A method and system for predicting a gust front based on a physical information residual neural network

Abstract

The application discloses a gust front prediction method and system based on a physical information residual neural network, and belongs to the technical field of radar detection and identification. The method comprises the following steps: establishing actual curves Curve1 and Curve2 of a gust front at two continuous time points before a to-be-predicted time point, establishing a point correspondence relationship to obtain sparse motion vector data corresponding to an optimal matching point pair, interpolating the sparse motion vector data to each point of Curve2 to generate a dense motion vector field, performing exponential attenuation, and performing arc length scaling correction and spline curvature smoothing adjustment on an actual curve of the gust front at the to-be-predicted time point to obtain a physical extrapolation result; inputting Curve1 and Curve2 into a residual neural network to output a residual correction amount; and fusing the physical extrapolation result and the residual correction amount to obtain a gust front prediction curve at the to-be-predicted time point. The application solves the problems of low multi-step extrapolation accuracy, shape distortion and poor small sample adaptability in the existing radar gust front prediction.

Description

Technical Field

[0001] This invention relates to a method and system for predicting gust fronts based on physical information residual neural networks, belonging to the field of radar detection and identification technology. Background Technology

[0002] A gust front is a small-to-medium-scale hazardous weather phenomenon generated by supercell storms, multicell storms, or squall line systems. When a storm develops to maturity, the descending airflow touches the ground and diverges in all directions. This diverging airflow converges with the surrounding warm, moist air, forming a convergence line. When the convergence reaches a certain intensity, it is called a gust front. Its formation mechanism is mainly as follows: the descending airflow of a convective storm is entrained, causing precipitation particles to evaporate or sublimate and cool drastically. This results in the descending airflow within the thunderstorm being much colder than the ambient temperature, forming a cold pool upon impact. The outflow leading edge of this cold pool, which diffuses outwards, is the gust front. On radar echoes, gust fronts often appear as thin, narrow band echoes accompanied by obvious radial velocity convergence characteristics. For the automatic identification of radar gust fronts, existing technologies have made some progress: early methods used template matching for comprehensive scoring, such as MIGFA (Machine Intelligent Gust Front Algorithm); subsequent developments included the bidirectional gradient method combined with radial convergence line identification, the function template correlation method for extracting thin line features, the fusion of movement features and mathematical morphology methods, and regional adaptation improvements to MIGFA; in recent years, deep learning models such as neurofuzzy theory, LBDT (Local Binary with Dual-Template), R2CNN (Rotational Region Convolutional Neural Network), and Mask R-CNN (Mask Region-based Convolutional Neural Network) have also been introduced, further improving the identification accuracy and automation level. In terms of gust front prediction, existing methods mainly rely on radar echo extrapolation, centroid tracking, optical flow analysis, or numerical model assistance to achieve short-term location forecasts.

[0003] However, existing gust front prediction technologies still suffer from the following major shortcomings: First, low accuracy of multi-step extrapolation. Traditional methods such as centroid matching and optical flow methods have large prediction errors when the front is curved, broken, or has a complex shape, and these errors accumulate rapidly with the extension of the prediction lead time. Meanwhile, existing evaluation methods mostly rely on single-point or centroid position error statistics, lacking a systematic, quantitative, and detailed evaluation of the similarity of the predicted gust front curve shape, making it difficult to comprehensively reflect the degree of matching between the predicted curve and the actual evolution curve in terms of shape, direction, and details. Second, shape distortion is common. Pure data-driven methods or simple physical models do not adequately constrain the geometric shape of the gust front curve, leading to unreasonable shape distortions in the predicted curve during its evolution, failing to maintain the smoothness and continuity that the front should have. Third, poor adaptability to small samples. Existing deep learning prediction methods heavily rely on large-scale labeled samples, while typical gust front processes, especially those accompanied by severe disasters, are relatively rare in actual operations, resulting in insufficient model generalization ability and limited support for refined and quantitative predictions under different weather conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a gust front prediction method and system based on physical information residual neural network. By combining physical constraint extrapolation and data-driven residual correction, it solves the problems of low multi-step extrapolation accuracy, easy shape distortion, and poor adaptability to small samples in existing radar gust front prediction.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for predicting gust fronts based on a physical information residual neural network, comprising:

[0007] At a preset elevation angle at the lower level of the radar, the gust front curve data of the two consecutive moments before the time to be predicted are obtained. Based on the gust front curve data, the actual gust front curves of the two consecutive moments before the time to be predicted are established by b-spline interpolation, and are respectively denoted as Curve1 and Curve2.

[0008] A feature correlation matching method based on a sliding window is used to establish the point correspondence between Curve1 and Curve2, and obtain the sparse motion vector data corresponding to the optimal matching point pair.

[0009] Using the radial basis function interpolation method, the sparse motion vector data is interpolated to each point of Curve2 to generate a dense motion vector field;

[0010] The initial prediction curve is obtained by exponentially decaying the dense motion vector field based on the Gompertz decay model. Then, the initial prediction curve is corrected by arc length scaling and spline curvature smoothing based on the arc length change rate and average curvature change rate of the first two consecutive time moments to obtain the physical extrapolation result.

[0011] Input Curve1 and Curve2 into a pre-trained residual neural network and output the residual correction amount;

[0012] The physical extrapolation results are weighted and fused with the residual correction to obtain the gust front prediction curve for the time to be predicted.

[0013] Furthermore, after obtaining the gust front prediction curve for the time to be predicted, it also includes:

[0014] At a preset elevation angle at the lower level of the radar, the gust front curve data for the time to be predicted is obtained in order to establish the actual gust front curve for the time to be predicted.

[0015] The predicted curve of the gust front and the actual curve of the gust front at the time to be predicted are quantitatively evaluated by using two-way average distance, root mean square error and normalized similarity score, so as to verify the accuracy of the prediction curve of the gust front at the time to be predicted.

[0016] Repeat the above steps to output the gust front prediction curve for the future forecast period.

[0017] Furthermore, the radar's lower-level preset elevation angle is 0.5°.

[0018] Furthermore, a feature correlation matching method based on a sliding window is used to establish the point correspondence between Curve1 and Curve2, obtaining the sparse motion vector data corresponding to the optimal matching point pair, including:

[0019] The cumulative arc length is calculated based on the curve coordinate data of Curve1 and Curve2 to obtain the arc length parameterized curve;

[0020] Based on the arc length parameterized curve, the sliding window features are extracted and centered to obtain local feature vectors;

[0021] The search region is determined based on the local feature vectors, and the correlation coefficient between the candidate window and the local feature vectors is calculated to obtain the optimal matching point pair.

[0022] The change in the corresponding position between Curve1 and Curve2 is calculated based on the optimal matching point pair to obtain sparse motion vector data.

[0023] Furthermore, using the radial basis function interpolation method, the sparse motion vector data is interpolated onto each point of Curve2 to generate a dense motion vector field, including:

[0024] The sparse motion vector data is used as input to a radial basis function interpolator to construct an RBF interpolation model.

[0025] The motion vector of each point in Curve2 is calculated using the RBF interpolation model, generating a dense distribution of motion vectors.

[0026] The motion vector distributions corresponding to all points of Curve2 are combined to generate a dense motion vector field.

[0027] Furthermore, based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain an initial prediction curve. Then, based on the arc length change rate and average curvature change rate of the first two consecutive time points, the initial prediction curve is subjected to arc length scaling correction and spline curvature smoothing adjustment to obtain the physical extrapolation result, including:

[0028] The dense motion vector field is exponentially decayed based on the Gompertz decay model to obtain the decayed motion vector.

[0029] The attenuated motion vector is applied to the actual curve of the gust front at the current moment to obtain the initial prediction curve;

[0030] Based on the initial prediction curve and the arc length change rate of the actual curve of the gust front at two consecutive moments before the time to be predicted, the target arc length is calculated and the initial prediction curve is scaled and corrected as a whole to obtain the arc length correction curve.

[0031] Based on the arc length correction curve and the average rate of curvature change of the actual curves of the gust fronts at the two consecutive moments before the time to be predicted, the target curvature is calculated and the initial prediction curve is smoothed using a spline function to obtain the physical extrapolation result.

[0032] Furthermore, the network structure of the residual neural network includes:

[0033] The input layer receives Curve1 and Curve2, and flattens and concatenates the coordinate vectors of Curve1 and Curve2 into a joint input vector;

[0034] Multiple hidden layers are used to extract features and perform nonlinear transformations on the joint input vector to obtain hidden layer features. Each hidden layer is a fully connected layer combined with ReLU activation function and Dropout regularization.

[0035] The output layer is used to generate a residual correction vector based on the features of the hidden layer, wherein the length of the residual correction vector is the same as the length of the Curve2 curve, and outputs the residual correction amount.

[0036] Furthermore, the training method for the residual neural network includes:

[0037] Training samples are extracted from a series of consecutive multi-frame gust front curves using a sliding window method. Each training sample consists of the actual gust front curves Curve1 and Curve2 at the first two consecutive moments as input to the residual neural network, and the actual gust front curve Curve3 at the third consecutive moment as a supervision label.

[0038] Flatten and concatenate Curve1 and Curve2 as input to the residual neural network. Calculate the residual vector through the residual neural network, and add the residual vector to the coordinates of Curve2 to obtain the initial prediction curve for the prediction time.

[0039] Using Curve3 as a supervision signal, the initial prediction curve is trained so that the residual neural network outputs the residual correction amount;

[0040] The mean squared error between the predicted curve and Curve3 is calculated as the data fitting term loss, and the physical fitting term loss is also calculated. The physical losses are then weighted and merged into the total loss function for gradient descent optimization. The physical fitting term loss includes arc length preservation loss and curvature smoothing loss.

[0041] The Adam optimizer is used for iterative training, and the weights of the residual neural network are updated based on the early stopping mechanism until the total loss converges, thus obtaining the trained residual neural network.

[0042] Furthermore, the total loss function is expressed as:

[0043] ;

[0044] In the formula, Represents the total loss function. This represents the loss of the data fitting term. The weighting coefficients represent the loss of the physical fit term. This indicates the loss due to arc length retention. Indicates curvature smoothing loss The weighting coefficients.

[0045] Secondly, the present invention provides a gust front prediction system based on a physical information residual neural network, comprising:

[0046] The data acquisition module is used to acquire the gust front curve data of the two consecutive moments before the time to be predicted at a preset elevation angle at the lower level of the radar. Based on the gust front curve data, the actual gust front curves of the two consecutive moments before the time to be predicted are established by b-spline interpolation, and are respectively denoted as Curve1 and Curve2.

[0047] The feature matching module is used to establish the point correspondence between Curve1 and Curve2 using a sliding window-based feature correlation matching method, and obtain the sparse motion vector data corresponding to the optimal matching point pair.

[0048] The motion vector interpolation module is used to interpolate the sparse motion vector data to each point of Curve2 using the radial basis function interpolation method to generate a dense motion vector field.

[0049] The physical extrapolation module is used to exponentially decay the dense motion vector field based on the Gompertz decay model to obtain an initial prediction curve, and to perform arc length scaling correction and spline curvature smoothing adjustment on the initial prediction curve according to the arc length change rate and average curvature change rate of the first two consecutive time moments to obtain the physical extrapolation result.

[0050] The residual prediction module is used to input Curve1 and Curve2 into a pre-trained residual neural network and output residual corrections.

[0051] The fusion prediction module is used to weight and fuse the physical extrapolation results with the residual correction to obtain the gust front prediction curve for the time to be predicted.

[0052] The iterative control module is used to repeatedly call the above modules and output the gust front prediction curve for the future prediction period.

[0053] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0054] 1. This invention utilizes radial basis function interpolation to interpolate sparse motion vector data onto each point of Curve2 to generate a dense motion vector field. Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain an initial prediction curve. The initial prediction curve is then subjected to arc length scaling correction and spline curvature smoothing adjustment based on the arc length change rate and average curvature change rate of the first two consecutive time points to obtain a physical extrapolation result. The physical extrapolation result is then weighted and fused with the residual correction amount output by the residual neural network to obtain the gust front prediction curve at the time to be predicted. This allows the physical extrapolation result and the residual correction amount to compensate each other at the curve position, achieving coordinated prediction of positional and morphological information. This reduces the error accumulation effect in multi-step recursive prediction and improves the reliability of gust front prediction results.

[0055] 2. This invention utilizes the radial basis function interpolation method to interpolate sparse motion vector data onto each point of Curve2 to generate a dense motion vector field. Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain an initial prediction curve. According to the arc length change rate and the average curvature change rate of the first two consecutive time points, the initial prediction curve is subjected to arc length scaling correction and spline curvature smoothing adjustment to obtain the physical extrapolation result. This allows the physical extrapolation process to simultaneously take into account the exponential decay law of the motion vector and the geometric constraints of the curve arc length and curvature, thus solving the problems of curve length distortion and shape unsmoothness in pure geometric extrapolation.

[0056] 3. This invention constructs an input layer that receives Curve1 and Curve2 and flattens and concatenates the coordinate vectors into a joint input vector. After passing through multiple fully connected hidden layers combined with ReLU activation function and Dropout regularization, the output layer generates a residual correction vector with the same length as Curve2. Curve3 is used as the supervision signal, and the mean square error between the predicted curve and Curve3 is used as the data fitting term loss. At the same time, the arc length preservation loss and curvature smoothing loss are weighted and merged into the total loss function for gradient descent optimization. This allows the residual neural network to learn both the data fitting error and the geometric constraints of the curve during the training process, improving the accuracy of the residual correction amount output by the residual neural network after training and the rationality of the curve shape. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a gust front prediction method based on a physical information residual neural network provided in an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of hybrid extrapolation provided in an embodiment of the present invention, wherein (a) is a schematic diagram of regional feature curve matching, (b) is a schematic diagram of single-step hybrid extrapolation, and (c) is a schematic diagram of multi-step hybrid extrapolation;

[0059] Figure 3 This is a schematic diagram of the hybrid extrapolation distance distribution provided in an embodiment of the present invention, wherein (a) is a schematic diagram of the distance distribution at time t+3, (b) is a schematic diagram of the distance distribution at time t+4, and (c) is a schematic diagram of the distance distribution at time t+5;

[0060] Figure 4 This is a schematic diagram of centroid extrapolation provided in an embodiment of the present invention, wherein (a) is a schematic diagram of the centroid movement trajectory, (b) is a schematic diagram of single-step centroid extrapolation, and (c) is a schematic diagram of multi-step centroid extrapolation;

[0061] Figure 5This is a schematic diagram of the centroid extrapolation distance distribution provided in an embodiment of the present invention, wherein (a) is a schematic diagram of the distance distribution at time t+3, (b) is a schematic diagram of the distance distribution at time t+4, and (c) is a schematic diagram of the distance distribution at time t+5;

[0062] Figure 6 This is a schematic diagram of the actual location of the gust front at time t on the radar image provided in an embodiment of the present invention;

[0063] Figure 7 This is a schematic diagram of the actual location of the gust front on the radar image at time t+1 provided in an embodiment of the present invention;

[0064] Figure 8 This is a schematic diagram of the actual location of the gust front on the radar image at time t+2 provided in an embodiment of the present invention;

[0065] Figure 9 This is a schematic diagram of the actual location of the gust front on the radar image at time t+3 provided in an embodiment of the present invention;

[0066] Figure 10 This is a schematic diagram of the actual location of the gust front on the radar image at time t+4 provided in an embodiment of the present invention;

[0067] Figure 11 This is a schematic diagram of the actual location of the gust front on the radar image at time t+5 provided in an embodiment of the present invention;

[0068] Figure 12 This is a schematic diagram of gust front prediction at time t+3 based on physical information residual neural network provided in an embodiment of the present invention;

[0069] Figure 13 This is a schematic diagram of gust front prediction at time t+4 based on physical information residual neural network provided in an embodiment of the present invention;

[0070] Figure 14 This is a schematic diagram of gust front prediction at time t+5 based on physical information residual neural network provided in an embodiment of the present invention;

[0071] Figure 15 This is a schematic diagram of the prediction based on the centroid gust front at time t+3 provided in an embodiment of the present invention;

[0072] Figure 16 This is a schematic diagram of the prediction based on the centroid gust front at time t+4 provided in an embodiment of the present invention;

[0073] Figure 17 This is a schematic diagram of the prediction based on the centroid gust front at time t+5 provided in an embodiment of the present invention. Detailed Implementation

[0074] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0075] Example 1

[0076] like Figure 1 As shown in the figure, this embodiment introduces a gust front prediction method based on physical information residual neural network, including:

[0077] Step 1: At the preset elevation angle of the lower radar layer, acquire the gust front curve data of the two consecutive moments before the time to be predicted. Based on the gust front curve data, establish the actual gust front curves of the two consecutive moments before the time to be predicted using the b-spline interpolation method, and denot them as Curve1 and Curve2 respectively.

[0078] This embodiment obtains gust front curve data for two consecutive moments before the time to be predicted at a preset elevation angle at the lower level of the radar, and establishes the actual gust front curves Curve1 and Curve2 for the two consecutive moments before the time to be predicted using b-spline interpolation based on the gust front curve data. This can preserve the geometric integrity of the gust front with the actual curve shape of the two consecutive moments.

[0079] Step 2: Using a feature correlation matching method based on a sliding window, establish the point correspondence between Curve1 and Curve2 to obtain the sparse motion vector data corresponding to the optimal matching point pair.

[0080] This embodiment establishes the point correspondence between Curve1 and Curve2 by using a feature correlation matching method based on a sliding window and obtains the sparse motion vector data corresponding to the optimal matching point pair. It can achieve accurate point-by-point association under non-rigid deformation conditions of the curve and avoid matching errors caused by curve stretching or shrinking.

[0081] Step 3: Using the radial basis function interpolation method, the sparse motion vector data is interpolated to each point of Curve2 to generate a dense motion vector field.

[0082] This embodiment uses the radial basis function interpolation method to interpolate the sparse motion vector data onto each point of Curve2 to generate a dense motion vector field. This can smoothly and continuously spread the motion information at the finite matching points to the entire curve, ensuring that each curve point obtains a reasonable motion vector, thereby completely depicting the motion trend of each local area of ​​the gust front.

[0083] Step 4: Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain the initial prediction curve. Then, based on the arc length change rate and the average curvature change rate of the first two consecutive time moments, the initial prediction curve is subjected to arc length scaling correction and spline curvature smoothing adjustment to obtain the physical extrapolation result.

[0084] This embodiment obtains an initial prediction curve by exponentially decaying the dense motion vector field based on the Gompertz decay model. Then, based on the arc length change rate and average curvature change rate of the first two consecutive moments, the initial prediction curve is corrected by arc length scaling and spline curvature smoothing to obtain a physical extrapolation result. This allows the prediction curve to simultaneously conform to the inherent decay law of the gust front and the actual geometric deformation characteristics, significantly suppressing the abnormal expansion or distortion of the curve caused by pure numerical extrapolation.

[0085] Step 5: Input Curve1 and Curve2 into the pre-trained residual neural network and output the residual correction amount.

[0086] This embodiment uses Curve1 and Curve2 as inputs to a pre-trained residual neural network to output residual correction values. It can learn complex nonlinear residual characteristics in the evolution of historical curves that are not described by the physical model in a data-driven manner, thereby compensating for the system bias of physical extrapolation in strongly nonlinear deformation scenarios.

[0087] Step 6: Weight and fuse the physical extrapolation results with the residual correction to obtain the gust front prediction curve for the time to be predicted.

[0088] This embodiment obtains the gust front prediction curve for the time to be predicted by weighted fusion of physical extrapolation results and residual corrections. It can organically combine the determinism of physical laws with the adaptability of neural networks, improving the prediction accuracy of complex gust front evolution processes while ensuring the physical rationality of the prediction curve. By repeatedly executing the above steps to output the gust front prediction curve for the future time period to be predicted, it is possible to achieve step-by-step correction of errors in multi-step rolling prediction, so that the long-term series prediction results maintain a stable and continuous geometric shape.

[0089] Repeat the above steps to output the gust front prediction curve for the future forecast period.

[0090] This embodiment uses the gust front prediction curve output from the previous time step as the historical input for Curve2 in the next prediction round, and sequentially moves the time windows of Curve1 and Curve2 forward, repeating steps one to seven. This allows for the rolling generation of gust front curve sequences for multiple consecutive future time steps without additional manual intervention. Each prediction round inherits the attenuation and deformation constraints of physical extrapolation, as well as the ability of the residual neural network to correct the evolution trend. This effectively suppresses the frame-by-frame accumulation of errors and curve divergence in long-term prediction, and finally outputs a set of gust front prediction curves that maintain geometric continuity, reasonable shape, and smooth motion within the future prediction time period.

[0091] Example 2

[0092] Based on the same inventive concept as Embodiment 1, this embodiment introduces the implementation steps of a gust front prediction method based on a physical information residual neural network, including:

[0093] Step 1: At the preset elevation angle of the lower radar layer, acquire the gust front curve data of the two consecutive moments before the time to be predicted. Based on the gust front curve data, establish the actual gust front curves of the two consecutive moments before the time to be predicted using the b-spline interpolation method, and denot them as Curve1 and Curve2 respectively.

[0094] This embodiment identifies gust front regions based on gust front curve data, and establishes the actual gust front curves for the two consecutive moments before the moment to be predicted using b-spline interpolation based on the gust front regions.

[0095] This embodiment selects the gust front process observed by the KINX radar station, and the radar's lower-level preset elevation angle is 0.5°.

[0096] Step 2: Using a feature correlation matching method based on a sliding window, establish the point correspondence between Curve1 and Curve2 to obtain the sparse motion vector data corresponding to the optimal matching point pair.

[0097] Step 2.1: Calculate the cumulative arc length based on the curve coordinate data of Curve1 and Curve2 to obtain the arc length parameterized curve.

[0098] Step 2.2: Extract the sliding window features based on the arc length parameterized curve and perform centering processing to obtain the local feature vector.

[0099] Step 2.3: Determine the search region based on the local feature vectors, and calculate the correlation coefficient between the candidate window and the local feature vectors to obtain the optimal matching point pair.

[0100] Step 2.4: Determine the motion relationship between Curve1 and Curve2 based on the optimal matching point to obtain sparse motion vector data.

[0101] Step 3: Using the radial basis function interpolation method, the sparse motion vector data is interpolated to each point of Curve2 to generate a dense motion vector field.

[0102] Step 3.1: Use the sparse motion vector data as input to the radial basis function interpolator to construct the RBF interpolation model.

[0103] Step 3.2: Calculate the motion vector of each point in Curve2 using the RBF interpolation model to generate a dense motion vector distribution.

[0104] Step 3.3: Combine the motion vector distributions corresponding to all points of Curve2 to generate a dense motion vector field.

[0105] Step 4: Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain the initial prediction curve. Then, based on the arc length change rate and the average curvature change rate of the first two consecutive time moments, the initial prediction curve is subjected to arc length scaling correction and spline curvature smoothing adjustment to obtain the physical extrapolation result.

[0106] Step 4.1: Perform exponential decay on the dense motion vector field based on the Gompertz decay model to obtain the decayed motion vector.

[0107] Step 4.2: Apply the attenuated motion vector to the actual curve of the gust front at the current moment to obtain the initial prediction curve.

[0108] Step 4.3: Based on the initial prediction curve and the arc length change rate of the actual curve of the gust front at two consecutive moments before the time to be predicted, calculate the target arc length and perform overall scaling correction on the initial prediction curve to obtain the arc length correction curve.

[0109] Step 4.4: Based on the arc length correction curve and the average rate of curvature change of the actual curves of the gust fronts at the two consecutive moments before the time to be predicted, calculate the target curvature and use a spline function to smooth the initial prediction curve to obtain the physical extrapolation result.

[0110] Step 5: Input Curve1 and Curve2 into the pre-trained residual neural network and output the residual correction amount.

[0111] Step 5.1: Construct the residual neural network.

[0112] In this embodiment, the residual neural network structure includes an input layer, multiple hidden layers, and an output layer. The input layer receives Curve1 and Curve2, flattens their coordinate vectors, and concatenates them into a joint input vector. The multiple hidden layers perform feature extraction and nonlinear transformation on the joint input vector to obtain hidden layer features. Each hidden layer is a fully connected layer combined with ReLU activation and Dropout regularization. The output layer generates a residual correction vector based on the hidden layer features. The length of the residual correction vector is the same as the length of Curve2, and outputs the residual correction amount.

[0113] Step 5.2: Train the residual neural network.

[0114] Step 5.2.1: Construct time-series supervised training samples based on the actual curve data of the gust front at three consecutive time points.

[0115] In this embodiment, the actual curve data of the gust front at the three consecutive moments include the actual curve data of the gust front at the first consecutive moment, the actual curve data of the gust front at the second consecutive moment, and the actual curve data of the gust front at the third consecutive moment.

[0116] Step 5.2.2: Based on the actual gust front curve data at the first and second consecutive time points, establish the actual gust front curves at the first and second consecutive time points respectively using the b-spline interpolation method.

[0117] Step 5.2.3: Based on the actual gust front curve data at the third consecutive time point, establish the actual gust front curves at the third consecutive time point using the b-spline interpolation method.

[0118] Step 5.2.4: In the time-series supervised training samples, the actual curves of the gust front at the first and second consecutive time moments are used as input data for the residual neural network, and the actual curve data of the gust front at the third consecutive time moment are used as supervision labels.

[0119] Step 5.2.5: Flatten and splice the actual curves of the gust front at the first and second consecutive time points as input to the residual neural network. Calculate the residual vector through the residual neural network and add the residual vector to the coordinates of the actual curves of the gust front at the second consecutive time point to obtain the initial prediction curve at the prediction time.

[0120] Step 5.2.6: Use the actual curve of the gust front at the third consecutive time step as a supervision signal to train the initial prediction curve so that the residual neural network outputs the residual correction amount.

[0121] Step 5.2.7: Calculate the mean square error between the predicted curve and the actual curve of the gust front at the third consecutive time as the data fitting term loss, and at the same time calculate the physical fitting term loss, and then weight and merge the physical losses into the total loss function for gradient descent optimization.

[0122] In this embodiment, the physical fitting term loss includes arc length preservation loss and curvature smoothing loss. The total loss function is expressed as:

[0123] ;

[0124] In the formula, Represents the total loss function. This represents the loss of the data fitting term. The weighting coefficients representing the loss of the physical fit term. This indicates the loss due to arc length retention. Indicates curvature smoothing loss The weighting coefficients.

[0125] Step 5.2.8: Use the Adam optimizer for iterative training and update the weights of the residual neural network based on the early stopping mechanism until the total loss converges to obtain the trained residual neural network.

[0126] Step 6: Weight and fuse the physical extrapolation results with the residual correction to obtain the gust front prediction curve for the time to be predicted.

[0127] Repeat the above steps to output the gust front prediction curve for the future forecast period.

[0128] Step 7: Output the prediction accuracy of the gust front prediction curve at the time to be predicted.

[0129] Step 7.1: At the preset elevation angle at the lower level of the radar, acquire the gust front curve data for the time to be predicted, so as to establish the actual gust front curve for the time to be predicted.

[0130] Step 7.2: Use bidirectional average distance, root mean square error and normalized similarity score to quantitatively evaluate the predicted curve and actual curve of the gust front at the time to be predicted, so as to verify the accuracy of the prediction curve of the gust front at the time to be predicted.

[0131] Step 8: Comparative experiment.

[0132] To verify the effectiveness of this embodiment, a comparative experiment was conducted with the traditional gust front extrapolation method based solely on centroid displacement. Experimental results show that this embodiment significantly outperforms the baseline method relying solely on centroid translation in terms of single-step and multi-step prediction accuracy, as well as curve shape preservation capability, fully demonstrating the effectiveness of the hybrid strategy integrating physical constraints and residual neural networks. The hybrid extrapolation prediction evaluation table and the centroid extrapolation prediction evaluation table obtained through the comparative experiment are shown in Table 1 and Table 2, respectively:

[0133] Table 1: Evaluation Table of Hybrid Extrapolation Prediction

[0134] Table 2: Centroid Extrapolation Prediction and Evaluation Table

[0135] As shown in Tables 1 and 2, the gust front prediction method based on physical information residual neural networks proposed in this invention outperforms the traditional centroid extrapolation method in terms of evaluation metrics such as average distance, RMSE, and curve similarity. Specifically, the average distance is reduced by more than 64%, the RMSE by more than 17%, and the curve similarity is improved by approximately 30%–970%. Particularly noteworthy is that at time t+5, the similarity remains at 0.7514, while the centroid extrapolation decreases to 0.0699, indicating that this invention can effectively characterize the motion evolution and morphological changes of gust fronts, significantly improving the accuracy and stability of gust front nowcasting.

[0136] Figure 2 This is a schematic diagram of hybrid extrapolation provided in an embodiment of the present invention. Figure 2 From (a) we can see the point pair relationship of the curve matching between the two time points; from Figure 2 From (b) we can see the difference between the single-step extrapolation curve and the actual curve; from Figure 2 As shown in (c), the multi-step extrapolation curve gradually approximates the actual gust front position. Therefore, this invention can effectively establish the correspondence between gust front curves and, combined with the evolution characteristics of gust front arc length and curvature, better characterize the motion evolution process of gust fronts, achieving coordinated prediction of gust front position and morphology.

[0137] Figure 3 This is a schematic diagram of the hybrid extrapolation distance distribution provided in an embodiment of the present invention. Figure 3 From (a), we can see the distribution of distances between the extrapolated points and the actual curve; from Figure 3 From (b) we can see the trend of extrapolation error over time; Figure 3 (c) shows the error accumulation during multi-step extrapolation. Therefore, it can be seen that the present invention can maintain a small prediction error and effectively suppress error accumulation during multi-step prediction.

[0138] Figure 4 This is a schematic diagram of centroid extrapolation provided in an embodiment of the present invention. Figure 4 From (a), we can know the direction and speed of the gust front's center of mass movement; from Figure 4 From (b) we can see the matching between the single-step extrapolation prediction and the actual curve; from Figure 4 As shown in (c), the multi-step extrapolation curve approximates the actual gust front position. This indicates that the traditional centroid extrapolation method mainly relies on overall translation for prediction, making it difficult to characterize the local morphological changes of the gust front.

[0139] Figure 5 This is a schematic diagram of the centroid extrapolation distance distribution provided in an embodiment of the present invention. Figure 5 From (a) we can see the distribution of the centroid extrapolation error; from Figure 5 From (b) we can see the trend of error change over time; from Figure 5 (c) shows the cumulative error of multi-step centroid extrapolation. This indicates that the error of the traditional centroid extrapolation method gradually increases with the increase of prediction lead time, resulting in poor prediction stability.

[0140] Figure 6 – Figure 11 The following are schematic diagrams of the actual positions of the gust front on the radar image from time t to t+5. Figure 6 – Figure 11 This reveals the actual spatial positional variation of the gust front at each moment. Therefore, it shows that the traditional centroid extrapolation method exhibits increasing error and poor prediction stability as the prediction lead time increases.

[0141] Figure 12 – Figure 14 The diagrams shown are, in order, schematics of gust front predictions at times t+3 to t+5 based on a physical information residual neural network. Figure 12 – Figure 14 The results show the fitting effect and error trend of the residual neural network prediction curve on the actual curve. This indicates that the gust front not only undergoes overall displacement during its evolution but also exhibits significant morphological changes.

[0142] Figure 15 – Figure 17 The following are schematic diagrams of predictions based on the centroid gust front from time t+3 to t+5. Figure 15 – Figure 17 It can be seen that the centroid extrapolation prediction curve approximates the actual gust front position and varies in error. Therefore, compared to the method of this invention, the traditional centroid extrapolation method is more prone to positional shifts and morphological distortions in long-term predictions.

[0143] Example 3

[0144] Based on the same inventive concept as other embodiments, this embodiment introduces a gust front prediction system based on a physical information residual neural network, comprising:

[0145] The data acquisition module is used to acquire the gust front curve data of the two consecutive moments before the time to be predicted at a preset elevation angle at the lower level of the radar. Based on the gust front curve data, the actual gust front curves of the two consecutive moments before the time to be predicted are established by b-spline interpolation, and are respectively denoted as Curve1 and Curve2.

[0146] The feature matching module is used to establish the point correspondence between Curve1 and Curve2 using a sliding window-based feature correlation matching method, and obtain the sparse motion vector data corresponding to the optimal matching point pair.

[0147] The motion vector interpolation module is used to interpolate the sparse motion vector data to each point of Curve2 using the radial basis function interpolation method to generate a dense motion vector field.

[0148] The physical extrapolation module is used to exponentially decay the dense motion vector field based on the Gompertz decay model to obtain an initial prediction curve, and to perform arc length scaling correction and spline curvature smoothing adjustment on the initial prediction curve according to the arc length change rate and average curvature change rate of the first two consecutive time moments to obtain the physical extrapolation result.

[0149] The residual prediction module is used to input Curve1 and Curve2 into a pre-trained residual neural network and output residual corrections.

[0150] The fusion prediction module is used to weight and fuse the physical extrapolation results with the residual correction to obtain the gust front prediction curve for the time to be predicted.

[0151] The iterative control module is used to repeatedly call the above modules and output the gust front prediction curve for the future prediction period.

[0152] The specific functions of each module described above are explained in the relevant content of Embodiment 1 or 2, and will not be repeated here.

[0153] In summary, this invention utilizes radial basis function interpolation to interpolate sparse motion vector data onto each point of Curve2 to generate a dense motion vector field. Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain an initial prediction curve. The initial prediction curve is then subjected to arc length scaling correction and spline curvature smoothing adjustment based on the arc length change rate and average curvature change rate of the first two consecutive time points to obtain a physical extrapolation result. The physical extrapolation result is then weighted and fused with the residual correction amount output by the residual neural network to obtain the gust front prediction curve at the time to be predicted. This allows the physical extrapolation result and the residual correction amount to compensate each other at the curve position, achieving coordinated prediction of positional and morphological information. This reduces the error accumulation effect in multi-step recursive prediction and improves the reliability of gust front prediction results.

[0154] This invention utilizes the radial basis function interpolation method to interpolate sparse motion vector data onto each point of Curve2 to generate a dense motion vector field. Based on the Gompertz decay model, the dense motion vector field is exponentially decayed to obtain an initial prediction curve. According to the arc length change rate and the average curvature change rate of the first two consecutive time points, the initial prediction curve is subjected to arc length scaling correction and spline curvature smoothing adjustment to obtain the physical extrapolation result. This allows the physical extrapolation process to simultaneously take into account the exponential decay law of motion vectors and the geometric constraints of curve arc length and curvature, thus solving the problems of curve length distortion and shape unsmoothness in pure geometric extrapolation.

[0155] This invention constructs an input layer that receives Curve1 and Curve2 and flattens and concatenates the coordinate vectors into a joint input vector. After passing through multiple fully connected hidden layers combined with ReLU activation function and Dropout regularization, the output layer generates a residual correction vector with the same length as Curve2. Curve3 is used as the supervision signal, and the mean square error between the predicted curve and Curve3 is used as the data fitting term loss. At the same time, the arc length preservation loss and curvature smoothing loss are weighted and merged into the total loss function for gradient descent optimization. This allows the residual neural network to learn both the data fitting error and the geometric constraints of the curve during the training process, improving the accuracy of the residual correction amount output by the residual neural network after training and the rationality of the curve shape.

[0156] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0157] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0158] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting gust fronts based on physical information residual neural networks, characterized in that, include: At a preset elevation angle at the lower level of the radar, the gust front curve data of the two consecutive moments before the time to be predicted are obtained. Based on the gust front curve data, the actual gust front curves of the two consecutive moments before the time to be predicted are established by b-spline interpolation, and are respectively denoted as Curve1 and Curve2. A feature correlation matching method based on a sliding window is used to establish the point correspondence between Curve1 and Curve2, and obtain the sparse motion vector data corresponding to the optimal matching point pair, including: The cumulative arc length is calculated based on the curve coordinate data of Curve1 and Curve2 to obtain the arc length parameterized curve; Based on the arc length parameterized curve, the sliding window features are extracted and centered to obtain local feature vectors; The search region is determined based on the local feature vectors, and the correlation coefficient between the candidate window and the local feature vectors is calculated to obtain the optimal matching point pair. Based on the optimal matching point pair, the change in the corresponding position between Curve1 and Curve2 is calculated to obtain sparse motion vector data; Using the radial basis function interpolation method, the sparse motion vector data is interpolated onto each point of Curve2 to generate a dense motion vector field, including: The sparse motion vector data is used as input to a radial basis function interpolator to construct an RBF interpolation model. The motion vector of each point in Curve2 is calculated using the RBF interpolation model, generating a dense distribution of motion vectors. Combine the motion vector distributions corresponding to all points of Curve2 to generate a dense motion vector field; The initial prediction curve is obtained by exponentially decaying the dense motion vector field based on the Gompertz decay model. Then, based on the arc length change rate and average curvature change rate of the first two consecutive time points, the initial prediction curve is adjusted by arc length scaling and spline curvature smoothing to obtain the physical extrapolation result, including: The dense motion vector field is exponentially decayed based on the Gompertz decay model to obtain the decayed motion vector. The attenuated motion vector is applied to the actual curve of the gust front at the current moment to obtain the initial prediction curve; Based on the initial prediction curve and the arc length change rate of the actual curve of the gust front at two consecutive moments before the time to be predicted, the target arc length is calculated and the initial prediction curve is scaled and corrected as a whole to obtain the arc length correction curve. Based on the arc length correction curve and the average rate of curvature change of the actual curve of the gust front at the two consecutive moments before the time to be predicted, the target curvature is calculated and the initial prediction curve is smoothed using a spline function to obtain the physical extrapolation result. Input Curve1 and Curve2 into a pre-trained residual neural network and output the residual correction amount; The physical extrapolation results are weighted and fused with the residual correction to obtain the gust front prediction curve for the time to be predicted. Repeat the above steps to output the gust front prediction curve for the future time period to be predicted. The training method for the residual neural network includes: Training samples are extracted from a series of consecutive multi-frame gust front curves using a sliding window method. Each training sample consists of the actual gust front curves Curve1 and Curve2 at the first two consecutive moments as input to the residual neural network, and the actual gust front curve Curve3 at the third consecutive moment as a supervision label. Flatten and concatenate Curve1 and Curve2 as input to the residual neural network. Calculate the residual vector through the residual neural network, and add the residual vector to the coordinates of Curve2 to obtain the initial prediction curve for the prediction time. Using Curve3 as a supervision signal, the initial prediction curve is trained so that the residual neural network outputs the residual correction amount; The mean squared error between the predicted curve and Curve3 is calculated as the data fitting term loss, and the physical fitting term loss is also calculated. The physical losses are then weighted and merged into the total loss function for gradient descent optimization. The physical fitting term loss includes arc length preservation loss and curvature smoothing loss. The Adam optimizer is used for iterative training, and the weights of the residual neural network are updated based on the early stopping mechanism until the total loss converges, thus obtaining the trained residual neural network.

2. The gust front prediction method based on physical information residual neural network according to claim 1, characterized in that, After obtaining the gust front prediction curve for the time to be predicted, the following is also included: At a preset elevation angle at the lower level of the radar, the gust front curve data for the time to be predicted is obtained in order to establish the actual gust front curve for the time to be predicted. The predicted curves of the gust front and the actual curves of the gust front at the time to be predicted are quantitatively evaluated by using two-way average distance, root mean square error, and normalized similarity score, so as to verify the accuracy of the prediction of the gust front curve at the time to be predicted.

3. The gust front prediction method based on physical information residual neural network according to claim 1, characterized in that, The radar's lower-level preset elevation angle is 0.5°.

4. The gust front prediction method based on physical information residual neural network according to claim 1, characterized in that, The total loss function is expressed as: ； In the formula, Represents the total loss function. This represents the loss of the data fitting term. The weighting coefficients represent the loss of the physical fit term. This indicates the loss due to arc length retention. Indicates curvature smoothing loss The weighting coefficients.

5. A gust front prediction system based on a physical information residual neural network, characterized in that, The method for predicting gust fronts based on physical information residual neural networks as described in any one of claims 1 to 4 includes: The data acquisition module is used to acquire the gust front curve data of the two consecutive moments before the time to be predicted at a preset elevation angle at the lower level of the radar. Based on the gust front curve data, the actual gust front curves of the two consecutive moments before the time to be predicted are established by b-spline interpolation, and are respectively denoted as Curve1 and Curve2. The feature matching module is used to establish the point correspondence between Curve1 and Curve2 using a sliding window-based feature correlation matching method, and obtain the sparse motion vector data corresponding to the optimal matching point pair. The motion vector interpolation module is used to interpolate the sparse motion vector data to each point of Curve2 using the radial basis function interpolation method to generate a dense motion vector field. The physical extrapolation module is used to exponentially decay the dense motion vector field based on the Gompertz decay model to obtain an initial prediction curve, and to perform arc length scaling correction and spline curvature smoothing adjustment on the initial prediction curve according to the arc length change rate and average curvature change rate of the first two consecutive time moments to obtain the physical extrapolation result. The residual prediction module is used to input Curve1 and Curve2 into a pre-trained residual neural network and output residual corrections. The fusion prediction module is used to weight and fuse the physical extrapolation results with the residual correction to obtain the gust front prediction curve for the time to be predicted. The iterative control module is used to repeatedly call the above modules and output the gust front prediction curve for the future prediction period.

Images