A sub-seasonal extreme precipitation prediction method based on deep fusion of precursor factors
By constructing a deep learning model of spatiotemporal transformation network, integrating numerical models and historical observation data, and combining observation precursor factors, the problem of single dependence in the existing technology is solved, significantly improving the accuracy and comprehensiveness of prediction of extreme precipitation in the second season.
Patent Information
- Application Number
- CN202411305389.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-09-19
AI Technical Summary
The existing subseason extreme precipitation prediction methods rely on a single numerical mode output or a single observation precursor factor, and fail to fully integrate potential correlations and spatiotemporal characteristics in multi-source information, resulting in limited prediction accuracy and comprehensiveness.
Build a deep learning model of spatiotemporal transformation network based on optimized loss function, integrate numerical mode precipitation prediction, historical precipitation observation and observation precursor factor data, extract spatiotemporal features through time sequence encoding, spatial encoding, multi-head self-attention mechanism and feedforward neural network to perform The forecast of extreme precipitation in the next season.
The accuracy and reliability of prediction of extreme precipitation in the second season are significantly improved. By deeply fusion of multi-source information, complex space-time dependencies are captured, and the generalization ability and practical application effect of the model are enhanced.
Smart Images

Figure CN118822048B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of sub-seasonal extreme precipitation prediction, and in particular to a sub-seasonal extreme precipitation prediction method with deep integration of precursor factors. Background Art
[0002] Effective prediction of sub-seasonal extreme precipitation can provide important support for disaster prevention and mitigation, and reduce the negative impact of extreme weather. In fact, changes in some precursor factors (such as sea temperature, sea ice, and circulation during the same period and previous periods) often significantly affect the occurrence of extreme precipitation events by affecting atmospheric movement, water vapor transport, and local climate conditions. At the same time, they can also provide some effective information for the accurate prediction of extreme precipitation.
[0003] The current prediction of sub-seasonal extreme precipitation still faces many technical difficulties. For example, traditional sub-seasonal prediction methods often rely on a single numerical model output or a single observational precursor factor statistic, but fail to fully integrate and utilize such effective information, and are often unable to fully explore and utilize the potential correlations and spatiotemporal characteristics in multi-source information, resulting in limitations on the accuracy and comprehensiveness of the prediction. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a sub-seasonal extreme precipitation prediction method with deep integration of precursor factors that can significantly improve the accuracy and comprehensiveness of sub-seasonal extreme precipitation prediction.
[0005] Technical solution: The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors described in the present invention comprises the following steps:
[0006] Based on the optimized loss function, a spatiotemporal transformation network deep learning model integrating numerical model precipitation prediction, historical precipitation observation and observed precursor factors is constructed. Model training and optimization are carried out based on numerical model precipitation prediction dataset, historical precipitation observation dataset and observed precursor factor dataset. Using the trained model, the numerical model precipitation prediction and precursor factor data of the target time are input for prediction, and the sub-seasonal extreme precipitation prediction results are output.
[0007] Among them, the spatiotemporal transformation network deep learning model includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module and an output module; the input module obtains the numerical model precipitation prediction data set, the historical precipitation observation data set and the observation precursor factor data set, and splices them into the model input feature matrix; the spatiotemporal encoding module extracts the spatiotemporal features of the input feature matrix, wherein the temporal encoding embeds the time information through the position encoding method to capture the temporal relationship of the input feature matrix, and the spatial encoding further extracts the spatial relationship between the input feature matrices of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module is used to gradually restore the spatial resolution of the data to obtain the spatiotemporal decoded feature matrix; the output module converts the generated spatiotemporal decoded feature matrix into prediction features.
[0008] Furthermore, the method for producing the numerical model precipitation prediction dataset and the historical precipitation observation dataset is as follows:
[0009] First, the numerical model precipitation prediction data and historical precipitation observation data are denoised and then standardized preprocessed. The formula is:
[0010] ,
[0011] ,
[0012] in, and are the denoised numerical model precipitation prediction data and historical precipitation observation data matrices, and are the means of the denoised numerical model precipitation prediction data and the historical precipitation observation data, and are the standard deviations of the denoised numerical model precipitation prediction data and historical precipitation observation data, and They are the standardized numerical model precipitation prediction data and historical precipitation observation data matrices respectively.
[0013] Furthermore, the method for producing the observation precursor factor data set is: adaptively annotating and encoding the observation precursor factor data to generate a feature vector in a unified format, specifically:
[0014] The observed precursor factor data include the sea temperature, sea ice, and circulation during the same period and before the reporting time. The adaptive annotation coding preprocessing calculation method is:
[0015] ,
[0016] ,
[0017] ,
[0018] Among them, Embedding is the annotation coding embedding function, SST , SeaIce and Circulation are the sea temperature, sea ice and circulation data matrices for the same period and previous period of the reporting time, , and They are the sea temperature, sea ice and circulation data matrices after adaptive annotation and encoding;
[0019] The normalization preprocessing calculation method is:
[0020] ,
[0021] ,
[0022] ,
[0023] in, , and are the minimum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the maximum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling encoding and normalization preprocessing respectively.
[0024] Furthermore, the optimized loss function is:
[0025] ,
[0026] ,
[0027] ,
[0028] ,
[0029] in, Loss is the optimized loss function, is the weight of the mean square error loss term, is the weight of the critical success index loss term, is the weight of the structural similarity index loss term, MSE is the mean square error, CSI is the critical success index, SSIM is the structural similarity index, N is the evaluation sample size in model training, and Respectively The observed and predicted values of samples, TP is the actual number of samples, FP is the number of false positive samples, FN is the number of false negative samples, and They are N Group and The average value of and They are N Group and The variance of for N Group and The covariance between and is a stabilizing constant used to avoid the denominator being zero.
[0030] Furthermore, the processing of the spatiotemporal coding module is as follows:
[0031] ,
[0032] ,
[0033] ,
[0034] ,
[0035] ,
[0036] in, Indicates at location The first The position encoding value in the dimension is calculated using the sine function; is the time step, indicating the position of the current data point in the time series; is the feature dimension, which indicates the index of the current feature in the feature matrix; is the total dimension of the input feature matrix; Indicates at location The first The position encoding value in the dimension is calculated using the cosine function; PERepresents the time position encoding matrix, where each row corresponds to a time step and each column corresponds to a feature dimension; Describe the input feature matrix after temporal encoding, represents the model input feature matrix, , and are the standardized numerical model precipitation prediction data and historical precipitation observation data matrices, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling coding and normalization preprocessing; represents the input feature matrix after spatiotemporal encoding; ReLU represents the activation function, is the convolution kernel weight, represents the convolution operation, is the bias term.
[0037] Furthermore, the multi-head self-attention mechanism module includes multiple self-attention heads, which focus on the input feature matrix after spatiotemporal encoding. , through multiple self-attention heads Self-attention is calculated in parallel to calculate the correlation between input features and capture the dependencies between different time and space points; its specific form is:
[0038] ,
[0039] ,
[0040] ,
[0041] in, represents the feature matrix of the multi-head self-attention mechanism, concat represents the concatenation function, Indicates A self-attention head, , is the number of self-attention heads, is the output weight matrix, , and denote the query matrix, key matrix and value matrix respectively, , , , , and Respectively The weight matrix of the query matrix, key matrix and value matrix in the self-attention head, Attention represents the self-attention weight mechanism, and softmax is the activation function. For each head dimension size,d is the total dimension of the input feature matrix, T Represents the transpose of a matrix.
[0042] Furthermore, the feedforward neural network module includes two linear transformation layers and an activation function, which is used to calculate the multi-head self-attention mechanism feature matrix Perform nonlinear transformation to enhance the model's ability to express complex features; its specific form is:
[0043] ,
[0044] ,
[0045] ,
[0046] in, represents the intermediate feature, represents the feature matrix of the multi-head self-attention mechanism, and are the weight matrix and bias term of the first layer linear transformation respectively; ReLU represents the activation function, Represents the activated features; represents the feedforward neural network feature matrix, and are the weight matrix and bias term of the second layer linear transformation respectively.
[0047] Furthermore, the spatiotemporal decoding module includes a deconvolution layer, which is used to transform the feature matrix of the feedforward neural network Upsample to restore the original spatial resolution; its specific form is:
[0048] ,
[0049] in, represents the feature matrix after spatiotemporal decoding, ReLU represents the activation function, and are the deconvolution kernel weight and bias term respectively, Represents a deconvolution operation.
[0050] Furthermore, the output module includes a fully connected layer to map the high-dimensional features to the target output space and transform the feature matrix after spatiotemporal decoding into Converted into prediction features; its specific form is:
[0051] ,
[0052] in, is the precipitation prediction feature output by the model, and are the weight matrix and bias term of the fully connected layer respectively.
[0053] The system corresponding to the above method includes:
[0054] The model building and training unit is used to build a spatiotemporal transformation network deep learning model that integrates numerical model precipitation prediction, historical precipitation observation and observation precursor factors based on an optimized loss function, and to carry out model training and optimization based on the numerical model precipitation prediction data set, historical precipitation observation data set and observation precursor factor data set;
[0055] The prediction unit is used to use the trained model, input the numerical model precipitation prediction and precursor factor data of the target time for prediction, and output the sub-seasonal extreme precipitation prediction results;
[0056] Among them, the spatiotemporal transformation network deep learning model includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module and an output module; the input module obtains the numerical model precipitation prediction data set, the historical precipitation observation data set and the observation precursor factor data set, and splices them into the model input feature matrix; the spatiotemporal encoding module extracts the spatiotemporal features of the input feature matrix, wherein the temporal encoding embeds the time information through the position encoding method to capture the temporal relationship of the input feature matrix, and the spatial encoding further extracts the spatial relationship between the input feature matrices of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module is used to gradually restore the spatial resolution of the data to obtain the spatiotemporal decoded feature matrix; the output module converts the generated spatiotemporal decoded feature matrix into prediction features.
[0057] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are: (1) by deeply integrating numerical model precipitation forecasts and historical precipitation observation data, as well as observed precursor factors (such as sea temperature, sea ice, and circulation at the same time and before the onset time), the potential correlation and spatiotemporal characteristics in these data are fully utilized; the multi-source information can be comprehensively processed through the spatiotemporal transformation network deep learning model, thereby significantly improving the accuracy and reliability of sub-seasonal extreme precipitation prediction; (2) the constructed spatiotemporal transformation network deep learning model extracts the spatiotemporal characteristics of the input data through temporal coding and spatial coding, and uses a multi-head self-attention mechanism and a feedforward neural network to capture complex spatiotemporal dependencies. The spatiotemporal coding module of the model can effectively capture the temporal and spatial relationships in the data. The multi-head self-attention mechanism further enhances the model's ability to capture long-term series and large-spatial dependencies, while the feedforward neural network improves the model's ability to express complex features. The model has obvious advantages in feature extraction and representation, thereby improving the prediction effect; (3) In the model training process, an optimized loss function is adopted, which comprehensively considers the mean square error, critical success index and structural similarity index; this loss function design can not only minimize the difference between the predicted value and the observed value, but also optimize the model's prediction ability for extreme events and maintain the structural similarity of the prediction results; by adjusting the weights, the contribution of each error term is balanced, making the model training process more stable and effective, thereby improving the generalization ability of the model and the actual application effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 is a flow chart of the method of the present invention;
[0059] Figure 2 Schematic diagram of adaptive annotation coding and normalization preprocessing for observed precursor factors;
[0060] Figure 3 This is a schematic diagram of the deep learning model of the space-time transformation network;
[0061] Figure 4 Schematic diagram of model training. DETAILED DESCRIPTION
[0062] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0063] The present invention provides a method for predicting sub-seasonal extreme precipitation with deep integration of precursor factors, the process is as follows: Figure 1 As shown, the specific steps include:
[0064] S1. Obtain the numerical model precipitation prediction data and historical precipitation observation data, first perform denoising, that is, remove obvious outliers or incorrect data in the data, and then perform standardization preprocessing; the calculation method is:
[0065] (1),
[0066] (2),
[0067] in, and are the denoised numerical model precipitation prediction data and historical precipitation observation data matrices, and are the means of the denoised numerical model precipitation prediction data and the historical precipitation observation data, and are the standard deviations of the denoised numerical model precipitation prediction data and historical precipitation observation data, and They are the standardized numerical model precipitation prediction data and historical precipitation observation data matrices respectively.
[0068] Subsequently, the standardized numerical model precipitation prediction data and historical precipitation observation data matrix are divided into training set and validation set according to a certain ratio.
[0069] S2, obtain the observed precursor factor data (such as sea temperature, sea ice, and circulation at the same time and before the onset time), perform adaptive labeling coding and normalization preprocessing, and divide the precursor factor data set into a training set and a validation set according to the data set division scheme in step S1; Figure 2 As shown, the specific steps are as follows:
[0070] S21. Obtain the observed precursor factor data, including the sea temperature, sea ice, and circulation at the same time as and before the reporting time, perform adaptive labeling and coding preprocessing, encode the data from different sources, and generate a feature vector in a unified format, that is, each category of data is converted into a real vector of fixed size through an embedding function; the calculation method is:
[0071] (3),
[0072] (4),
[0073] (5),
[0074] Among them, Embedding is the annotation coding embedding function, SST , SeaIce and Circulation are the sea temperature, sea ice and circulation data matrices for the same period and previous period of the reporting time, , and They are the sea temperature, sea ice and circulation data matrices after adaptive annotation and encoding.
[0075] S22. Perform normalization preprocessing on the sea temperature, sea ice and circulation data matrix after adaptive labeling and encoding, and convert data of different dimensions into the same scale range; the calculation method is:
[0076] (6),
[0077] (7),
[0078] (8),
[0079] in, , and are the minimum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the maximum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling encoding and normalization preprocessing respectively.
[0080] S23. Corresponding to the data set division scheme in step S1, the adaptive annotation coding obtained in S22 and the sea temperature, sea ice and circulation data matrix after normalization preprocessing are also divided into a training set and a validation set.
[0081] S3. Based on the optimized loss function, a space-time transformation network deep learning model is constructed that integrates numerical model precipitation prediction, historical precipitation observation and observation precursor factors. The space-time characteristics of the input data are extracted through temporal coding and spatial coding, and the multi-head self-attention mechanism and feedforward neural network are used to capture the complex space-time dependencies. The generated space-time characteristics are then converted into prediction results through the decoder, and the model training and optimization are carried out using the training set and validation set. Among them, the space-time transformation network deep learning model mainly includes an input module, a space-time coding module, a multi-head self-attention mechanism module, a feedforward neural network module, a space-time decoding module and an output module. The input module obtains and distributes the data features of the numerical model precipitation prediction, historical precipitation observation and observation precursor factors obtained by S1 and S2, and integrates them into the model input feature matrix. The coding module extracts the spatiotemporal features of the input feature matrix, where the temporal coding embeds time information through the position coding method to capture the temporal relationship of the input feature matrix, and the spatial coding further extracts the spatial relationship between the data in the input feature matrix of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module includes a deconvolution layer, which is used to gradually restore the spatial resolution of the data output by the feedforward neural network module to obtain the spatiotemporal decoded feature matrix; the output module includes a fully connected layer, which converts the generated spatiotemporal decoded feature matrix into prediction features; the specific steps are as follows:
[0082] S31. Establish an optimized loss function, taking into account the mean square error ( MSE ), Critical Success Index ( CSI ) and the structural similarity index ( SSIM ), its specific form is:
[0083] (9),
[0084] (10),
[0085] (11),
[0086] (12),
[0087] in, Loss is the optimized loss function, is the weight of the mean square error loss term, is the weight of the critical success index loss term, is the weight of the structural similarity index loss term, N is the evaluation sample size in model training, and Respectively The observed and predicted values of samples, TP is the actual number of samples, FP is the number of false positive samples, FP is the number of false negative samples, and They are N Group and The average value of and They are N Group and The variance of for N Group and The covariance between and is a stabilizing constant used to avoid the denominator being zero.
[0088] The loss function design can not only minimize the difference between the predicted value and the observed value, but also optimize the model's ability to predict extreme events and maintain the structural similarity of the prediction results. By adjusting the weights and balancing the contribution of each error term, the model training process is made more stable and effective, thereby improving the model's generalization ability and practical application effect.
[0089] S32, using the loss function obtained in S31, construct a spatiotemporal transformation network deep learning model that integrates numerical model precipitation prediction, historical precipitation observation, and observed precursor factors. The model mainly includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module, and an output module; Figure 3 As shown, its specific form is:
[0090] Input module: The model reads the standardized numerical model precipitation forecast data and historical precipitation observation data obtained by S1 and S2, as well as the sea temperature, sea ice and circulation data after adaptive labeling and normalization preprocessing, and integrates them into the model input feature matrix X ; Its specific form is:
[0091] (13),
[0092] Spatiotemporal coding module: It includes a temporal coding layer and a spatial coding layer to extract the input feature matrix X The temporal and spatial features of the input feature matrix are captured by embedding temporal information through the position encoding method. XThe spatial encoding further extracts the input feature matrix of different regions through a convolutional neural network including a ReLU activation function. X The spatial relationship between them; its specific form is:
[0093] (14),
[0094] (15),
[0095] (16),
[0096] (17),
[0097] (18),
[0098] in, Indicates at location The first The position encoding value in the dimension is calculated using the sine function; is the time step, indicating the position of the current data point in the time series; is the feature dimension, which represents the index of the feature of the current data point in the feature matrix; is the total dimension of the input feature matrix; Indicates at location The first The position encoding value on the dimension is calculated using the cosine function. Finally, the time position encoding matrix is formed PE , each row corresponds to a time step, and each column corresponds to a feature dimension. Through the combination of sine and cosine functions, the encoding values of different positions and dimensions are different, so that the model can effectively distinguish the different time position features in the input feature matrix. PE Add input feature matrix X , forming the input feature matrix after temporal encoding , that is, embedding time information into the input feature matrix to help the model capture the temporal relationship of the data. Subsequently, a convolutional neural network is used to extract The spatial characteristics of is the convolution kernel weight, represents the convolution operation, is the bias term, which eventually forms the input feature matrix after spatiotemporal encoding , capturing the spatial relationships between different regions.
[0099] Multi-head self-attention mechanism module: includes multiple self-attention heads, targeting the input feature matrix after spatiotemporal encoding , through multiple self-attention heads Calculate self-attention in parallel and calculate the input feature matrix after spatiotemporal encoding The correlation between them captures the dependencies between different time and space points; its specific form is:
[0100] (19),
[0101] (20),
[0102] (twenty one),
[0103] Among them, concat means concatenating the outputs of all self-attention heads together. is the number of self-attention heads, is the output weight matrix. Self-Attention Head , , the input feature matrix after spatiotemporal encoding is calculated by the self-attention mechanism The correlation between them is the input feature matrix after spatiotemporal encoding Mapping to the query matrix , key matrix Sum Matrix ,in , and Respectively The weight matrix of the query matrix, key matrix and value matrix in the self-attention head. Attention represents the self-attention weight mechanism. The attention weight can then be calculated. ,in is the activation function, is the dimension size of each self-attention head, T Represents the transpose of the matrix, and finally forms the multi-head self-attention mechanism feature matrix .
[0104] Feedforward neural network module: includes two linear transformation layers and an activation function, which is used to calculate the feature matrix of the multi-head self-attention mechanism Perform nonlinear transformation to enhance the model's ability to express complex features; its specific form is:
[0105] (twenty two),
[0106] (twenty three),
[0107] (twenty four),
[0108] Among them, the multi-head self-attention mechanism feature matrix After the first layer of linear transformation, and are the weight matrix and bias term of the first layer linear transformation, forming the intermediate feature representation ; Intermediate feature representation go through Activation function, introduces nonlinear characteristics, and forms the feature representation after activation ; Feature representation after activation After the second layer of linear transformation, and They are the weight matrix and bias term of the second layer linear transformation, forming the feedforward neural network feature matrix .
[0109] Spatiotemporal decoding module: includes a deconvolution layer, which is the inverse process of a convolution operation. Upsample to restore the original spatial resolution; its specific form is:
[0110] (25),
[0111] in, is the feature matrix after spatiotemporal decoding, ReLU represents the activation function, and are the deconvolution kernel weight and bias term respectively, Represents a deconvolution operation.
[0112] Output module: includes a fully connected layer, which maps high-dimensional features to the target output space and decodes the feature matrix after spatiotemporal decoding. Converted into precipitation prediction features; its specific form is:
[0113] (26),
[0114] in, and are the weight matrix and bias term of the fully connected layer, respectively. It is the precipitation prediction characteristics output by the model.
[0115] S33, train the model constructed in S32, pass the data training set through the input module, spatiotemporal encoding module, multi-head self-attention mechanism module, feedforward neural network module, spatiotemporal decoding module and output module, generate precipitation prediction value, and calculate the loss function; use the back propagation algorithm to calculate the gradient of the loss function relative to the model parameters, and use the gradient descent method to update the model parameters to minimize the loss function; evaluate the model performance on the validation set, calculate the loss value and evaluation index on the validation set, and determine the generalization ability of the model; such as Figure 4 shown.
[0116] S4. Based on the deep learning model of the space-time transformation network trained in S3, the trained model is applied to the current prediction for the target time, the numerical model precipitation prediction and precursor factor data are input, and the sub-seasonal extreme precipitation prediction results are output; the specific steps are as follows:
[0117] S41, inputting the numerical model precipitation prediction and precursor factor data within the target time into the space-time transformation network deep learning model trained in S3 to form the precipitation prediction characteristics of the target time;
[0118] S42, corresponding to the standardized inverse process of S1, substitutes the precipitation prediction characteristics of the target time and outputs the final sub-seasonal extreme precipitation prediction result.
[0119] In summary, the precipitation prediction method described in the present invention, through the deep fusion of numerical model precipitation prediction, historical precipitation observation and observation precursor factor data, uses a new deep learning method of space-time transformation network, which significantly improves the accuracy and comprehensiveness of sub-seasonal extreme precipitation prediction, and has extremely high application value and broad application prospects.
[0120] Based on the same inventive concept, the precipitation prediction system corresponding to the above method includes:
[0121] The model building and training unit is used to build a spatiotemporal transformation network deep learning model that integrates numerical model precipitation prediction, historical precipitation observation and observation precursor factors based on an optimized loss function, and to carry out model training and optimization based on the numerical model precipitation prediction data set, historical precipitation observation data set and observation precursor factor data set;
[0122] The prediction unit is used to use the trained model, input the numerical model precipitation prediction and precursor factor data of the target time for prediction, and output the sub-seasonal extreme precipitation prediction results;
[0123] Among them, the spatiotemporal transformation network deep learning model includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module and an output module; the input module obtains the numerical model precipitation prediction data set, the historical precipitation observation data set and the observation precursor factor data set, and splices them into the model input feature matrix; the spatiotemporal encoding module extracts the spatiotemporal features of the input feature matrix, wherein the temporal encoding embeds the time information through the position encoding method to capture the temporal relationship of the input feature matrix, and the spatial encoding extracts the spatial relationship between the input feature matrices of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module includes a deconvolution layer, which is used to gradually restore the spatial resolution of the data and obtain the spatiotemporal decoded feature matrix; the output module includes a fully connected layer, which converts the generated spatiotemporal decoded feature matrix into prediction features.
Claims
1. A sub-seasonal extreme precipitation prediction method with deep fusion of precursor factors, It is characterized in that The following steps are involved: Based on the optimized loss function, a spatiotemporal transformation network deep learning model integrating numerical model precipitation prediction, historical precipitation observation and observed precursor factors is constructed. Model training and optimization are carried out based on numerical model precipitation prediction dataset, historical precipitation observation dataset and observed precursor factor dataset. Using the trained model, the numerical model precipitation prediction and precursor factor data of the target time are input for prediction, and the sub-seasonal extreme precipitation prediction results are output. The method for producing the observation precursor factor data set is: adaptively annotate and encode the observation precursor factor data to generate a feature vector in a unified format, specifically: The observed precursor factor data include the sea temperature, sea ice, and circulation during the same period and before the reporting time. The adaptive annotation coding preprocessing calculation method is: , , , Among them, Embedding is the annotation coding embedding function, SST , SeaIce and Circulation are the sea temperature, sea ice and circulation data matrices for the same period and previous period of the reporting time, , and They are the sea temperature, sea ice and circulation data matrices after adaptive annotation and encoding; The normalization preprocessing calculation method is: , , , in, , and are the minimum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the maximum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling coding and normalization preprocessing; The spatiotemporal transformation network deep learning model includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module and an output module; the input module obtains the numerical model precipitation prediction data set, the historical precipitation observation data set and the observation precursor factor data set, and splices them into the model input feature matrix; the spatiotemporal encoding module extracts the spatiotemporal features of the input feature matrix, wherein the temporal encoding embeds the time information through the position encoding method to capture the temporal relationship of the input feature matrix, and the spatial encoding further extracts the spatial relationship between the input feature matrices of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module is used to gradually restore the spatial resolution of the data to obtain the spatiotemporal decoded feature matrix; the output module converts the generated spatiotemporal decoded feature matrix into prediction features.
2. According to the method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors described in claim 1, It is characterized in that The preparation method of the numerical model precipitation prediction dataset and the historical precipitation observation dataset is as follows: First, the numerical model precipitation prediction data and historical precipitation observation data are denoised and then standardized preprocessed. The formula is: , , in, and are the denoised numerical model precipitation prediction data and historical precipitation observation data matrices, and are the means of the denoised numerical model precipitation prediction data and the historical precipitation observation data, and are the standard deviations of the denoised numerical model precipitation prediction data and historical precipitation observation data, and They are the standardized numerical model precipitation prediction data and historical precipitation observation data matrices respectively.
3. According to the method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors as described in claim 1, It is characterized in that The optimized loss function is: , , , , in, Loss is the optimized loss function, is the weight of the mean square error loss term, is the weight of the critical success index loss term, is the weight of the structural similarity index loss term, MSE is the mean square error, CSI is the critical success index, SSIM is the structural similarity index, N is the evaluation sample size in model training, and Respectively The observed and predicted values of samples, TP is the actual number of samples, FP is the number of false positive samples, FN is the number of false negative samples, and They are N Group and The average value of and They are N Group and The variance of for N Group and The covariance between and is a stability constant.
4. The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors according to claim 1, It is characterized in that The processing process of the spatiotemporal coding module is: , , , , , in, Indicates at location The first The position encoding value in the dimension is calculated using the sine function; is the time step, indicating the position of the current data point in the time series; is the feature dimension, which indicates the index of the current feature in the feature matrix; is the total dimension of the input feature matrix; Indicates at location The first The position encoding value in the dimension is calculated using the cosine function; PE Represents the time position encoding matrix, where each row corresponds to a time step and each column corresponds to a feature dimension; Describe the input feature matrix after temporal encoding, represents the model input feature matrix, , and are the standardized numerical model precipitation prediction data and historical precipitation observation data matrices, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling coding and normalization preprocessing; represents the input feature matrix after spatiotemporal encoding; ReLU represents the activation function, is the convolution kernel weight, represents the convolution operation, is the bias term.
5. The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors according to claim 1, It is characterized in that The multi-head self-attention mechanism module includes multiple self-attention heads, which focus on the input feature matrix after spatiotemporal encoding. , through multiple self-attention heads Compute self-attention in parallel, calculate the correlation between input features, and capture the dependencies between different time and space points; Its specific form is: , , , in, represents the feature matrix of the multi-head self-attention mechanism, concat represents the concatenation function, Indicates A self-attention head, , is the number of self-attention heads, is the output weight matrix, , and denote the query matrix, key matrix and value matrix respectively, , , , , and Respectively The weight matrix of the query matrix, key matrix and value matrix in the self-attention head, Attention represents the self-attention weight mechanism, and softmax is the activation function. For each head dimension size, d is the total dimension of the input feature matrix, T Represents the transpose of a matrix.
6. The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors according to claim 1, It is characterized in that The feedforward neural network module consists of two linear transformation layers and an activation function, which is used to calculate the feature matrix of the multi-head self-attention mechanism. Perform nonlinear transformation to enhance the model's ability to express complex features; its specific form is: , , , in, represents the intermediate feature, represents the feature matrix of the multi-head self-attention mechanism, and are the weight matrix and bias term of the first layer linear transformation respectively; ReLU represents the activation function, Represents the activated features; represents the feedforward neural network feature matrix, and are the weight matrix and bias term of the second layer linear transformation respectively.
7. The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors according to claim 1, It is characterized in that The spatiotemporal decoding module consists of a deconvolution layer that transforms the feature matrix of the feedforward neural network into Upsample to restore the original spatial resolution; its specific form is: , in, represents the feature matrix after spatiotemporal decoding, ReLU represents the activation function, and are the deconvolution kernel weight and bias term respectively, Represents a deconvolution operation.
8. The method for predicting sub-seasonal extreme precipitation with deep fusion of precursor factors according to claim 1, It is characterized in that The output module includes a fully connected layer that maps high-dimensional features to the target output space and transforms the feature matrix after spatiotemporal decoding into Converted into prediction features; its specific form is: , in, is the precipitation prediction feature output by the model, and are the weight matrix and bias term of the fully connected layer respectively.
9. A sub-seasonal extreme precipitation prediction system with deep integration of precursor factors, It is characterized in that include: The model building and training unit is used to build a spatiotemporal transformation network deep learning model that integrates numerical model precipitation prediction, historical precipitation observation and observation precursor factors based on an optimized loss function, and to carry out model training and optimization based on the numerical model precipitation prediction data set, historical precipitation observation data set and observation precursor factor data set; The prediction unit is used to use the trained model, input the numerical model precipitation prediction and precursor factor data of the target time for prediction, and output the sub-seasonal extreme precipitation prediction results; The method for producing the observation precursor factor data set is: adaptively annotate and encode the observation precursor factor data to generate a feature vector in a unified format, specifically: The observed precursor factor data include the sea temperature, sea ice, and circulation during the same period and before the reporting time. The adaptive annotation coding preprocessing calculation method is: , , , Among them, Embedding is the annotation coding embedding function, SST , SeaIce and Circulation are the sea temperature, sea ice and circulation data matrices for the same period and previous period of the reporting time, , and They are the sea temperature, sea ice and circulation data matrices after adaptive annotation and encoding; The normalization preprocessing calculation method is: , , , in, , and are the minimum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the maximum values of sea temperature, sea ice and circulation after adaptive labeling and coding, , and They are the sea temperature, sea ice and circulation data matrices after adaptive labeling coding and normalization preprocessing; The spatiotemporal transformation network deep learning model includes an input module, a spatiotemporal encoding module, a multi-head self-attention mechanism module, a feedforward neural network module, a spatiotemporal decoding module and an output module; the input module obtains the numerical model precipitation prediction data set, the historical precipitation observation data set and the observation precursor factor data set, and splices them into the model input feature matrix; the spatiotemporal encoding module extracts the spatiotemporal features of the input feature matrix, wherein the temporal encoding embeds the time information through the position encoding method to capture the temporal relationship of the input feature matrix, and the spatial encoding further extracts the spatial relationship between the input feature matrices of different regions through the convolutional neural network; the multi-head self-attention mechanism module captures the dependency between different time and space points by calculating the correlation between spatiotemporal features; the feedforward neural network module further processes the output of the multi-head self-attention mechanism module to enhance the expression ability of complex features; the spatiotemporal decoding module is used to gradually restore the spatial resolution of the data to obtain the spatiotemporal decoded feature matrix; the output module converts the generated spatiotemporal decoded feature matrix into prediction features.
