A regional precipitation prediction method and system based on multi-source heterogeneous deep integration

Abstract

This invention discloses a regional precipitation prediction method and system based on multi-source heterogeneous deep integration, belonging to the field of precipitation prediction technology. The method includes: classifying multi-source climate factors into atmospheric circulation, sea surface temperature, and comprehensive factor groups according to their physical properties; performing correlation screening, standardization, and PCA dimensionality reduction on each group of factors; constructing base models for each group of factors on a station-by-station basis using various machine learning algorithms, and employing a station-by-station hyperparameter search strategy; concatenating the out-of-place predictions of all base models and using the result as input to the ensemble layer of a deep neural network, learning dynamic ensemble weights station-by-station; and finally outputting the precipitation prediction results. This invention effectively eliminates the systematic bias of a single algorithm through heterogeneous integration and improves the model's adaptability to spatial heterogeneity through station-by-station hyperparameter search, exhibiting advantages such as high prediction accuracy, good interdecadal stability, and strong predictive ability for extreme precipitation events.

Description

Technical Field

[0001] This invention relates to the field of precipitation prediction technology, specifically to a regional precipitation prediction method and system based on multi-source heterogeneous deep integration. Background Technology

[0002] Subseasonal-seasonal regional precipitation forecasting is a crucial requirement for meteorological disaster prevention and mitigation. In recent years, statistical-machine learning precipitation forecasting methods based on large-scale climate factors (such as atmospheric circulation indices and sea surface temperature indices) have gradually developed. Representative existing schemes are as follows:

[0003] Option 1: Traditional statistical methods; using methods such as multiple linear regression and stepwise regression to establish the statistical relationship between climate factors and precipitation. These methods assume a linear relationship between factors and precipitation, making it difficult to capture the nonlinear influence of factors such as atmospheric circulation on precipitation.

[0004] Option 2: Single deep learning method; a deep neural network (DNN) is used to predict precipitation at multiple stations, and multiple base models are built using the same algorithm and then integrated (Jiang Wei, Liu Yunyun, Chen Peng, Zhang Zhiwei. 2021. An objective prediction method for summer precipitation in Jiangsu using deep neural networks and precursor signals. Acta Meteorologica Sinica, 79(6): 1035-1048). This method divides climate factors into multiple groups such as atmospheric circulation and sea surface temperature, and trains a DNN model for each group of factors, and then learns dynamic integration weights through DNN. All base models are of the same type of algorithm (DNN), which belongs to "homogeneous integration". The error patterns of each base model in homogeneous integration are highly correlated, and the integration gain is limited. When the training samples are limited (such as only more than 60 years of meteorological data), similar algorithms are prone to overfitting in the same direction, resulting in systematic bias after integration.

[0005] Option 3: Single machine learning ensemble method; using ensemble learning algorithms such as random forest and XGBoost for precipitation prediction. This type of method has a good ability to capture feature interactions, but it is not as good as deep learning in extracting deep nonlinear patterns. Summary of the Invention

[0006] The purpose of this invention is to provide a regional precipitation prediction method and system based on multi-source heterogeneous deep integration, so as to solve the above-mentioned problems existing in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The regional precipitation prediction method based on multi-source heterogeneous deep integration includes the following steps: Step S1, acquire multi-source climate factor data and historical precipitation observation data of multiple meteorological stations in the target area. The multi-source climate factors include at least atmospheric circulation index and sea surface temperature index. The climate factors are divided into at least two groups according to their physical properties. Step S2: Perform feature preprocessing on each group of climate factors; Step S3: For each group of preprocessed factors, at least two different types of machine learning algorithms are used to build base models. Each meteorological station is modeled independently, and K-fold cross-validation is used to generate the out-of-fold prediction of each base model during the training process. That is, each training sample is generated by the model that has not used the sample for training, so as to obtain a leak-free prediction sequence of the same length as the training set. Step S4: Using the out-of-range predictions generated by all base models as input, train a deep neural network as an ensemble layer for each meteorological station to learn the optimal nonlinear mapping relationship between the predicted values ​​of each base model and the actual precipitation observation values, thereby achieving adaptive learning of dynamic ensemble weights. Step S5: Input the predicted values ​​of each base model for the future period into the trained ensemble layer, and output the final precipitation prediction result.

[0008] Furthermore, in step S1, the climate factors are divided into atmospheric circulation factor group, sea surface temperature factor group and comprehensive factor group containing all factors according to their physical properties.

[0009] Furthermore, in step S2, the preprocessing includes: correlation pre-screening, standardization processing, and principal component analysis dimensionality reduction.

[0010] Furthermore, in step S3, the machine learning algorithms include deep neural network algorithms, tree model algorithms, and linear model algorithms. Among them, the deep neural network base model uses a station-by-station hyperparameter search strategy to determine the optimal hyperparameters, while the tree model and linear regression base models select the optimal hyperparameters for each station through cross-validation.

[0011] Furthermore, in step S4, the integration layer uses a station-by-station hyperparameter search strategy to determine the optimal network structure.

[0012] Furthermore, the hyperparameter search space of the ensemble layer deep neural network is consistent with the hyperparameter search space of the deep neural network algorithm of the base model in step S3.

[0013] Furthermore, the station-by-station hyperparameter search space of the deep neural network algorithm includes the number of hidden layers, the number of neurons per layer, and the Dropout rate; the tree model algorithm selects the maximum depth through cross-validation; and the linear model algorithm selects the regularization parameter through cross-validation.

[0014] Furthermore, in step S3, a time exponential weight is applied to the base model constructed from the sea surface temperature factor group.

[0015] Furthermore, during the base model training process in step S3 and the ensemble layer training process in step S4, Gaussian noise is added to the training data for data augmentation to alleviate the overfitting problem under small sample conditions; wherein, the augmentation factor of the base model training data is 3 times and the augmentation factor of the ensemble layer training data is 5 times.

[0016] A regional precipitation prediction system based on multi-source heterogeneous deep integration includes: The data acquisition and grouping module is used to acquire multi-source climate factor data and historical precipitation observation data from multiple meteorological stations in the target area, and to divide the climate factors into at least two groups according to their physical attributes. The feature processing module, connected to the data acquisition and grouping module, is used to perform feature preprocessing on each group of climate factors; The heterogeneous base model construction module, connected to the feature processing module, is used to construct a base model for each group of preprocessed factors using at least two different types of machine learning algorithms, perform independent modeling and hyperparameter optimization on a site-by-site basis, and generate out-of-bounds predictions. The deep ensemble layer module, connected to the heterogeneous base model construction module, is used to train the deep neural network site by site, taking the out-of-fold predictions of all base models as input, to learn dynamic ensemble weights. The prediction output module, connected to the deep integration layer module, is used to input the base model's predictions for future periods into the trained integration layer and output the final precipitation prediction result.

[0017] The beneficial effects of this invention are: Heterogeneous ensemble significantly outperforms homogeneous ensemble: This invention employs multiple different types of machine learning algorithms to construct the base model (heterogeneous ensemble). The error patterns of each algorithm are orthogonal, and the ensemble effectively eliminates the systematic bias of a single algorithm. Station-by-station adaptive hyperparameter search enhances the ability to model spatial heterogeneity: This invention independently determines the optimal hyperparameters for each meteorological station, enabling the model to adaptively match the precipitation-factor relationships of different stations. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the regional precipitation prediction method based on multi-source heterogeneous deep integration of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0020] Embodiments of the present invention: like Figure 1 As shown, the regional precipitation prediction method based on multi-source heterogeneous deep integration includes the following steps: Step S1: Data acquisition and grouping Acquire multi-source climate factor data and historical precipitation observation data from multiple meteorological stations in the target area. The multi-source climate factors include at least the atmospheric circulation index and the sea surface temperature index. The climate factors are categorized according to their physical properties into an atmospheric circulation factor group, a sea surface temperature factor group, and a comprehensive factor group containing all factors.

[0021] Step S2: Feature Selection and Dimensionality Reduction Each group of climate factors underwent characteristic preprocessing; the preprocessing included: Correlation pre-screening is performed by calculating the correlation coefficient between each factor and the average precipitation of stations in the target area, retaining factors whose absolute correlation coefficient exceeds a preset threshold, and removing redundant factors that are not significantly related to precipitation.

[0022] Standardization processing involves standardizing the screened factors to eliminate dimensional differences.

[0023] Principal component analysis (PCA) is used to reduce dimensionality. After standardization, PCA is performed on the factors to retain the principal components whose cumulative variance contribution rate reaches a preset proportion. This significantly reduces the feature dimensionality while keeping information loss under control, thus alleviating the curse of dimensionality problem in small samples.

[0024] Step S3: Construct the base model For each group of preprocessed factors, a base model is constructed using at least two different types of machine learning algorithms. These machine learning algorithms include deep neural network algorithms, tree model algorithms, and linear model algorithms. Different types of algorithms have complementary modeling characteristics, allowing different algorithms to adapt to the mathematical characteristics of different physical processes.

[0025] Each meteorological station is modeled independently, meaning a separate prediction model is trained for the precipitation sequence of each station.

[0026] For deep neural network base models, a station-by-station hyperparameter grid search strategy is employed: for each station, the predictive performance of each hyperparameter configuration is evaluated through K-fold cross-validation from a predefined hyperparameter search space, and the configuration with the minimum validation loss is selected as the optimal hyperparameter for that station. The station-by-station hyperparameter search space for deep neural network algorithms includes the number of hidden layers, the number of neurons per layer, and the Dropout rate; tree model algorithms select the maximum depth through cross-validation; and linear model algorithms select the regularization parameter through cross-validation.

[0027] During the K-fold cross-validation training process, the out-of-fold predictions (OOF) of each base model are generated simultaneously. That is, each training sample is predicted by a model that has not used the sample for training, thereby obtaining a leak-free prediction sequence of the same length as the training set, which serves as the input feature for subsequent ensemble layers.

[0028] The final model is retrained on all training data using the optimal hyperparameters for each site, and used for predictions of future time periods.

[0029] Step S4: Dynamic Weight Learning The out-of-division prediction vectors generated by all base models are concatenated column-wise to form the input feature matrix of the ensemble layer. A deep neural network is trained as the ensemble layer for each meteorological station to learn the optimal nonlinear mapping relationship between the predicted values ​​of each base model and the actual precipitation observation values, thereby achieving adaptive learning of the dynamic ensemble weights.

[0030] The ensemble layer employs a station-by-station hyperparameter search strategy to determine the optimal network structure. The hyperparameter search space of the ensemble layer's deep neural network is consistent with the hyperparameter search space of the deep neural network-like algorithms of the base model in step S3.

[0031] Step S5: Precipitation Forecast Output The predictions of each base model for future periods are input into the trained ensemble layer, which outputs the final precipitation prediction result.

[0032] Taking the precipitation forecast for the period from late July to early August from 102 national meteorological stations in Henan Province as an example.

[0033] Dataset: Climate factors: 89 factors (63 atmospheric circulation factors and 26 sea surface temperature factors), spanning from 1961 to 2022 (62 years), with each factor expanded into an annual characteristic vector on a monthly basis.

[0034] Precipitation data: Cumulative precipitation during the "late July to early August" period from 1961 to 2022 at 102 national meteorological stations in Henan Province.

[0035] Forecast target: The percentage of precipitation anomalies during the "late July to early August" period at each station from 2023 to 2025.

[0036] Step S1: The 89 climate factors were divided into three groups: atmospheric circulation factors (63 factors), sea surface temperature factors (26 factors), and comprehensive factors (all 89 factors). Each factor was further divided into 12 months: atmospheric factors (756 dimensions), sea surface temperature factors (312 dimensions), and comprehensive factors (1068 dimensions).

[0037] Step S2: Perform the following steps for each group of factors in sequence: Correlation screening: retain factors with an absolute correlation coefficient ≥0.10 (atmosphere, all) or ≥0.15 (sea surface temperature) with the station average precipitation.

[0038] Standardization: z-score standardization is performed on the screened factors.

[0039] PCA dimensionality reduction: retain principal components with a cumulative variance contribution rate of 85% (atmosphere) or 90% (sea surface temperature, all).

[0040] The dimensions of each factor group after processing are as follows: atmosphere 27 dimensions, sea surface temperature 12 dimensions, and all 30 dimensions.

[0041] Step S3: In this embodiment, three different types of machine learning algorithms are used: (a) Deep Neural Network (DNN): A multi-layer fully connected network with BatchNorm and Dropout regularization. The hyperparameter search space is: number of hidden layers {5,6,7,8,9,10} × number of neurons per layer {10,20,30,50,100} × Dropout rate {0.1,0.2,0.3}, for a total of 90 configurations. Each configuration is evaluated using 10-fold cross-validation, and the configuration with the minimum validation loss is selected as the optimal hyperparameter for that site. Training is performed for 200 epochs, with the learning rate decreasing from 0.01 cosine annealing to 0.0001. Gaussian noise is added to the training set during training for 3x data augmentation to alleviate overfitting.

[0042] (b) Random Forest (RF): The optimal value is selected from the maximum depth {3, 5, 7, unlimited} through cross-validation. An additional time exponential weight is applied to the sea surface temperature factor group to accommodate the non-stationarity of the sea surface temperature teleconnection relationship.

[0043] (c) Ridge regression: The optimal value of the regularization parameter α∈{0.01,0.1,1,10,100,1000} is selected through cross-validation, and the time exponential weight is also applied to the sea surface temperature factor group.

[0044] Three sets of factors × three algorithms = 9 base models. Each base model is trained independently for 102 sites and generates OOF predictions synchronously.

[0045] Step S4: The OOF predictions of the nine base models are concatenated column-wise to form a (62,9)-dimensional input matrix for the ensemble layer (62 years × 9 base models). A DNN ensemble layer is trained for each site, with the hyperparameter search space consistent with the base model DNN in step S3(a) (90 configurations). The optimal configuration is selected using 10-fold cross-validation. The training set is augmented 5-fold during ensemble layer training.

[0046] Step S5: The predictions for 2023-2025 from the nine base models are input into the ensemble layer model, and the final prediction results are output.

[0047] The application scenarios of the method of the present invention are not limited to the "July-August" precipitation forecast in Henan Province, but can be extended to the forecasting of precipitation or other meteorological elements (such as temperature, wind speed, etc.) in other regions and at other time scales (such as months, seasons, years).

[0048] In other embodiments, other algorithms may also be used in step S3, such as deep learning models like LSTM and CNN, tree models like XGBoost and LightGBM, and linear models like Lasso and ElasticNet.

[0049] The prediction method of this invention has significant technical advantages: (1) Heterogeneous integration is significantly better than homogeneous integration A base model is constructed using multiple different types of machine learning algorithms (heterogeneous ensemble). The error patterns of each algorithm are orthogonal, and the ensemble can effectively eliminate the systematic bias of a single algorithm.

[0050]

[0051] The experimental results show that each addition of a heterogeneous algorithm brings a positive improvement, and the 9-dimensional heterogeneous ensemble (PS=75.3) is significantly better than the 3-dimensional homogeneous ensemble (PS=73.4), with a PS improvement of 1.9 and an accuracy (ACC) improvement of 0.043.

[0052] (2) Enhance the ability to model spatial heterogeneity This invention independently searches for optimal hyperparameters for each meteorological station, enabling the model to adaptively match the precipitation-factor relationships of different stations. The optimal configurations selected from the 102 stations in the experiment were widely distributed, verifying the necessity of station-by-station search. Compared to a fixed uniform parameter scheme, the PS of the station-by-station search scheme increased from 60.4 to 75.3.

[0053] (3) Good interdecadal stability It has shown stable performance in cross-validation over a period of 62 years, from 1961 to 2022:

[0054] The PS score remained stable between 72 and 77 across different decadals, with no significant interdecadal drift, indicating that the model is robust to climate change.

[0055] (4) Strong ability to predict extreme precipitation events In 2021, an extreme precipitation year in Henan Province (the year of the "7.20" torrential rainstorm), the ACC of the method of this invention reached 0.894 and the PS reached 80.4, indicating that it has a strong predictive ability for the spatial distribution of extreme precipitation events.

[0056] The regional precipitation prediction system based on multi-source heterogeneous deep integration of the present invention includes: The data acquisition and grouping module is used to acquire multi-source climate factor data and historical precipitation observation data from multiple meteorological stations in the target area, and to divide the climate factors into at least two groups according to their physical attributes. The feature processing module, connected to the data acquisition and grouping module, is used to perform feature preprocessing on each group of climate factors; The heterogeneous base model construction module, connected to the feature processing module, is used to construct a base model for each group of preprocessed factors using at least two different types of machine learning algorithms, perform independent modeling and hyperparameter optimization on a site-by-site basis, and generate out-of-bounds predictions. The deep ensemble layer module, connected to the heterogeneous base model construction module, is used to train the deep neural network site by site, taking the out-of-fold predictions of all base models as input, to learn dynamic ensemble weights. The prediction output module, connected to the deep integration layer module, is used to input the base model's predictions for future periods into the trained integration layer and output the final precipitation prediction result.

[0057] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0058] The modules or units described in this invention can be implemented by software programs, hardware circuits, or a combination of both, and their specific implementation does not constitute a limitation on the scope of protection of this invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of this invention without departing from the spirit and substance of this invention, and all such modifications and substitutions should be covered within the scope of protection of this invention.

[0059] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A regional precipitation prediction method based on multi-source heterogeneous deep integration, characterized in that, Includes the following steps: Step S1: Obtain multi-source climate factor data and historical precipitation observation data from multiple meteorological stations in the target area. The multi-source climate factors include at least the atmospheric circulation index and the sea surface temperature index. The climate factors are divided into at least two groups according to their physical properties. Step S2: Perform feature preprocessing on each group of climate factors; Step S3: For each group of preprocessed factors, at least two different types of machine learning algorithms are used to build base models. Each meteorological station is modeled independently, and K-fold cross-validation is used to generate the out-of-fold prediction of each base model during the training process. That is, each training sample is generated by the model that has not used the sample for training, so as to obtain a leak-free prediction sequence of the same length as the training set. Step S4: Using the out-of-range predictions generated by all base models as input, train a deep neural network as an ensemble layer for each meteorological station to learn the optimal nonlinear mapping relationship between the predicted values ​​of each base model and the actual precipitation observation values, thereby achieving adaptive learning of dynamic ensemble weights. Step S5: Input the predicted values ​​of each base model for the future period into the trained ensemble layer, and output the final precipitation prediction result.

2. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S1, the climate factors are divided into atmospheric circulation factor group, sea surface temperature factor group and comprehensive factor group containing all factors according to their physical properties.

3. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S2, the preprocessing includes: correlation pre-screening, standardization processing, and principal component analysis dimensionality reduction.

4. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S3, the machine learning algorithms include deep neural network algorithms, tree model algorithms, and linear model algorithms. Among them, the deep neural network base model uses a station-by-station hyperparameter search strategy to determine the optimal hyperparameters, while the tree model and linear regression base models select the optimal hyperparameters for each station through cross-validation.

5. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S4, the integration layer uses a station-by-station hyperparameter search strategy to determine the optimal network structure.

6. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 5, characterized in that: The hyperparameter search space of the ensemble deep neural network is consistent with the hyperparameter search space of the deep neural network algorithm of the base model in step S3.

7. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: The station-by-station hyperparameter search space of the deep neural network algorithm includes the number of hidden layers, the number of neurons per layer, and the Dropout rate; the tree model algorithm selects the maximum depth through cross-validation; and the linear model algorithm selects the regularization parameter through cross-validation.

8. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S3, a time exponential weight is applied to the base model constructed from the sea surface temperature factor group.

9. The regional precipitation prediction method based on multi-source heterogeneous deep integration according to claim 1, characterized in that: In step S3, during the training of the base model, and in step S4, during the training of the ensemble layer, Gaussian noise is added to the training data for data augmentation to alleviate the overfitting problem under small sample conditions; the augmentation factor for the base model training data is 3 times, and the augmentation factor for the ensemble layer training data is 5 times.

10. A regional precipitation prediction system based on multi-source heterogeneous deep integration, characterized in that, include: The data acquisition and grouping module is used to acquire multi-source climate factor data and historical precipitation observation data from multiple meteorological stations in the target area, and to divide the climate factors into at least two groups according to their physical attributes. The feature processing module, connected to the data acquisition and grouping module, is used to perform feature preprocessing on each group of climate factors; The heterogeneous base model construction module, connected to the feature processing module, is used to construct a base model for each group of preprocessed factors using at least two different types of machine learning algorithms, perform independent modeling and hyperparameter optimization on a site-by-site basis, and generate out-of-bounds predictions. The deep ensemble layer module, connected to the heterogeneous base model construction module, is used to train the deep neural network site by site, taking the out-of-fold predictions of all base models as input, to learn dynamic ensemble weights. The prediction output module, connected to the deep integration layer module, is used to input the base model's predictions for future periods into the trained integration layer and output the final precipitation prediction result.

Images