A method for forecasting continuous distribution of precipitation over an extended period based on fine-grained terrain features

Abstract

The present invention discloses a method for forecasting the continuous distribution probability of extended-period precipitation by integrating fine terrain features, including: obtaining standardized multi-member multi-factor forecast data of the extended-period numerical model ensemble forecast system and regional historical long-term series precipitation observation data, and extracting the continuous distribution characteristics of historical precipitation; obtaining a descaled key terrain multidimensional feature factor library based on regional high-resolution fine terrain data; constructing an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model, and carrying out model training and optimization; using the trained model to generate an improved forecast result of the continuous distribution probability of extended-period precipitation. The present invention integrates fine terrain data, adopts a weighted cross entropy loss function, effectively handles the problem of imbalanced precipitation distribution categories, improves the generalization ability of the model, and improves the forecast performance of the continuous distribution probability forecast of precipitation, which has extremely high application value.

Description

Technical Field

[0001] The invention belongs to the field of extended-period precipitation continuous distribution probability forecasting, in particular to an extended-period precipitation continuous distribution probability forecasting method integrating fine terrain features. Background Art

[0002] Extended-range forecast refers to weather forecasts with a time scale of between weeks and months, which is between short-term weather forecasts and long-term climate predictions. It has important application value in disaster prevention and mitigation, agricultural production, water resources management and other fields. However, since the complex physical mechanisms of the Earth system at the extended-range scale have not yet been clarified, extended-range forecasts face the challenges of complex weather systems and high uncertainty. At present, the numerical model ensemble forecast system is operated in meteorological operational forecast centers in many countries around the world. It uses a variety of different initial conditions and different physical parameterization schemes, and provides a series of possible weather scenarios through multi-member forecasts. In this context, probabilistic forecasts can provide probability information on the occurrence of weather events, can quantitatively describe the uncertainty characteristics of forecasts, and improve the scientificity and reliability of decision-making.

[0003] At the same time, topography often has a significant impact on the occurrence and development of precipitation, especially in complex terrain areas (such as mountainous and hilly areas), and areas with different terrain characteristics are often accompanied by different temporal and spatial characteristics of precipitation. Incorporating fine terrain features (such as elevation, slope aspect, slope, undulation, and roughness) into precipitation forecast models helps to accurately capture the impact mechanism of terrain on precipitation. On the other hand, traditional precipitation probability forecasts are mainly carried out for specific magnitude precipitation forecasts. When facing areas with different climate characteristics, their differences are not sufficiently reflected, making it difficult to meet the needs of emergency disaster reduction and decision-making services. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for forecasting the continuous distribution probability of precipitation over an extended period of time by integrating fine terrain features.

[0005] Technical solution: The method for predicting the continuous distribution probability of precipitation over an extended period by integrating fine terrain features of the present invention comprises the following steps:

[0006] Obtain standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and regional historical long-term series precipitation observation data, and extract the continuous distribution characteristics of historical precipitation;

[0007] Based on the regional high-resolution fine terrain data, key terrain features are extracted, and the standardized key terrain feature data are combined to form a de-scaled key terrain multi-dimensional feature factor library;

[0008] The extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model is constructed, and the model training and optimization are carried out; the model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module and an output module. The input module is used to combine the standardized extended-period numerical model ensemble forecast system multi-member multi-factor forecast data, the regional historical long-term series precipitation observation data, the historical precipitation continuous distribution characteristics and the descaled key terrain multi-dimensional feature factor library to form the original input features; the encoder module is used to extract the multi-scale spatial features of the original input features; the intermediate layer module is used to transfer and process the multi-scale spatial features to obtain a higher-level feature matrix; the feature mapping module is used to use the extreme gradient boosting model to learn the mapping relationship between higher-level features and precipitation probability, generate a feature matrix, and finally obtain preliminary forecast results; the feature analysis module is used to calculate the importance of different features generated by the extreme gradient boosting model; the feature correction module is used to correct the feature matrix generated by the extreme gradient boosting model; the decoder module is used to reconstruct the high-resolution corrected feature matrix based on the corrected feature matrix and multi-scale spatial features, and the output module generates the final precipitation continuous distribution probability based on the high-resolution corrected feature matrix;

[0009] By using the trained model, for the target time, based on the multi-member and multi-factor forecast data of the extended-range numerical model ensemble forecast system and the regional high-resolution and fine terrain data, an improved extended-range precipitation continuous distribution probability forecast result is generated.

[0010] Furthermore, based on the standardized regional historical long-term series precipitation observation data, the continuous distribution characteristics of historical precipitation are extracted, including the precipitation observation data itself, as well as its mean, standard deviation, skewness and kurtosis.

[0011] Furthermore, the calculation formulas for mean, standard deviation, skewness, and kurtosis are:

[0012] ,

[0013] ,

[0014] ,

[0015] ,

[0016] in, For the The precipitation observation data of samples, is the number of samples of regional historical long-term precipitation observation data after preprocessing, is the mean precipitation observation value, is the standard deviation of precipitation observations, is the precipitation observation skewness, is the precipitation observation kurtosis.

[0017] Furthermore, the extracted key terrain features include elevation, aspect, slope, undulation and roughness, which are standardized so that the characteristic values ​​are within the same scale range, and the standardized elevation, aspect, slope, undulation and roughness are combined together to form a de-scaled key terrain multidimensional feature factor library.

[0018] Furthermore, the calculation formulas for elevation, aspect, slope, relief and roughness are:

[0019] ,

[0020] ,

[0021] ,

[0022] ,

[0023] ,

[0024] in, and Respectively represent the longitude and latitude grid positions, is the elevation value library, is the elevation, which indicates the height of the ground surface relative to sea level; and are the different directional gradients of elevation, is the slope direction, which indicates the azimuth of the ground surface. is the slope, which indicates the steepness of the terrain; is the actual surface area of ​​the region, is the projected plane area of ​​the regional terrain, is the relief, which indicates the complexity of the terrain surface; For the The elevation of the latitude and longitude grid points, is the average elevation, is the total number of longitude and latitude grid points, Roughness refers to the degree of undulation of the terrain surface.

[0025] Furthermore,

[0026] Input module: It contains a feature integration layer, which reads the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and regional historical long-term series precipitation observation data, historical precipitation continuous distribution characteristics, and de-scaled key terrain multi-dimensional feature factor library to form the original input features;

[0027] Encoder module: contains multiple convolutional layers and pooling layers, implemented as;

[0028] ,

[0029] ,

[0030] in, Indicates The local feature matrix of the convolutional layer, Indicates The local feature matrix of the convolutional layer, For the The convolution weights of the convolutional layers, is the convolution operation, For the The convolutional layer corresponds to the bias term, is the activation function, For the pooling operation, Indicates The high-level feature matrix output by layer pooling, i.e., the multi-scale spatial features obtained by the encoder module;

[0031] The middle layer module: contains a convolution layer and a batch normalization layer, and is implemented as follows:

[0032] ,

[0033] ,

[0034] in, is the convolution kernel weight, is the corresponding bias term, is the feature matrix output by the convolutional layer, is the scaling parameter, is the translation parameter, is the mean of the small batch data, is the variance of the mini-batch data, To prevent division by zero for small constants, It is the feature matrix after batch normalization, and also the higher-level feature matrix obtained by the intermediate layer module;

[0035] Feature mapping module: The implementation form is:

[0036] ,

[0037] ,

[0038] in, represents the extreme gradient boosting model operation, The feature matrix generated for the extreme gradient boosting model, is the loss function of the extreme gradient boosting model, It is The loss between the predicted value and the true value of samples, is a regularization term for the complexity of the extreme gradient boosting model, is the first decision tree feature function, and finally Substitute into the extreme gradient boosting model to form a preliminary forecast result ;

[0039] Feature analysis module: The implementation form is:

[0040] ,

[0041] in, Represents a decision tree in the extreme gradient boosting model The loss caused by using specific features in is the set of all decision trees, is a measure of the importance of a particular feature;

[0042] Feature correction module: It includes a feature fusion layer and a matrix correction layer, and is implemented as follows:

[0043] ,

[0044] in, is the correction factor, represents the Hadamard product, is the modified feature matrix;

[0045] Decoder module: Contains a transposed convolution layer and a skip connection layer, implemented as follows:

[0046] ,

[0047] ,

[0048] in, is the transposed convolution operation, is the feature matrix of the transposed convolutional layer output, For the jump connection of the corresponding encoder module, The high-resolution corrected feature matrix output by the decoder module;

[0049] Output module: contains an activation layer, implemented as:

[0050] ,

[0051] in, is the activation function, is the final precipitation continuous distribution probability.

[0052] Furthermore, the model training and optimization method is:

[0053] Through forward propagation, features are extracted in the input module, encoder module, intermediate layer module, feature mapping module, feature analysis module, feature correction module, decoder module, and output module in turn to generate the continuous distribution probability of precipitation. The full-scale weighted cross entropy loss function is used to calculate the loss, and the Adam optimizer is used to update the model parameters during the back propagation process. After multiple iterative updates, the model converges or reaches the preset number of iterations to ensure the optimization of model performance.

[0054] Furthermore, the expression of the full-scale weighted cross entropy loss function is:

[0055] ,

[0056] in, is the evaluation sample size in model training, is the number of continuous distribution categories of precipitation; Indicates samples belong to the category The true label of The model forecast samples belong to the category probability; For Category The weight of .

[0057] The extended-period precipitation continuous distribution probability forecasting system integrating fine terrain features of the present invention comprises:

[0058] The first data processing unit is used to obtain the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and the regional historical long-term series precipitation observation data, and extract the continuous distribution characteristics of historical precipitation;

[0059] The second data processing unit is used to extract key terrain features based on the regional high-resolution fine terrain data, and combine the standardized key terrain feature data to form a de-scaled key terrain multi-dimensional feature factor library;

[0060] The model building and training unit is used to build an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model, and to carry out model training and optimization; wherein, the model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module and an output module. The input module is used to combine the standardized extended-period numerical model ensemble forecast system multi-member multi-factor forecast data, the regional historical long-term series precipitation observation data, the historical precipitation continuous distribution characteristics and the de-scaled key terrain multi-dimensional feature factor library to form the original input features; the encoder module is used to extract the multi-scale space of the original input features. Features; The middle layer module is used to transfer and process multi-scale spatial features to obtain a higher-level feature matrix; The feature mapping module is used to use the extreme gradient boosting model to learn the mapping relationship between higher-level features and precipitation probability, generate a feature matrix, and finally obtain preliminary forecast results; The feature analysis module is used to calculate the importance of different features generated by the extreme gradient boosting model; The feature correction module is used to correct the feature matrix generated by the extreme gradient boosting model; The decoder module is used to reconstruct a high-resolution corrected feature matrix based on the corrected feature matrix and multi-scale spatial features, and the output module generates the final continuous distribution probability of precipitation based on the high-resolution corrected feature matrix;

[0061] The forecast unit is used to use the trained model to generate improved extended-range precipitation continuous distribution probability forecast results for the target time based on the multi-member and multi-factor forecast data of the extended-range numerical model ensemble forecast system and the regional high-resolution fine terrain data.

[0062] The electronic device of the present invention comprises:

[0063] A memory storing executable program code;

[0064] a processor coupled to the memory;

[0065] The processor calls the executable program code stored in the memory to execute the steps of the method for forecasting the continuous distribution probability of precipitation over an extended period by integrating fine terrain features.

[0066] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are: (1) by integrating high-resolution fine terrain features (such as elevation, aspect, slope, undulation and roughness) and introducing a key terrain multi-dimensional feature factor library, the extended-term precipitation continuous distribution probability forecast can more accurately capture the precipitation distribution characteristics of complex terrain areas, and improve the model's sensitivity to terrain changes compared to traditional precipitation forecasting methods, thereby significantly improving the accuracy of the extended-term precipitation probability forecast; (2) an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model is established, which effectively extracts and maps multi-layer features. By capturing the nonlinear relationship between multiple features and precipitation, and enhancing the ability to extract and reconstruct multi-scale features, a precipitation probability forecasting model with strong generalization ability is constructed, which enhances the adaptability and stability of the model in different regions and at different times. (3) A full-scale weighted cross entropy loss function is proposed for the probability forecasting of continuous distribution of precipitation, which optimizes the imbalance problem of precipitation data and gives higher weights to extreme precipitation events of greater concern, so that the model can learn and forecast key events more effectively, thereby improving the overall forecasting performance while also effectively improving the forecasting skills of key event characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2 To integrate the detailed topographic features;

[0069] Figure 3 Schematic diagram of the extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model;

[0070] Figure 4 Optimize the diagram for model training. DETAILED DESCRIPTION

[0071] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0072] In the process of realizing the extended-range precipitation probability forecast, we can carry out targeted analysis of the regional precipitation continuous distribution characteristics, and combine the terrain characteristics to construct an extended-range precipitation continuous distribution probability forecast that integrates fine terrain features, providing a more comprehensive regional precipitation forecast risk assessment and response reference.

[0073] The method for forecasting the continuous distribution probability of precipitation over an extended period by integrating fine terrain features provided by the present invention has the following process: Figure 1 As shown, the specific steps include:

[0074] S1. Obtain the multi-member and multi-factor forecast data of the extended-range numerical model ensemble forecast system and the regional historical long-term series precipitation observation data, and perform data cleaning and standardization preprocessing to extract the continuous distribution characteristics of historical precipitation; divide them into training set and validation set according to a certain ratio; the specific steps are as follows:

[0075] S11. Obtain the multi-member and multi-factor forecast data (precipitation, temperature, air pressure, humidity, wind speed) of the extended-range numerical model ensemble forecast system and the regional historical long-term series precipitation observation data, and perform data cleaning and standardization preprocessing; data cleaning first fills in missing values ​​or deletes samples with too many missing values, and then uses the Z-score method to detect and process outliers; standardization processing is to standardize the data to make the scales of different features consistent.

[0076] S12. Based on the preprocessed regional historical long-term series precipitation observation data, extract the historical precipitation continuous distribution characteristics, including the precipitation observation data itself, as well as the mean, standard deviation, skewness, and kurtosis. The calculation method is:

[0077] ,

[0078] ,

[0079] ,

[0080] ,

[0081] in, For the The precipitation observation data of samples, is the number of samples of regional historical long-term precipitation observation data after preprocessing, is the mean precipitation observation value, is the standard deviation of precipitation observations, is the precipitation observation skewness, is the precipitation observation kurtosis.

[0082] S13. Divide the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data, regional historical long-term precipitation observation data, and historical precipitation continuous distribution characteristics into training set and validation set in a ratio of 8:2, respectively.

[0083] S2. Obtain high-resolution and fine terrain data of the region, extract key terrain features, including elevation, aspect, slope, undulation, and roughness, and perform normalization preprocessing; integrate the preprocessed key terrain feature data to form a descaled key terrain multidimensional feature factor library; Figure 2 As shown, the specific steps are as follows;

[0084] S21. Obtain high-resolution and fine terrain data of the region and extract key terrain features, including elevation, aspect, slope, undulation, and roughness. The calculation method is as follows:

[0085] ,

[0086] ,

[0087] ,

[0088] ,

[0089] ,

[0090] in, and Respectively represent the longitude and latitude grid positions, is the elevation value library, is the elevation, which indicates the height of the ground surface relative to sea level; and are the different directional gradients of elevation, is the slope direction, which indicates the azimuth of the ground surface. is the slope, which indicates the steepness of the terrain; is the actual surface area of ​​the region, is the projected plane area of ​​the regional terrain, is the relief, which indicates the complexity of the terrain surface; For the The elevation of the latitude and longitude grid points, is the average elevation, is the total number of longitude and latitude grid points, Roughness refers to the degree of undulation of the terrain surface.

[0091] S22. Perform normalization preprocessing on the extracted key terrain features so that the feature values ​​are within the same scale range for easy model processing. The calculation method is:

[0092] ,

[0093] in, The original key terrain features, and are the maximum and minimum values ​​of the original key terrain features, are the normalized key terrain features.

[0094] S23. All normalized key terrain feature data are combined together to form a de-scaled key terrain multi-dimensional feature factor library for model calling.

[0095] S3. Establish a full-scale weighted cross entropy loss function for the continuous distribution probability forecast of precipitation, construct an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model that integrates the multi-member and multi-factor forecast data of the extended-period numerical model ensemble forecast system, the regional long-term precipitation observation data, and the regional high-resolution fine terrain data, and carry out model training and optimization; wherein, the extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module, and an output module; the input module is used to obtain and integrate the standardized extended-period numerical model ensemble forecast system multi-member and multi-factor forecast data, the regional historical long-term precipitation observation data, the historical precipitation continuous distribution characteristics, and the descaled ones in step S2. The original input features are obtained from the multi-dimensional feature factor library of key terrain; the encoder module is used to extract the multi-scale spatial features of the original input features; the intermediate layer module is used to transfer and process the multi-scale spatial features to generate a higher-level feature matrix; the feature mapping module is used to use the extreme gradient boosting model to learn the mapping relationship between higher-level features and precipitation probability, generate a feature matrix, and finally obtain preliminary forecast results; the feature analysis module is used to calculate the importance of different features generated by the extreme gradient boosting model and identify the most important and contributing features to precipitation forecast; the feature correction module is used to correct the feature matrix generated by the extreme gradient boosting model; the decoder module is used to reconstruct a high-resolution corrected feature matrix based on the corrected feature matrix and multi-scale spatial features; the output module generates the final continuous distribution probability of precipitation based on the high-resolution corrected feature matrix; the specific steps are as follows:

[0096] S31. Establish a full-scale weighted cross entropy loss function for the continuous distribution probability forecast of precipitation. By assigning different weights to different precipitation levels, the model is optimized to achieve a comprehensive, accurate and efficient correction of the full-scale precipitation forecast. The specific form is:

[0097] ,

[0098] in, is the evaluation sample size in model training, is the number of continuous distribution categories of precipitation; Indicates samples belong to the category The true label is a binary variable, which takes 0 or 1, representing the sample Not belonging to or belonging to the category ; The model forecast samples belong to the category The probability is a real number between 0 and 1, indicating the model's Belongs to category confidence; For Category The weight of is used to deal with the problem of imbalanced categories. This loss function can solve the problem of imbalanced precipitation distribution categories during model training and improve the model's generalization ability and forecasting performance.

[0099] S32. Using the full-scale weighted cross entropy loss function obtained in S31, an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecasting model is constructed by integrating multi-member and multi-factor forecast data of the numerical model ensemble forecasting system, regional long-term precipitation observation data, and regional high-resolution fine terrain data. The model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module, and an output module; Figure 3 As shown, its specific form is:

[0100] Input module: It contains a feature integration layer. The model reads the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and regional historical long-term series precipitation observation data and historical precipitation continuous distribution characteristics obtained in step S1, as well as the descaled key terrain multidimensional feature factor library obtained in step S2, and integrates them to form the original input features.

[0101] Encoder module: contains multiple convolutional layers and pooling layers, which are used to extract multi-scale spatial features of the original input features; the convolutional layer is used to extract local features, and the pooling layer is used to reduce the resolution of the feature matrix and enhance the abstraction ability of the features. Its specific form is:

[0102] ,

[0103] ,

[0104] in, Indicates The local feature matrix of the convolutional layer, Indicates The local feature matrix of the convolutional layer, For the The convolution weights of the convolutional layers, is the convolution operation, For the The convolutional layers correspond to the bias terms, is the activation function, For the pooling operation, Indicates The high-level feature matrix output by layer pooling is the multi-scale spatial features obtained by the encoder module.

[0105] Middle layer module: contains a convolution layer and a batch normalization layer, which are used to transfer and process the multi-scale spatial features output by the encoder module; the convolution layer is used to further process the multi-scale spatial features output by the encoder module, and the batch normalization layer is used to stabilize the training process and accelerate convergence; the middle layer module can generate a higher-level feature matrix based on the encoder module. Its specific form is:

[0106] ,

[0107] ,

[0108] in, is the convolution kernel weight, is the corresponding bias term, is the feature matrix output by the convolutional layer, is the scaling parameter, is the translation parameter, is the mean of the data in the current batch in batch normalization, is the variance of the data in the current batch in batch normalization, To prevent division by zero for small constants, It is the feature matrix after batch normalization, and also the higher-level feature matrix obtained by the intermediate layer module.

[0109] Feature mapping module: Analyze the higher-level feature matrix obtained by the intermediate layer module, use extreme gradient boosting to learn the mapping relationship between the higher-level feature matrix and precipitation probability, perform feature selection, and generate preliminary forecast results. Its specific form is:

[0110] ,

[0111] ,

[0112] in, represents the extreme gradient boosting model operation, The feature matrix generated for the extreme gradient boosting model, is the loss function of the extreme gradient boosting model, It is The loss between the predicted value and the true value of samples, is a regularization term for the complexity of the extreme gradient boosting model, is the first decision tree feature function, and finally Substitute into the extreme gradient boosting model to form a preliminary forecast result .

[0113] Feature analysis module: Calculate the importance of different features generated by the extreme gradient boosting model in the feature mapping module, and identify the most important and contributing features for precipitation forecasting. Its specific form is:

[0114] ,

[0115] in, Represents a decision tree in the extreme gradient boosting model The loss caused by using specific features in is the set of all decision trees, is the importance measure for that particular feature.

[0116] Feature correction module: It includes a feature fusion layer and a matrix correction layer. It uses the feature importance measurement output by the feature analysis module and the preliminary prediction results formed by the extreme gradient boosting model of the feature mapping module to correct the feature matrix generated by the extreme gradient boosting model of the feature mapping module, generate a corrected feature matrix, and enhance the prediction ability of the subsequent U-Net decoder of the model. Its specific form is:

[0117] ,

[0118] in, is the correction factor, represents the Hadamard product, is the modified feature matrix.

[0119] Decoder module: It contains a transposed convolution layer and a skip connection layer, which are used to reconstruct high-resolution feature outputs; the transposed convolution layer restores the resolution of the corrected feature matrix to the high resolution of the original input features, and the skip connection layer directly transfers the multi-scale spatial features obtained by the encoder module to the decoder module to generate a high-resolution corrected feature matrix to retain more detailed information. Its specific form is:

[0120] ,

[0121] ,

[0122] in, is the transposed convolution operation, is the feature matrix of the transposed convolutional layer output, For the jump connection of the corresponding encoder module, It is the high-resolution rectified feature matrix output by the decoder module.

[0123] Output module: contains an activation layer, based on the high-resolution corrected feature matrix generated by the decoder module, using The function generates the continuous distribution probability of precipitation. Its specific form is:

[0124] ,

[0125] in, is the activation function, is the final precipitation continuous distribution probability.

[0126] S33, train the extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model constructed in S32, extract features and generate precipitation continuous distribution probability in the input module, encoder module, intermediate layer module, feature mapping module, feature analysis module, feature correction module, decoder module, and output module in sequence through forward propagation, use the full-scale weighted cross entropy loss function to calculate the loss, and use the Adam optimizer to update the model parameters during the back propagation process, and update the model after multiple iterations until the model converges or reaches the preset number of iterations to ensure the optimization of model performance; Figure 4 shown.

[0127] S4. Use the extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model trained in step S3 to generate improved extended-term precipitation continuous distribution probability forecast results for the target time based on the multi-member and multi-factor forecast data of the extended-term numerical model ensemble forecast system and the regional high-resolution fine terrain data.

Claims

1. A method for forecasting continuous distribution probability of precipitation over an extended period integrating fine terrain features, characterized in that: The following steps are involved: Obtain standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and regional historical long-term series precipitation observation data, and extract the continuous distribution characteristics of historical precipitation; Based on the regional high-resolution fine terrain data, key terrain features are extracted, including elevation, aspect, slope, undulation and roughness. The elevation, aspect, slope, undulation and roughness are standardized so that the characteristic values ​​are within the same scale range, and the standardized elevation, aspect, slope, undulation and roughness are combined together to form a de-scaled key terrain multi-dimensional feature factor library; The extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model is constructed, and the model training and optimization are carried out; the model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module and an output module. The input module is used to combine the standardized extended-period numerical model ensemble forecast system multi-member multi-factor forecast data, the regional historical long-term series precipitation observation data, the historical precipitation continuous distribution characteristics and the descaled key terrain multi-dimensional feature factor library to form the original input features; the encoder module is used to extract the multi-scale spatial features of the original input features; the intermediate layer module is used to transfer and process the multi-scale spatial features to obtain a higher-level feature matrix; the feature mapping module is used to use the extreme gradient boosting model to learn the mapping relationship between higher-level features and precipitation probability, generate a feature matrix, and finally obtain preliminary forecast results; the feature analysis module is used to calculate the importance of different features generated by the extreme gradient boosting model; the feature correction module is used to correct the feature matrix generated by the extreme gradient boosting model; the decoder module is used to reconstruct the high-resolution corrected feature matrix based on the corrected feature matrix and multi-scale spatial features, and the output module generates the final precipitation continuous distribution probability based on the high-resolution corrected feature matrix; By using the trained model, for the target time, based on the multi-member and multi-factor forecast data of the extended-range numerical model ensemble forecast system and the regional high-resolution and fine terrain data, an improved extended-range precipitation continuous distribution probability forecast result is generated.

2. The method for forecasting continuous distribution probability of precipitation over an extended period by integrating fine terrain features according to claim 1 is characterized in that: Based on the standardized regional historical long-term series precipitation observation data, the continuous distribution characteristics of historical precipitation are extracted, including the precipitation observation data itself, as well as its mean, standard deviation, skewness and kurtosis.

3. The method for forecasting continuous distribution probability of precipitation over an extended period by integrating fine terrain features according to claim 2 is characterized in that: The calculation formulas for mean, standard deviation, skewness, and kurtosis are: ， ， ， ， in, For the The precipitation observation data of samples, is the number of samples of regional historical long-term precipitation observation data after preprocessing, is the mean of precipitation observations, is the standard deviation of precipitation observations, is the precipitation observation skewness, is the precipitation observation kurtosis.

4. The method for predicting continuous distribution probability of precipitation over an extended period by integrating fine terrain features according to claim 1 is characterized in that: The calculation formulas for elevation, aspect, slope, relief and roughness are: ， ， ， ， ， in, and Respectively represent the longitude and latitude grid positions, is the elevation value library, is the elevation, which indicates the height of the ground surface relative to sea level; and are the different directional gradients of elevation, is the slope direction, which indicates the azimuth of the ground surface. is the slope, which indicates the steepness of the terrain; is the actual surface area of ​​the region, is the projected plane area of ​​the regional terrain, is the relief, which indicates the complexity of the terrain surface; For the The elevation of the latitude and longitude grid points, is the average elevation, is the total number of longitude and latitude grid points, Roughness refers to the degree of undulation of the terrain surface.

5. The method for forecasting continuous distribution probability of precipitation over an extended period integrating fine terrain features according to claim 1 is characterized by: Input module: It contains a feature integration layer, which reads the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and regional historical long-term series precipitation observation data, historical precipitation continuous distribution characteristics, and de-scaled key terrain multi-dimensional feature factor library to form the original input features; Encoder module: contains multiple convolutional layers and pooling layers, implemented as; ， ， in, Indicates The local feature matrix of the convolutional layer, Indicates The local feature matrix of the convolutional layer, For the The convolution weights of the convolutional layers, is the convolution operation, For the The convolutional layer corresponds to the bias term, is the activation function, For the pooling operation, Indicates The high-level feature matrix output by layer pooling, i.e., the multi-scale spatial features obtained by the encoder module; The middle layer module: contains a convolution layer and a batch normalization layer, and is implemented as follows: ， ， in, is the convolution kernel weight, is the corresponding bias term, is the feature matrix output by the convolutional layer, is the scaling parameter, is the translation parameter, is the mean of the small batch data, is the variance of the mini-batch data, To prevent division by zero for small constants, It is the feature matrix after batch normalization, and also the higher-level feature matrix obtained by the intermediate layer module; Feature mapping module: The implementation form is: ， ， in, represents the extreme gradient boosting model operation, The feature matrix generated for the extreme gradient boosting model, is the loss function of the extreme gradient boosting model, It is The loss between the predicted value and the true value of samples, is a regularization term for the complexity of the extreme gradient boosting model, is the first decision tree feature function, and finally Substitute into the extreme gradient boosting model to form a preliminary forecast result ; Feature analysis module: The implementation form is: ， in, Represents a decision tree in the extreme gradient boosting model The loss caused by using specific features in is the set of all decision trees, is a measure of the importance of a particular feature; Feature correction module: It includes a feature fusion layer and a matrix correction layer, and is implemented as follows: ， in, is the correction factor, represents the Hadamard product, is the corrected feature matrix; Decoder module: Contains a transposed convolution layer and a skip connection layer, implemented as follows: ， ， in, is the transposed convolution operation, is the feature matrix of the transposed convolutional layer output, For the jump connection of the corresponding encoder module, The high-resolution corrected feature matrix output by the decoder module; Output module: contains an activation layer, implemented as: ， in, is the activation function, is the final precipitation continuous distribution probability.

6. The method for forecasting continuous distribution probability of precipitation over an extended period integrating fine terrain features according to claim 1 is characterized by: The model training and optimization methods are: Through forward propagation, features are extracted in the input module, encoder module, intermediate layer module, feature mapping module, feature analysis module, feature correction module, decoder module, and output module in turn to generate the continuous distribution probability of precipitation. The full-scale weighted cross entropy loss function is used to calculate the loss, and the Adam optimizer is used to update the model parameters during the back propagation process. After multiple iterative updates, the model converges or reaches the preset number of iterations to ensure the optimization of model performance.

7. The method for forecasting continuous distribution probability of precipitation over an extended period by integrating fine terrain features according to claim 6 is characterized in that: The expression of the full-scale weighted cross entropy loss function is: ， in, is the evaluation sample size in model training, is the number of continuous distribution categories of precipitation; Indicates samples belong to the category The true label of The model forecast samples belong to the category The probability of For Category The weight of .

8. The extended-period precipitation continuous distribution probability forecast system integrating fine terrain features is characterized by: include: The first data processing unit is used to obtain the standardized extended-range numerical model ensemble forecast system multi-member multi-factor forecast data and the regional historical long-term series precipitation observation data, and extract the continuous distribution characteristics of historical precipitation; The second data processing unit is used to extract key terrain features based on the regional high-resolution fine terrain data, the extracted key terrain features including elevation, aspect, slope, undulation and roughness, standardize the elevation, aspect, slope, undulation and roughness so that the characteristic values ​​are within the same scale range, and combine the standardized elevation, aspect, slope, undulation and roughness to form a de-scaled key terrain multi-dimensional feature factor library; The model building and training unit is used to build an extreme gradient boosting U-Net hybrid deep learning precipitation probability forecast model, and to carry out model training and optimization; wherein, the model includes an input module, an encoder module, an intermediate layer module, a feature mapping module, a feature analysis module, a feature correction module, a decoder module and an output module. The input module is used to combine the standardized extended-period numerical model ensemble forecast system multi-member multi-factor forecast data, the regional historical long-term series precipitation observation data, the historical precipitation continuous distribution characteristics and the de-scaled key terrain multi-dimensional feature factor library to form the original input features; the encoder module is used to extract the multi-scale space of the original input features. Features; The middle layer module is used to transfer and process multi-scale spatial features to obtain a higher-level feature matrix; The feature mapping module is used to use the extreme gradient boosting model to learn the mapping relationship between higher-level features and precipitation probability, generate a feature matrix, and finally obtain preliminary forecast results; The feature analysis module is used to calculate the importance of different features generated by the extreme gradient boosting model; The feature correction module is used to correct the feature matrix generated by the extreme gradient boosting model; The decoder module is used to reconstruct a high-resolution corrected feature matrix based on the corrected feature matrix and multi-scale spatial features, and the output module generates the final continuous distribution probability of precipitation based on the high-resolution corrected feature matrix; The forecast unit is used to use the trained model to generate improved extended-range precipitation continuous distribution probability forecast results for the target time based on the multi-member and multi-factor forecast data of the extended-range numerical model ensemble forecast system and the regional high-resolution fine terrain data.

9. An electronic device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the method for forecasting the continuous distribution probability of precipitation over an extended period integrating fine terrain features as described in any one of claims 1 to 7.

Images