorological data of the target area to obtain the daily average meteorological data, calculating the pentad mean of the subtropical high pressure data in the northwest Pacific through the historical meteorological data in the subtropical high pressure area in the northwest Pacific, and resampling the digital elevation data to obtain the altitude data and slope aspect data at the same resolution;
[0103] In the embodiment of the present application, historical meteorological data and digital elevation data of the target area are obtained, and historical meteorological data of the subtropical high pressure (hereinafter referred to as subtropical high) area in the northwest Pacific is obtained. The longitude and latitude range of the subtropical high area is selected from 110° to 150° east longitude and 10° to 50° north latitude; the historical meteorological data and digital elevation data are both grid data, wherein the spatial resolution of the meteorological data is 0.25°×0.25°, the time resolution is 1 hour, and the spatial resolution of the digital elevation data is 30m;
[0104] In the embodiment of the present application, the data is preprocessed. First, the historical meteorological data of the target area is calculated to obtain the daily average meteorological data; secondly, the pentad (every 5 days is 1 pentad) mean of the subtropical high data is calculated; finally, the digital elevation data is resampled to obtain the altitude data with a spatial resolution of 1km×1km, and the slope aspect data at the same resolution is calculated;
[0105] In the embodiment of the present application, the meteorological data of the subtropical high region in the past 10 years are used to train the MSConvLSTM (Multistep ConvLSTM) model. The task objective is to predict the meteorological data of the next 9 pentads (45 days) by using the data of the previous 18 pentads (90 days).
[0106] S2: Based on the pentad mean of the standardized Northwest Pacific subtropical high pressure data, the correlation feature map between the Northwest Pacific subtropical high pressure and the target area is obtained through the pre-trained MSConvLSTM model and SCW model;
[0107] In the embodiment of the present application, the standardized data of the first 18 phases of subtropical high are input into the subtropical high characteristic factor fusion module, and the change characteristics of the subtropical high in the next 9 phases are obtained through the pre-trained MSConvLSTM model in the module, and then the correlation feature map of the subtropical high and the target area is obtained through the SCW (Spatial Correlation Weighting) model;
[0108] In the embodiment of the present application, the subtropical high characteristic factor fusion module is composed of MSConvLSTM and SCW model in series, in which MSConvLSTM uses a pre-training method to obtain model parameters first, that is, using subtropical high data and taking the average value of the pentad, predicting the potential height of the next 9 pentads through the previous 18 pentads; the reason why the subtropical high area is selected in the range of 110°~150° east longitude and 10°~50° north latitude is that it is the main part that affects the year-round activity. A similar rolling prediction method is used, such as Figure 7 As shown, the first step uses the previous 18 Hou to predict the next step G T , G T The output is divided into two branches, one is direct output, and the other is obtained through an encoder-decoder structure. T This structure is nested in the connection layer of each layer, which can effectively reduce the impact of cumulative errors. In the second step, the G predicted by the previous 17 candidates and the previous step is used. T The data of 18 phases are used to predict the next step G T+1 , which is also divided into two branch outputs; and so on, the potential height prediction for the next 9 seasons is finally obtained; the encoder-decoder structure consists of two 3×3 convolutional layers and two upsampling layers, such as Figure 8 As shown;
[0109] In the embodiment of the present application, the SCW model has the following workflow: the potential height data of the first three phases of the subtropical high and the next eight phases, as well as the longitude and latitude data of the subtropical high area and the target area are input in parallel, and the potential height data is subjected to convolution-pooling operations to obtain the corresponding feature map F; the longitude and latitude data of the subtropical high are subjected to two layers of convolution layers to obtain the feature map F1 of this stage; the longitude and latitude data of the target area are first subjected to upsampling to expand the data size, and then are subjected to the convolution layer to obtain the feature map F2 of this stage; then F1 and F2 enter the shared convolution layer to obtain their respective spatial feature weights W1 and W2, which are added to obtain the spatial correlation feature weight W. The shared convolution layer includes the pooling layer-convolution layer-BN layer, and the shared convolution layer parameters are shared. It can learn both the spatial feature information of the subtropical high and the spatial feature information of the target area, and realize the mutual correlation of geographic spatial information. Finally, the model multiplies F and W by weight to obtain the relevant feature map of the wind speed of the subtropical high and the target area;
[0110] S3: Based on the standardized daily average meteorological data, the characteristic graph of each meteorological variable is obtained through the 3D-CBAM-ConvLSTM model;
[0111] In the embodiment of the present application, the standardized historical meteorological data of the target area in the previous 90 days is input into the historical real-time feature extraction module, and the characteristic graph of each meteorological variable is obtained through the 3D-CBAM-ConvLSTM (3D Convolutional Block Attention Module ConvLSTM) in the module;
[0112] In the embodiment of the present application, the historical live feature extraction module is composed of 3D-CBAM-ConvLSTM, where 3D-CBAM is composed of a 3D channel attention module and a 3D spatiotemporal attention module, such as Figure 4 As shown, the input data first passes through the 3D channel attention module, such as Figure 5 As shown, the feature map F is obtained C , then the feature map F C Enter the 3D spatiotemporal attention module and obtain the feature map F CS .
[0113]
[0114] Where F is the input data, is the vector cross product, M c is the channel attention weight, F c is the feature map after channel weighting, M s is the spatiotemporal attention weight, F cs It is a feature map after channel and spatiotemporal weighting.
[0115] In the embodiment of the present application, the 3D channel attention module is composed of two pooling layers, a multi-layer perceptron (MLP), and an activation function. The specific steps of the 3D channel attention module are: first, the spatial information of the input data is gathered by using 3D maximum pooling and average pooling, and then the feature graphs that pass through the MLP are summed based on the corresponding elements, and then the feature attention weight is obtained by the sigmoid activation function. The calculation formula of the 3D channel attention module is:
[0116] M C (F)=σ(MLP(AvgPool3D(F))+MLP(MaxPool3D(F))) (3)
[0117] Where MLP stands for multi-layer perceptron operation and σ stands for sigmoid activation function operation.
[0118] In the embodiment of the present application, the spatiotemporal attention module is composed of two pooling layers, a convolutional layer and an activation function, such as Figure 6As shown in the figure, the specific steps of the spatiotemporal attention module are as follows: first, the output feature map of the channel attention module is subjected to channel-based maximum pooling and average pooling operations, then the extracted features are merged through channels, and then a convolution operation with a convolution kernel of 7×7×7 is performed and the dimension is reduced to 1 channel number, and finally the spatiotemporal attention weight map is output through the sigmoid activation function. The calculation formula of the 3D spatiotemporal attention module is:
[0119] M S (F) = σ(f 7×7×7 ([AvgPool3D(F);MaxPool3D(F)])) (4)
[0120] In the formula, f 7×7×7 represents the convolution operation with a convolution kernel of 7×7×7, and σ represents the sigmoid activation function.
[0121] It should be noted that the present invention improves the channel attention by changing the original 2D pooling and convolution into 3D pooling and convolution. On the one hand, it adapts to the needs of input data, and on the other hand, the spatiotemporal attention mechanism can effectively focus on important spatial and temporal information, and weightedly highlight key features in the spatial and temporal dimensions.
[0122] In the embodiment of the present application, the ConvLSTM operation process in the 3D-CBAM-ConvLSTM model is as follows:
[0123] I t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i ) (5)
[0124] F t =σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f ) (6)
[0125] O t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t-1 +b o ) (7)
[0126] Ct =F t ⊙C t-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c ) (8)
[0127] H t =O t ⊙tanh(C t ) (9)
[0128] Among them, X t represents the input data at time t, I t 、F t and O t It is converted into a gating coefficient between 0 and 1 through the sigmoid operation. t The input data at time t is gated according to the gating coefficient. t According to the gating coefficient, the meteorological status information transmitted at time t-1 is selectively forgotten, forgetting the historical meteorological characteristics that are not important in long-term memory. Output gate O t Represents the historical meteorological characteristics output according to the gating coefficient. t It is a memory unit used to memorize the time-varying meteorological characteristics of multiple time steps. It recalculates a candidate value at each moment and scales the unit state according to the information updated by the forget gate and the input gate. t represents the hidden state of the ConvLSTM cell output at time t; W and b represent the weight and bias of the gating mechanism; ⊙, * and σ represent Hadamard, Matmul and Sigmoid operations respectively, as shown in Figure 3 After the model calculation, the characteristic graph of each meteorological variable is output.
[0129] S4: Based on the standardized elevation data and slope data, the terrain feature map is obtained through the Res-FPN model;
[0130] In the embodiment of the present application, the standardized altitude and aspect data are input into the terrain feature fusion module, and the terrain feature map is obtained through the Res-FPN (Residual-Feamre Pyramid Net) model in the module;
[0131] In the embodiment of the present application, the terrain feature fusion module is composed of a Res-FPN network model, such as Figure 2As shown in the figure, the Res-FPN model consists of 4 layers of residual blocks. The number of channels of the output data of each layer will be expanded to twice the original number, and the size will be reduced to half. The output of each layer from left to right is called C1, C2, C3, and C4. The C4 data is then divided into two branches for parallel processing. One branch obtains the feature map D4 of this layer through the residual connection module. The backbone of the residual connection module is a convolutional neural network, and the residual connection is composed of global average pooling. The other branch obtains features of the same size as C3 through upsampling, and the C3 data is added point by point with the upsampled C4 data after adjusting the number of channels through 1x1 convolution, and then the feature map D3 of this layer is obtained through the residual connection module. The subsequent fusion of C3 and C2 data is similar. Res-FPN can enhance the fusion of multi-scale terrain features and strengthen generalization capabilities.
[0132] S5: Channel splicing of the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map;
[0133] S6: Input the channel spliced data into the prediction model to predict the future daily average wind speed.
[0134] It should be noted that the present invention includes grid data of multiple meteorological elements and terrain factors in the area, as well as telerelated subtropical high data; different modules are used to process heterogeneous data to better extract the characteristic information hidden in the data to meet the prediction needs; a pre-training method is used to obtain the changing characteristics of the subtropical high, and the external conditions of the region are obtained, effectively improving the accuracy of wind speed prediction; the U-Net end-to-end prediction method is used, and the model has an efficient training and reasoning process, which can meet the needs of the prediction task.
[0135] This embodiment also provides a wind speed prediction system integrating teleconnection factors, including:
[0136] A preprocessing module is used to obtain historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, preprocess the data, and standardize the preprocessed data. The preprocessing includes calculating the historical meteorological data of the target area to obtain daily average meteorological data, calculating the pentad mean of the subtropical high pressure data of the northwest Pacific Ocean through the historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, and resampling the digital elevation data to obtain altitude data and aspect data at the same resolution;
[0137] The subtropical high characteristic factor fusion module is used to obtain the correlation characteristic map between the subtropical high pressure in the northwest Pacific and the target area through the pre-trained MSConvLSTM model and SCW model based on the pentad value of the standardized subtropical high pressure data in the northwest Pacific;
[0138] The historical real-time feature extraction module is used to obtain the feature map of each meteorological variable through the 3D-CBAM-ConvLSTM model based on the standardized daily average meteorological data;
[0139] The terrain feature fusion module is used to obtain the terrain feature map through the Res-FPN model based on the standardized altitude data and slope aspect data;
[0140] The splicing module is used to splice the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map;
[0141] The prediction module is used to input the channel spliced data into the prediction model to predict the future daily average wind speed.
[0142] Furthermore, it also includes:
[0143] Memory, used to store programs;
[0144] A processor is used to load the program to execute the wind speed prediction method integrating teleconnection factors.
[0145] This embodiment also provides a computer-readable storage medium storing a program, and when the program is executed by a processor, the wind speed prediction method integrating teleconnection factors is implemented.
[0146] The storage medium proposed in this embodiment and the wind speed prediction method integrating telecorrelation factors proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.
[0147] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ReadOnly, Memory, ROM), random access memory (RandomAccess Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform the methods of various embodiments of the present invention.
[0148] Example 2
[0149] Reference Figure 9-10, which is an embodiment of the present invention, provides a wind speed prediction method integrating teleconnection factors. In order to verify its beneficial effects, comparison results of multiple schemes are provided.
[0150] Step 1: Obtain historical meteorological data and digital elevation data of the target area, and obtain historical meteorological data of the subtropical high pressure (hereinafter referred to as subtropical high) area in the northwest Pacific. The longitude and latitude range of the subtropical high area is selected between 110° and 150° east longitude and 10° and 50° north latitude. The historical meteorological data and digital elevation data are both grid data, in which the spatial resolution of meteorological data is 0.25° × 0.25°, the time resolution is 1 hour, and the spatial resolution of digital elevation data is 30m. The specific operations include:
[0151] In this embodiment, Guizhou Province is located in the range of 103.5°~109.5° east longitude and 24.5°~29.5° north latitude. The historical meteorological data of the region from 2019 to 2023 is obtained from the ERA5 dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF). The time resolution is 1 hour and the spatial resolution is 0.25°. The meteorological elements include 2-meter temperature, 2-meter dew point temperature, 10-meter U and V components of wind, 100-meter U and V components of wind, and surface pressure. The digital elevation data comes from the ASTER GDEM V2 global digital elevation data product with a spatial resolution of 30m;
[0152] The subtropical high area is located in the range of 110°~150°E and 10°~50°N. The geopotential height on the 500hPa equipotential surface in the region from 2009 to 2023 is obtained from ERA5 with a temporal resolution of 1 hour and a spatial resolution of 0.5°×0.5°.
[0153] Step 2: Preprocess the data. First, calculate the historical meteorological data of the target area to obtain the daily average meteorological data; secondly, calculate the pentad mean (every 5 days is 1 pentad) of the subtropical high data; finally, resample the digital elevation data to obtain the altitude data with a spatial resolution of 1km×1km, and calculate the slope aspect data at the same resolution; the specific operations include:
[0154] In this embodiment, the daily mean of meteorological data is first calculated, and then the mean of every 5 days is calculated when processing the subtropical high data; the digital elevation data is processed using ArcGis software, and the digital elevation data is resampled into altitude data with a spatial resolution of 1km×1km using the built-in toolbox, and the slope aspect data with the same resolution is calculated through the altitude data;
[0155] Step 3: Use the meteorological data of the subtropical high region in the past 10 years to train the MSConvLSTM (MultistepConvLSTM) model. The task goal is to predict the meteorological data of the next 9 pentads (45 days) based on the data of the previous 18 pentads (90 days).
[0156] The MSConvLSTM predicts the potential height of the next 9 Hou by using the previous 18 Hou, using a method similar to rolling prediction. The first step uses the previous 18 Hou to predict the situation of the next step G T , G T The output is divided into two branches, one is direct output, and the other is obtained through an encoder-decoder structure. T This structure is nested in the connection layer of each layer, which can effectively reduce the impact of cumulative errors. In the second step, the G predicted by the previous 17 candidates and the previous step is used. T The data of 18 phases are used to predict the next step G T+1 , which is also divided into two branch outputs; and so on, the potential height prediction for the next 9 phases is finally obtained; the encoder-decoder structure consists of two 3×3 convolutional layers and two upsampling layers;
[0157] Specific operations include:
[0158] The input data is the normalized geopotential height of the first 18 pentads in the subtropical high region. The normalization formula is:
[0159] X out =(XX mean ) / X std (10)
[0160] In the formula, X is the input data, X out is the standardized data, X mean is the mean of the input data, X std is the standard deviation of the input data.
[0161] The training labels are the potential heights for the next 9 pentads, and the model is trained using the Adam optimizer.
[0162] Step 4: Input the standardized subtropical high data of the first 18 pentads into the subtropical high characteristic factor fusion module, and obtain the variation characteristics of the subtropical high in the next 9 pentads through the pre-trained MSConvLSTM model in the module. Then, the correlation feature map between the subtropical high and the target area is obtained through the SCW (Spatial Correlation Weighting) model.
[0163] SCW model, such as Fig. 9As shown in the figure, the workflow is as follows: the potential height data of the first three phases of the subtropical high and the next nine phases, as well as the longitude and latitude data of the subtropical high area and the target area are input in parallel, and the potential height data is subjected to convolution-pooling operations to obtain the corresponding feature map F; the longitude and latitude data of the subtropical high are subjected to two layers of convolution to obtain the feature map F1 of this stage; the longitude and latitude data of the target area are first subjected to upsampling to expand the data size, and then passed through the convolution layer to obtain the feature map F2 of this stage; then F1 and F2 enter the shared convolution layer to obtain their respective spatial feature weights W1 and W2, and add them to obtain the spatial correlation feature weight W. The shared convolution layer includes the pooling layer-convolution layer-BN layer, and the shared convolution layer parameters are shared. It can learn the spatial feature information of both the subtropical high and the target area, and realize the mutual correlation of geographic spatial information. Finally, the model multiplies F and W by weight to obtain the relevant feature map of the wind speed of the subtropical high and the target area, and its dimension is (batchsize, 12, 81, 81);
[0164] Step 5: Input the standardized historical meteorological data of the target area in the first 90 days into the historical real-time feature extraction module, and obtain the feature map of each meteorological variable through the 3D-CBAM-ConvLSTM (3D Convolutional Block Attention ModuleConvLSTM) in the module;
[0165] The historical live data includes 2-meter temperature, 2-meter dew point temperature, 10-meter U and V components of wind, 100-meter U and V components of wind, and surface pressure. After calculation and processing in step 2, the daily mean is obtained. The dimension is (batchsize, 7, 90, 25, 21), where batchsize represents the batch size, 7 represents the number of variables, and (90, 25, 21) represents the daily average meteorological data of Guizhou in the previous 90 days. After standardization, it is input into the historical live feature extraction module;
[0166] The historical feature extraction module is composed of 3D-CBAM-ConvLSTM, where 3D-CBAM is composed of 3D channel attention module and 3D spatiotemporal attention module. First, the input data passes through the 3D channel attention module to obtain the feature map F C , then the feature map F C Enter the 3D spatiotemporal attention module and obtain the feature map F CS After the model calculation, the characteristic map of each meteorological variable is output, and its dimension is (batchsize, 7, 25, 21).
[0167] Step 6: Input the standardized altitude and aspect data into the terrain feature fusion module, and obtain the terrain feature map through the Res-FPN (Residual-Feature Pyramid Net) model in the module;
[0168] The terrain feature fusion module is composed of a Res-FPN network model. The Res-FPN model consists of 4 layers of residual blocks. The number of channels of the output data of each layer will be expanded to twice the original number, and the size will be reduced to half. The output of each layer from left to right is called C1, C2, C3, and C4. The C4 data is then divided into two branches for parallel processing. One branch obtains the feature map D4 of this layer through a residual connection module. The backbone of the residual connection module is a convolutional neural network, and the residual connection is composed of global average pooling. The other branch obtains features of the same size as C3 through upsampling, and the C3 data is added point by point with the upsampled C4 data after adjusting the number of channels through a 1x1 convolution, and then the feature map D3 of this layer is obtained through a residual connection module. The subsequent fusion of C3 and C2 data is similar. The final output terrain feature map dimension is (2, 25, 21).
[0169] Step 7: Channel stitching the correlation feature map between the subtropical high and the target area, the feature map of each meteorological variable and the terrain feature map; the specific operation process is as follows:
[0170] The terrain feature map is expanded, that is, processed into data with a dimension of (batchsize, 2, 25, 21), and then the outputs of the three modules are merged in the channel dimension to obtain a feature set with a dimension of (batchsize, 21, 25, 21);
[0171] Step 8: Input the channel spliced data into the U-net model to predict the daily average wind speed for the next 45 days;
[0172] The feature set in step 7 is sent to the U-Net model to obtain the output result. The U-Net is an encoder-decoder network model with a unique U-shaped structure. Its structure cleverly reflects the layer-by-layer encoding and decoding process of information. This design not only takes into account the upper and lower layer information, but also retains important spatial information. The left side of the U-shaped is the encoding part, which is mainly responsible for extracting the features of the input data. There are 3 layers of downsampling, and each downsampling layer contains 2 layers of 3×3 convolutional layers and a 2×2 maximum pooling layer; the right side of the U-shaped is the decoding part, which is mainly responsible for upsampling the feature map extracted by the encoder layer by layer. At the same time, each layer uses jump connections to transfer the encoder features to the decoder, which can alleviate the information loss in the end-to-end learning process.
[0173] The ratio of training set to test set is 4:1. The training label is the average wind speed of each day in the next 45 days. The model is trained using Adam optimizer with a learning rate of 0.01 and a loss function of root mean square error (RMSE). The formula is:
[0174]
[0175] In the formula, n represents the number of samples, m is the number of grid points, x is the predicted wind speed result, and y represents the actual wind speed of the sample. The smaller the RMSE value, the better the prediction result.
[0176] Use the trained model to predict the test set data and determine the model accuracy.
[0177] This example sets up an experiment to test three models. Case 1 is a model that only contains historical live data in this example, Case 2 is a model that contains historical live and terrain data in this example, and Case 3 is a model that contains all modules and corresponding data in this example. Fig.10 The RMSE errors of the three model schemes on the test set show that the RMSE index of the method of the present invention is the lowest compared with other schemes. Compared with only providing historical meteorological data, the RMSE index is improved by 9.6%; compared with providing historical meteorological data and terrain data, the RMSE index is improved by 5.1%; and it can be found that with the gradual addition of different data, the RMSE index of the model is significantly reduced, which reflects the effectiveness of both the data and the model.
[0178] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A wind speed prediction method integrating teleconnection factors, characterized in that: include: Acquire historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, preprocess the data, and standardize the preprocessed data, wherein the preprocessing includes calculating the historical meteorological data of the target area to obtain daily average meteorological data, calculating the pentad mean of the subtropical high pressure data of the northwest Pacific Ocean through the historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, and resampling the digital elevation data to obtain altitude data and aspect data at the same resolution; Based on the standardized pentad value of the Northwest Pacific subtropical high pressure data, the correlation feature map between the Northwest Pacific subtropical high pressure and the target area is obtained through the pre-trained MSConvLSTM model and SCW model; Based on the standardized daily average meteorological data, the characteristic graphs of each meteorological variable are obtained through the 3D-CBAM-ConvLSTM model; Based on the standardized elevation data and slope aspect data, the terrain feature map is obtained through the Res-FPN model; Channel splicing is performed on the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map; The channel spliced data is input into the prediction model to predict the average daily wind speed in the future.
2. The wind speed prediction method integrating teleconnection factors according to claim 1, characterized in that: The MSConvLSTM model includes the following steps: The pentad mean of the standardized Northwest Pacific subtropical high pressure data for the previous 18 pentads was used to construct the data set as the input of the model; The MSConvLSTM model is pre-trained using the dataset to learn and optimize model parameters; The pre-trained MSConvLSTM model adopts a rolling prediction strategy to form a nine-phase potential height prediction sequence; The rolling prediction strategy includes the following steps: Using the data of the previous 18 Hou as input, the potential height G of the 19th Hou is predicted. T , G T The output is divided into two branches. The first output is to directly output the preliminary prediction results. The second output is to introduce a deep learning module consisting of two layers of 3×3 convolutional layers and two layers of upsampling layers. T Perform deep feature extraction and reconstruction to obtain the optimized prediction result G′ T ,This module is embedded in the connection layer of each layer; The potential height G of the 19th Hou T Combined with the actual geopotential height data of the previous 17 Hou, it is used as the prediction of G at the next time point. T+1 Input The above process is repeated until a complete prediction sequence for the next 9 phases is obtained.
3. The wind speed prediction method integrating teleconnection factors according to claim 1 or 2, characterized in that: The SCW model includes the following steps: The geopotential height data of the first three phases of the subtropical high and the next eight phases, as well as the longitude and latitude data of the subtropical high area and the target area were input in parallel. The geopotential height data was used to obtain the corresponding feature map F through convolution-pooling operations; The subtropical high latitude and longitude data are processed through two convolutional layers to obtain the feature map F1 of this stage; The longitude and latitude data of the target area are upsampled to expand the data size, and the feature map F2 is obtained through the convolution layer; F1 and F2 enter the shared convolutional layer to obtain their respective spatial feature weights W1 and W2, and add them together to obtain the spatial correlation feature weight W; Multiply F and W and weight them to obtain the correlation characteristic map between the subtropical high and the wind speed in the target area.
4. The wind speed prediction method integrating teleconnection factors as claimed in claim 3 is characterized in that: The 3D-CBAM-ConvLSTM, wherein the 3D-CBAM is composed of a 3D channel attention module and a 3D spatiotemporal attention module, the 3D-CBAM, comprises the following steps: The input data passes through the 3D channel attention module to obtain the feature map F C ; Feature map F C Enter the 3D spatiotemporal attention module and obtain the feature map F CS ; The 3D channel attention module comprises the following steps: Use 3D maximum pooling and average pooling to aggregate the spatial information of the input data; Perform element-based addition operations on the feature maps that have passed through MLP respectively; The summed result is processed by the sigmoid activation function to obtain the attention weight of each channel; The attention weight of each channel is multiplied element-by-element with the corresponding channel in the original input data, and the output feature map F is C ; The 3D spatiotemporal attention module comprises the following steps: Perform channel-based maximum pooling and average pooling operations on the output feature map of the channel attention module; Merge channels of the extracted features; Through a convolution operation with a convolution kernel of 7×7×7 and reduce the dimension to 1 channel; Generate the spatiotemporal attention weights for each position through the sigmoid activation function; Combine the spatiotemporal attention weights with F C Perform element-by-element multiplication to generate the spatiotemporal attention feature map F CS .
5. The wind speed prediction method integrating teleconnection factors as claimed in claim 4, characterized in that: The calculation formula of the 3D channel attention module is expressed as: M C (F)=σ(MLP(AvgPool3D(F))+MLP(MaxPool3D(F))) Among them, MLP represents multi-layer perceptron operation, σ represents sigmoid activation function operation; The calculation formula of the 3D spatiotemporal attention module is expressed as: M S (F)=σ(f 7×7×7 ([AvgPool3D(F);MaxPool3D(F)])) Among them, f 7×7×7 represents the convolution operation with a convolution kernel of 7×7×7, and σ represents the sigmoid activation function.
6. The method for wind speed prediction integrating teleconnection factors as claimed in claim 5, characterized in that: The ConvLSTM operation process in the 3D-CBAM-ConvLSTM model is expressed as: I t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i ) F t =σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f ) O t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t-1 +b o ) C t =F t ⊙C t-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c ) H t =O t ⊙tanh(C t ) Among them, X t represents the input data at time t, I t 、F t and O t is converted into a gating coefficient between 0 and 1 through the sigmoid operation, C t is a memory unit, H t represents the hidden state of the ConvLSTM cell output at time t; W and b represent the weight and bias of the gating mechanism; ⊙, *, and σ represent Hadamard, Matmul, and Sigmoid operations, respectively.
7. The method for wind speed prediction integrating teleconnection factors as claimed in claim 6, characterized in that: The Res-FPN model includes 4 layers of residual blocks. The number of channels of each layer of output data is expanded to twice the original number and the size is halved to obtain four feature maps of different resolutions: C1, C2, C3, and C4. The C4 feature map is processed by the residual connection module to obtain the feature map D4; The C4 feature map is upsampled to the same size as C3. At the same time, the C3 feature map adjusts the number of channels through 1x1 convolution. The C3 feature map with adjusted channel number is added point-to-point with the upsampled C4 feature map, and the feature map D3 is obtained through the residual connection module; The D3 feature map is upsampled to the same size as C2. The C2 feature map also adjusts the number of channels through 1x1 convolution. The C2 feature map with adjusted channel number is added point-to-point with the upsampled D3 feature map, and the feature map D2 is obtained through the residual connection module. The D2 feature map is upsampled to the same size as C1, and the C1 feature map adjusts the number of channels through 1x1 convolution. The C1 feature map with adjusted channel number is added point-to-point with the upsampled D2 feature map, and then the feature map D1 is obtained through the residual connection module; All generated feature maps D1, D2, D3, and D4 are fused to obtain a terrain feature map.
8. A system based on the wind speed prediction method integrating teleconnection factors according to claim 1, characterized in that: A preprocessing module is used to obtain historical meteorological data and digital elevation data of the target area and historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, preprocess the data, and standardize the preprocessed data. The preprocessing includes calculating the historical meteorological data of the target area to obtain daily average meteorological data, calculating the pentad mean of the subtropical high pressure data of the northwest Pacific Ocean through the historical meteorological data of the subtropical high pressure area of the northwest Pacific Ocean, and resampling the digital elevation data to obtain altitude data and aspect data at the same resolution; The subtropical high characteristic factor fusion module is used to obtain the correlation characteristic map between the subtropical high pressure in the northwest Pacific and the target area through the pre-trained MSConvLSTM model and SCW model based on the pentad value of the standardized subtropical high pressure data in the northwest Pacific; The historical real-time feature extraction module is used to obtain the feature map of each meteorological variable through the 3D-CBAM-ConvLSTM model based on the standardized daily average meteorological data; The terrain feature fusion module is used to obtain the terrain feature map through the Res-FPN model based on the standardized altitude data and slope aspect data; The splicing module is used to splice the correlation characteristic map of the Northwest Pacific subtropical high pressure and the target area, the characteristic map of each meteorological variable and the terrain characteristic map; The prediction module is used to input the channel spliced data into the prediction model to predict the future daily average wind speed.
9. An electronic device, characterized in that: include: Memory, used to store programs; A processor is used to load the program to execute the steps of the wind speed prediction method integrating teleconnection factors as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the steps of the wind speed prediction method integrating teleconnection factors as described in any one of claims 1 to 7 are implemented.