Long-term wind speed forecast correction method based on network fusion

Abstract

The invention discloses a long-term wind speed forecast correction method based on network convergence. The method comprises the following steps: collecting multi-source meteorological data, and preprocessing the collected forecast data; a dynamic spatio-temporal feature matrix is constructed, the time features, the spatial features and the meteorological element features are input into a wind speed correction model which is qualified in training, and the wind speed correction model comprises a spatial feature extraction network, a time sequence feature extraction network and a feature fusion network; the spatial feature extraction network extracts spatial feature vectors from the spatial features and the meteorological element features, the time sequence feature extraction network extracts time sequence feature vectors from the time features and the meteorological element features, and the feature fusion network outputs a wind speed correction value. Through a multi-task learning strategy and in combination with a fusion architecture of CNN and LSTM, the problem of insufficient precision of a traditional single model in complex terrain and multi-element prediction is solved, and the space-time modeling capability and complex region adaptability of wind speed prediction are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather forecasting, and particularly relates to a long-term wind speed forecasting correction method based on network fusion. BACKGROUND

[0002] With the continuous development of weather forecasting technology, the accuracy requirement of wind speed forecasting is getting higher and higher. At present, in the 0-168 hour per hour forecast and the 168-360 hour per 3 hour forecast, there are still certain errors in the prediction of wind speed elements. These errors mainly come from the following aspects: Firstly, the uneven spatio-temporal distribution of meteorological data is an important reason for the prediction error. Due to the uneven distribution of observation sites and the error of observation instruments, there is a certain spatio-temporal unevenness in meteorological data, which will gradually magnify with the extension of the prediction time.

[0003] Secondly, the complex relationship between meteorological elements is also an important factor affecting the prediction accuracy. There are complex interactions and influences between meteorological elements, and these relationships are often nonlinear, which is difficult to describe with a simple mathematical model.

[0004] Thirdly, the influence of terrain and underlying surface characteristics cannot be ignored. Complex terrain and underlying surface characteristics will have a significant impact on near-surface wind speed, and these influences are often nonlinear, which is difficult to describe with a simple mathematical model.

[0005] Traditional correction methods mainly rely on statistical methods, such as linear regression, Kalman filtering, etc. These methods have the following limitations when dealing with nonlinear and non-stationary time series: 1. Assuming that the wind speed change conforms to the linear relationship, it is difficult to handle the nonlinear coupling between meteorological elements (such as the nonlinear response of temperature field to wind speed).

[0006] 2. Unable to model the non-uniform influence of complex terrain (such as mountains, coastal areas) on wind speed, resulting in significant increase in local area prediction error.

[0007] 3. Relying on manual feature engineering, it is difficult to automatically extract spatio-temporal features (such as wind speed gradient, periodic change).

[0008] Although deep learning models can solve the above problems of statistical methods, single deep learning models (such as LSTM or CNN) still have the following limitations: 1. LSTM can capture temporal dependencies, but its modeling ability for spatial features (such as terrain distribution) is weak, resulting in decreased prediction accuracy in mountainous areas.

[0009] 2. CNN can extract spatial features, but lacks the ability to model long-term time dependence, leading to serious error accumulation in long-term prediction (>168 hours).

[0010] 3. The correlation of multi-wind speed elements (gust10, W10, W100) is not considered, and multiple models need to be trained independently, which is high in calculation cost and low in efficiency. SUMMARY

[0011] The purpose of the present application is to overcome the shortcomings of the prior art and provide a long-term wind speed prediction correction method based on network fusion.

[0012] To achieve the above purpose, the present application provides a long-term wind speed prediction correction method based on network fusion, comprising: Collecting multi-source meteorological data, the multi-source meteorological data including short-term prediction data and long-term prediction data; Pretreating the collected multi-source meteorological data; Constructing a dynamic spatio-temporal feature matrix according to the pretreated multi-source meteorological data, including time features, spatial features and meteorological element features; Inputting the time features, spatial features and meteorological element features into a trained wind speed correction model respectively, the wind speed correction model including a spatial feature extraction network, a time series feature extraction network and a feature fusion network; the spatial feature extraction network is used to extract spatial feature vectors from spatial features and meteorological element features, the time series feature extraction network is used to extract time series feature vectors from time features and meteorological element features, and the feature fusion network is used to splice the spatial feature vectors and time series feature vectors to form a joint feature vector, fuse the spatial features and time series features through a learnable weighting mechanism, and output a wind speed correction value.

[0013] Further, the short-term prediction data is 0-168 hours per hour, and the long-term prediction data is 168-360 hours per 3 hours prediction data, the short-term prediction data and long-term prediction data both include 10 meter gust wind speed, 10 meter wind speed, 100 meter wind speed and auxiliary meteorological element data, and the auxiliary meteorological element data includes temperature field, pressure field, humidity field and terrain height.

[0014] Further, the pretreatment includes data cleaning, spatio-temporal alignment and data standardization in sequence; Further, the data cleaning is as follows: Using 3σ criterion to eliminate abnormal values exceeding ±3 times of standard deviation; Using cubic spline interpolation to fill in missing time series data; The spatio-temporal alignment is as follows: The forecast data after data cleaning is aligned with Coordinated Universal Time (UTC) and the time resolution is unified to 1 hour; The forecast data is spatially aligned by bilinear interpolation; The data standardization is as follows: The continuous variables in the forecast data are standardized by the following method: ; Wherein, is the continuous variable after standardization, is the continuous variable before standardization, is the mean of the continuous variable, is the standard deviation of the continuous variable; The discrete variables in the forecast data are normalized by the following method: ; Wherein, is the discrete variable after normalization, is the discrete variable before normalization, is the upper threshold of the discrete variable, is the lower threshold of the discrete variable.

[0015] Further, the time features include forecast time, time period features, and historical trend features, the spatial features include grid point position features, spatial gradient features, and terrain normalization features, and the meteorological element features include temperature field, pressure field, and humidity field data processed by standardization.

[0016] Further, the spatial feature extraction network comprises: A first input layer for receiving input spatial features and meteorological element features; A multi-scale convolution structure comprising a first convolution layer, a pooling layer, a second convolution layer, and a global pooling layer connected in sequence, wherein the first convolution layer extracts local spatial features through a convolution kernel and a ReLU activation function, the pooling layer reduces the spatial dimension and retains key features using maximum pooling, the second convolution layer uses the same convolution kernel as the first convolution layer but increases the number of output channels to extract higher-level spatial features, and the global pooling layer compresses the spatial features into a one-dimensional vector through global average pooling; A terrain effect modeling unit for introducing dynamic normalization features of terrain height in the convolution structure to enhance the spatial feature extraction capability of complex regions; A first output layer for outputting compressed spatial feature vectors containing the spatial distribution pattern of meteorological elements at the current time.

[0017] Further, the time sequence feature extraction network comprises: a second input layer configured to receive inputted time features and meteorological element features; a multi-layer recurrent neural network structure, comprising a first layer recurrent neural network and a second layer recurrent neural network, wherein the first layer recurrent neural network contains a plurality of hidden units and returns a sequence to retain time step information, and the second layer recurrent neural network contains fewer hidden units and extracts final time-dependent features, and a gradient vanishing problem is avoided through a nonlinear activation function; a time period encoding module configured to introduce sine / cosine encoding in the inputted time features to capture hourly / daily periodic variation rules; a second output layer configured to output a time sequence feature vector to capture the evolution rules of meteorological elements over time.

[0018] Further, the feature fusion network comprises: a dynamic fusion module configured to fuse the spatial feature vector output by the spatial feature extraction network and the time sequence feature vector output by the time feature extraction network through a learnable weighting mechanism to form a joint feature vector; a multi-task mapping layer configured to map the joint feature vector to the prediction dimension of different wind speed elements through a fully connected layer with shared weights; a third output layer configured to output continuous wind speed correction values using a linear activation function.

[0019] Further, the first N layers of the spatial feature extraction network and the time feature extraction network share weights to extract general spatiotemporal features, and the model adaptability of complex regions is enhanced through terrain normalization features.

[0020] Further, a plurality of independent prediction branches are arranged after the feature fusion network, each branch is adapted to the prediction requirements of different wind speed elements, and shared feature layers are used to reduce redundant calculations.

[0021] Beneficial effects: 1. The application first combines a convolutional neural network (CNN) with a long short-term memory network (LSTM), extracts spatial features through the CNN, and captures time sequence dependencies through the LSTM, thereby solving the technical bottleneck that traditional single models cannot simultaneously model spatiotemporal complexity, and providing more accurate prediction correction results. The application can reduce wind speed prediction error by more than 40% compared with traditional methods; 2. The application designs a feature sharing method to extract general spatiotemporal features, and designs an independent branch network for each wind speed element to adapt to the prediction requirements of different elements, thereby significantly improving the multi-target correction efficiency; 3. The dynamic normalization features of terrain height are introduced, and the wind speed gradient of adjacent grids is extracted through a spatial convolution kernel, thereby realizing end-to-end modeling of terrain effects and wind speed prediction for the first time, and solving the problem of low wind speed prediction accuracy in complex terrain regions. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the long-term wind speed forecast correction method based on network fusion according to an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0024] like Figure 1 As shown, this embodiment of the invention provides a long-term wind speed forecast correction method based on network fusion, including: Multi-source meteorological data is collected, including both short-term and long-term forecast data. Details are as follows: Short-term forecast data: hourly forecast data from 0 to 168 hours; Long-term forecast data: 3-hourly forecast data from 168 to 360 hours.

[0025] Short-term and long-term forecast data specifically include 10-meter gust wind speed, 10-meter wind speed, 100-meter wind speed, and auxiliary meteorological element data, which include temperature field, pressure field, humidity field, and terrain height.

[0026] The collected multi-source meteorological data undergoes preprocessing. Specifically, preprocessing includes data cleaning, spatiotemporal alignment, and data standardization.

[0027] The data cleaning methods described above are as follows: 1) Outlier detection: Use the 3σ criterion to remove outliers that exceed ±3 times the standard deviation; 2) Missing value handling: Cubic spline interpolation is used to fill missing time series data.

[0028] The spatiotemporal alignment methods mentioned above are as follows: 1) Time alignment: Align the cleaned historical forecast data with the forecast data according to Coordinated Universal Time (UTC), with a unified time resolution of 1 hour; 2) Spatially align historical forecast data with forecast data (at the same resolution) using bilinear interpolation.

[0029] The above data standardization methods are as follows: Z-Score standardization: Continuous variables such as wind speed and temperature in forecast data are standardized using the following methods: ; where, is the standardized continuous variable, is the continuous variable before standardization, is the mean of the continuous variable, is the standard deviation of the continuous variable. Discrete variables such as terrain height in the forecast data are normalized as follows: ; where, is the normalized discrete variable, is the discrete variable before normalization, is the upper threshold of the discrete variable, is the lower threshold of the discrete variable.

[0030] According to the pre-processed multi-source weather data, a dynamic spatio-temporal feature matrix is constructed, which specifically includes: 1) Time features: Forecast time (lead_time): the time difference between the current forecast time and the start time.

[0031] Time period features: extract the sine encoding of hours , the sine encoding of days .

[0032] Historical trend features: calculate the moving average of wind speed in the past 24 hours .

[0033] 2) Spatial features: Grid point position features: sine / cosine encoding of longitude / latitude coordinates (lon / lat).

[0034] Spatial gradient features: extract the wind speed gradient of adjacent grid points through a convolution kernel (kernel_size=3×3).

[0035] Terrain normalization features: normalize the terrain height (HGT) as an input layer channel.

[0036] 3) Meteorological element features: Temperature field (T2m), pressure field (MSL), humidity field (RH) data processed by standardization.

[0037] The time feature, the space feature and the meteorological element feature are respectively input into a trained wind speed correction model, and the wind speed correction model comprises a space feature extraction network, a time sequence feature extraction network and a feature fusion network. The space feature extraction network is used to extract a space feature vector from the space feature and the meteorological element feature, the time sequence feature extraction network is used to extract a time sequence feature vector from the time feature and the meteorological element feature, and the feature fusion network is used to splice the space feature vector and the time sequence feature vector to form a joint feature vector, and output a wind speed correction value according to the joint feature vector. The wind speed correction value can be saved as a MICAPS type 4 format file, and meteorological business platform can directly call and display.

[0038] The space feature extraction network of the embodiment of the application is a convolutional neural network (CNN), which specifically comprises: 1) first input layer Input dimension: the input is multi-channel two-dimensional spatial grid data, including historical wind speed, temperature, air pressure and other meteorological element features and constructed space features.

[0039] Example input: assuming that the input is a tensor with a shape of [time step, latitude, longitude, channel number], for example, [1 (current time), 100, 100, 12], wherein the 12 channels represent different meteorological elements (such as wind speed, temperature, terrain, space features, meteorological element features, etc.).

[0040] 2) multi-scale convolution structure The multi-scale convolution structure comprises a first convolution layer, a pooling layer, a second convolution layer and a global pooling layer connected in sequence, wherein the first convolution layer extracts local space features through a convolution kernel and a ReLU activation function, the pooling layer reduces the spatial dimension and retains key features by using maximum pooling, the second convolution layer uses the same convolution kernel as the first convolution layer but increases the output channel number to extract higher-level space features, and the global pooling layer compresses the space features into a one-dimensional vector through global average pooling. The first convolution layer preferably uses a 3x3 convolution kernel, and the step length thereof is 1, the padding of the first convolution layer is same, and the output channel number thereof is 32, so as to extract local space features (such as wind speed gradient, temperature field distribution). The output channel number of the second convolution layer is preferably 64.

[0041] 3) terrain effect modeling unit The dynamic normalization feature of the terrain height is introduced in the convolution structure to enhance the space feature extraction capability of the complex region.

[0042] 4) first output layer An compressed space feature vector is output, which contains the spatial distribution pattern of the meteorological elements at the current time.

[0043] The timing feature extraction network of the embodiment of the application is a long short-term memory network (LSTM), which specifically comprises: 1) a second input layer Input dimension: The input is the timing feature and the meteorological element feature, for example, the timing change of meteorological elements such as wind speed and temperature in the past 24 hours.

[0044] Example input: Assuming that the input is a tensor with a shape of [time step, feature number], for example, [24 (past 24 hours), 12], where the 12 features include wind speed, temperature, humidity, etc.

[0045] 2) a multi-layer recurrent neural network structure Including a first layer of recurrent neural network and a second layer of recurrent neural network, wherein the first layer of recurrent neural network contains a plurality of hidden units and returns a sequence to retain time step information, and the second layer of recurrent neural network contains fewer hidden units and extracts final time-dependent features, and avoids the problem of gradient disappearance through a nonlinear activation function.

[0046] 3) a time period encoding module Sinusoidal / cosine encoding is introduced in the input timing feature to capture the hourly / daily periodic change rule.

[0047] 4) a second output layer An compressed timing feature vector is output to capture the evolution rule of meteorological elements over time, such as wind speed trend and periodic change.

[0048] The feature fusion network of the embodiment of the application comprises: 1) a dynamic fusion module The spatial feature vector output by the spatial feature extraction network and the timing feature vector output by the timing feature extraction network are fused through a learnable weighting mechanism (such as an attention mechanism) to form a joint feature vector.

[0049] 2) a multi-task mapping layer The joint feature is mapped to the prediction dimension of different wind speed elements (such as the corrected wind speed value) through a fully connected layer with shared weights.

[0050] 3) a third output layer A linear activation function (activation='linear') is used to directly output continuous wind speed correction values.

[0051] In addition, the first two layers (the first input layer, the convolutional layer, the second input layer and the LSTM layer) of the spatial feature extraction network and the timing feature extraction network share weights to extract general spatio-temporal features. An independent branch network can also be designed for each wind speed element (such as gust10, W10, W100, etc.) after the feature fusion network to adapt to the prediction requirements of different elements.

[0052] The training and optimization mode of the above wind speed correction model is as follows: 1) Training data collection Historical prediction data and corresponding real-time observation data: 0-168 hour hourly and 168-360 hour 3-hourly prediction data are obtained from GFS numerical model, including 10 meter gust wind speed (gust10), 10 meter wind speed (W10), 100 meter wind speed (W100) and auxiliary meteorological elements: temperature field (T2m), pressure field (MSL), humidity field (RH), terrain height (HGT). Real-time observation data are obtained from hourly measured wind speed data of national meteorological observation stations, with a spatial resolution of 0.25°x0.25° grid. After data collection, the historical prediction data and corresponding real-time observation data need to be processed according to the preprocessing method of the prediction stage. The data set can be divided into training set, validation set and test set, and then model training, validation and testing are carried out, which will not be repeated here.

[0053] 2) Loss function Combined loss: weighted combination of mean square error (MSE) and mean absolute error (MAE) to balance the sensitivity to large and small errors:

[0054] Wherein, α=0.7, to preferentially focus on the main error direction.

[0055] 3) Optimizer Adam optimizer: learning rate is set to 1e-4, and decay rate is 1e-6 to balance convergence speed and stability.

[0056] Early stopping mechanism: terminate training when validation set loss does not decrease for 5 consecutive epochs to prevent overfitting.

[0057] 4) Training strategy Batch training: batch size is set to 64, and model parameters are updated in batches.

[0058] Data augmentation: simulate spatial heterogeneity by random rotation (0°-360°) and horizontal flip to improve model generalization ability.

[0059] Table 1: Quantitative comparison experiment results of the present application and other methods ; Wherein, the smaller the RMSE (root mean square error) and MAE (mean absolute error), the higher the prediction accuracy. It can be seen that the RMSE and MAE of the present application are better than those of the existing methods in W10 and W100, indicating the advantages of the present application in spatio-temporal modeling and complex terrain adaptability.

[0060] The above description is only the preferred embodiment of the present application, it should be pointed out that other parts not specifically described belong to the prior art or common knowledge for ordinary skilled in the art. Without departing from the principles of the present application, several improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A network fusion-based long-term wind speed forecast revision method, characterized in that, The method comprises the following steps: Collecting multi-source weather data, including short-term forecast data and long-term forecast data; Preprocessing the collected multi-source weather data; Constructing a dynamic spatio-temporal feature matrix according to the preprocessed multi-source weather data, including time features, space features and weather element features; Inputting the time features, space features and weather element features into a trained wind speed correction model respectively, the wind speed correction model comprising a space feature extraction network, a time series feature extraction network and a feature fusion network; the space feature extraction network is used to extract a space feature vector from the space features and weather element features, the time series feature extraction network is used to extract a time series feature vector from the time features and weather element features, and the feature fusion network is used to splice the space feature vector and the time series feature vector to form a joint feature vector, fuse the space features and time series features through a learnable weighting mechanism, and output a wind speed correction value. 2.The network fusion based long-term wind speed forecast revision method according to claim 1, characterized in that, The short-term forecast data is 0-168 hours per hour, and the long-term forecast data is 168-360 hours per 3 hours, both of which include 10-meter gust wind speed, 10-meter wind speed, 100-meter wind speed and auxiliary weather element data, and the auxiliary weather element data includes temperature field, pressure field, humidity field and terrain height. 3.The network fusion based long-term wind speed forecast revision method according to claim 1, wherein, The preprocessing includes data cleaning, spatio-temporal alignment and data standardization in sequence.

4. The network fusion-based long-term wind speed forecast revision method according to claim 3, characterized in that, The data cleaning method is as follows: Using the 3σ criterion to remove abnormal values exceeding ±3 times the standard deviation; Using cubic spline interpolation to fill in missing time series data; The spatio-temporal alignment method is as follows: Aligning the cleaned forecast data according to Coordinated Universal Time, and unifying the time resolution to 1 hour; Performing spatial alignment on the forecast data through bilinear interpolation; The data standardization method is as follows: Standardizing continuous variables in the forecast data by the following method: ； wherein, is a standardized continuous variable, is a continuous variable before standardization, is a mean of a continuous variable, is a standard deviation of a continuous variable; Normalizing discrete variables in the forecast data by the following method: ； wherein, is a normalized discrete variable, is a discrete variable before normalization, is an upper threshold value of the discrete variable, is a lower threshold value of the discrete variable.

5. The long-term wind speed forecast revision method based on network fusion according to claim 1, characterized in that, The time features include forecast time, time period features and historical trend features, the space features include grid point position features, spatial gradient features and terrain normalization features, and the weather element features include temperature field, pressure field and humidity field data after standardization.

6. The long-term wind speed forecast revision method based on network fusion according to claim 1, wherein, The space feature extraction network comprises: A first input layer for receiving input space features and weather element features; A multi-scale convolution structure comprising a first convolution layer, a pooling layer, a second convolution layer and a global pooling layer connected in sequence, wherein the first convolution layer extracts local space features through a convolution kernel and a ReLU activation function, the pooling layer reduces the spatial dimension and retains key features through maximum pooling, the second convolution layer uses the same convolution kernel as the first convolution layer but increases the number of output channels to extract higher-level space features, and the global pooling layer compresses the space features into a one-dimensional vector through global average pooling; A terrain effect modeling unit for introducing dynamic normalization features of terrain height in the convolution structure to enhance the space feature extraction capability of complex areas; The first output layer is configured to output compressed spatial feature vectors, including spatial distribution patterns of meteorological elements at the current time.

7. The long-term wind speed forecast revision method based on network fusion according to claim 1, wherein, The time sequence feature extraction network comprises: The second input layer is configured to receive input time features and meteorological element features. The multi-layer recurrent neural network structure comprises a first layer of recurrent neural network and a second layer of recurrent neural network, wherein the first layer of recurrent neural network comprises a plurality of hidden units and returns a sequence to retain time step information, and the second layer of recurrent neural network comprises fewer hidden units and extracts final time-dependent features, and a non-linear activation function is used to avoid the gradient vanishing problem. The time period encoding module is configured to introduce sine / cosine encoding in the input time features to capture the hourly / daily periodic variation rules. The second output layer is configured to output time sequence feature vectors to capture the evolution rules of meteorological elements over time.

8. The method of claim 1, wherein, The feature fusion network comprises: The dynamic fusion module is configured to fuse the spatial feature vectors output by the spatial feature extraction network and the time sequence feature vectors output by the time sequence feature extraction network through a learnable weighting mechanism to form joint feature vectors. The multi-task mapping layer is configured to map the joint feature vectors to the prediction dimensions of different wind speed elements through a fully connected layer with shared weights. The third output layer is configured to output continuous wind speed correction values using a linear activation function.

9. The long-term wind speed forecast revision method based on network fusion according to claim 1, wherein, The first N layers of the spatial feature extraction network and the time sequence feature extraction network share weights to extract general spatio-temporal features, and the model adaptability of complex areas is enhanced through terrain normalized features.

10. The long-term wind speed forecast revision method based on network fusion according to claim 1, wherein, A plurality of independent prediction branches are arranged after the feature fusion network, each branch is adapted to the prediction requirements of different wind speed elements, and shared feature layers are used to reduce redundant calculations.

Images