Precipitation space reconstruction method, system and equipment based on terrain and weather multi-factor fusion driving and medium

Abstract

The invention relates to the technical field of meteorology and hydrology, in particular to a rainfall space reconstruction method, system and device based on terrain and meteorological multi-factor fusion driving and a medium, and the method comprises the steps: obtaining space terrain factors and meteorological observation data of a target area, and employing a BP neural network to progressively interpolate missing measurement values of meteorological driving variables in a layered manner; setting a missing detection elimination rule based on a wet season and a dry season to process precipitation data; constructing a 14-dimensional high-dimensional input feature system fusing a space terrain factor, a time sequence factor and a meteorological driving factor; nonlinear models such as an XGBoost model, a BP neural network model or an LSTM model are used for training, and finally the monthly scale precipitation space reconstruction of the grid is achieved. According to the method, the problems that a traditional interpolation method is poor in adaptability in a complex terrain area and insufficient in multi-factor driving relation description are effectively solved, and the precision and the physical consistency of precipitation space distribution are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological hydrology, in particular to a precipitation spatial reconstruction method, system, device and medium based on terrain and meteorological multi-factor fusion driving. BACKGROUND

[0002] At present, the precipitation spatial interpolation method has important application value in the fields of meteorology, hydrology, ecology and resource regulation, and is commonly used to convert the observation data of discrete meteorological sites into continuous spatial precipitation distribution to support regional scale hydrological simulation, climate assessment and scheduling decision.

[0003] However, the existing technology still faces significant challenges in the application of complex terrain areas. On the one hand, traditional interpolation methods (such as inverse distance weighting, Kriging, multiple regression, etc.) are mostly based on the assumption of spatial continuity and stationarity, and are difficult to effectively represent the abruptness and nonlinear distribution characteristics of precipitation in areas with dramatic terrain changes, often leading to unnatural spatial transition and significant estimation error. On the other hand, precipitation processes are driven by terrain (such as altitude, elevation, latitude and longitude) and meteorological factors (such as temperature, pressure, humidity, and sunshine duration), showing a highly nonlinear and strong interactive response mechanism. Traditional methods have limited ability in multi-factor fusion modeling, making it difficult to accurately depict these complex driving relationships, resulting in insufficient model interpretation and generalization ability.

[0004] In addition, in complex terrain areas such as plateaus and mountains, due to the sparse and uneven distribution of meteorological stations, traditional methods are prone to the "isolated point dominance" effect, making the interpolation results highly sensitive to individual station abnormal observations, with poor spatial consistency and error amplification. At the same time, existing high-resolution precipitation data products (including remote sensing inversion and reanalysis data) generally face problems such as high uncertainty, significant terrain shielding effect, and obvious systematic bias in complex terrain areas, which cannot meet the needs of high-precision modeling and fine-grained climate research. In summary, the current precipitation interpolation technology generally has the problems of insufficient spatial distribution precision, weak multi-factor integration capability, poor terrain adaptability, and low reliability of high-resolution data in complex terrain areas. SUMMARY

[0005] The present application aims to provide a precipitation spatial reconstruction method, system, device and medium based on terrain and meteorological multi-factor fusion driving, which overcomes the defects of poor adaptability of traditional interpolation methods in terrain mutation areas and insufficient description of multi-factor driving relationships in the prior art.

[0006] The present application achieves the above-mentioned purpose through the following technical solutions: In a first aspect, the present application provides a precipitation spatial reconstruction method based on terrain and meteorological multi-factor fusion driving, which comprises: Obtaining meteorological observation data corresponding to spatial terrain factors and time series factors in a target area meteorological station, the meteorological observation data including meteorological factors and precipitation, and the meteorological factors including daily / montly scale air temperature, atmospheric pressure, relative humidity and sunshine duration; Determining meteorological driving variables according to the data interpolated by the hierarchical progressive method based on the meteorological factors, and determining target variables after processing the precipitation according to a preset missing data elimination rule, the missing data elimination rule being formulated based on the corresponding wet period and dry period; Establishing a grid input feature system based on the spatial terrain factors, time series factors and meteorological driving variables of the target area meteorological station, and forming a high-dimensional input feature vector; Training a pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variables, the reconstruction model being established based on a nonlinear fitting algorithm; Reconstructing the monthly scale precipitation of a meteorological station in a to-be-measured area by using the trained reconstruction model, and outputting the grid monthly scale precipitation.

[0007] Further, the determining of the meteorological driving variables according to the data interpolated by the hierarchical progressive method based on the meteorological factors includes: The missing data of temperature, atmospheric pressure, relative humidity and sunshine duration in the daily meteorological factors are interpolated by using a hierarchical progressive reconstruction method of BP neural network, the missing values of atmospheric pressure, relative humidity and sunshine duration are reconstructed in turn through a multi-layer perception structure, and the to-be-processed data after the missing data is recovered are formed; The to-be-processed data is monthly summarized to generate monthly scale meteorological driving variables, including: The average air temperature, average atmospheric pressure and average relative humidity determined based on the daily average values of the effective observation days in the month; The monthly maximum air temperature, monthly minimum air temperature, atmospheric pressure maximum value, atmospheric pressure minimum value and relative humidity maximum value determined based on the maximum values or minimum values of the effective observation days in the month; The monthly cumulative sunshine duration determined based on the summation of the sunshine durations of the effective observation days in the month.

[0008] Further, the determining of the target variables after the processing of the precipitation according to the preset missing data elimination rule includes: Determining the missing data days of the monthly precipitation when the regional meteorological station is in the wet period or the dry period; If the missing data days of the monthly precipitation in the wet period are greater than or equal to a first threshold value, or if the missing data days of the monthly precipitation in the dry period are greater than or equal to a second threshold value, the monthly precipitation sample is marked as missing; The precipitation of the effective observation days in the month whose missing data days do not exceed the corresponding threshold value is accumulated to obtain the monthly cumulative precipitation and serve as the target variable.

[0009] Further, the grid input feature system is established based on the spatial terrain factor, time series factor and meteorological driving variable in the target area weather station, forming a high-dimensional input feature vector, including: A grid system is constructed with a set spatial resolution, covering the target area weather station. The observation data of each weather station in each grid cell i is extracted, defined as the input sample as a high-dimensional input feature vector : : The spatial terrain factor is determined based on the static attributes of the target area weather station within a set period , wherein, represents the latitude of the weather station j, represents the longitude of the weather station j, represents the altitude of the weather station j; The time series factor is determined based on the acquisition time of the meteorological observation data , wherein, represents the year; represents the month; The meteorological driving factor is determined based on the monthly scale meteorological driving variable of the meteorological observation data , wherein, represents the monthly average temperature, , respectively represent the maximum / minimum value of the daily maximum / minimum temperature in the month, represents the monthly average air pressure, , respectively represent the maximum / minimum value of the daily air pressure in the month, represents the monthly average relative humidity, represents the maximum value of the daily average relative humidity in the month, represents the monthly cumulative sunshine duration.

[0010] Further, before training the pre-constructed reconstruction model, the method further comprises: determining the cumulative precipitation in the target area weather station ; The cumulative precipitation and the high-dimensional input feature vector are constructed as a sample set , , respectively as the start and end year, m unit as month; Based on the set division rule, the stations in the sample set are divided into training weather stations​ and verifying meteorological stations , as follows: .

[0011] Further, the set division rule comprises: the verification meteorological stations are distributed within a preset elevation range and a latitude and longitude gradient; the number of the training meteorological stations accounts for a first proportion value of the total number of stations, and the training meteorological stations are uniformly distributed in different climate zones; the number of the verification meteorological stations accounts for a second proportion value of the total number of stations, and the second proportion value is less than the first proportion value.

[0012] Further, the nonlinear fitting algorithm comprises an XGBoost model, a BP neural network model, and an LSTM model.

[0013] In a second aspect, the present application provides a precipitation spatial reconstruction system based on terrain and meteorological multi-factor fusion driving, which is used to implement the precipitation spatial reconstruction method as described above, and the system comprises: a data acquisition module, configured to acquire spatial terrain factors and time series factors corresponding to meteorological observation data in a target area meteorological station, the meteorological observation data comprising meteorological factors and precipitation, and the meteorological factors comprising daily / montly scale air temperature, atmospheric pressure, relative humidity, and sunshine duration; a data processing module, configured to determine meteorological driving variables according to the meteorological factors and the data after layered progressive interpolation, and determine target variables after processing the precipitation according to a preset missing data elimination rule, the missing data elimination rule being formulated based on corresponding wet period and dry period; a feature construction module, configured to establish a grid input feature system based on the spatial terrain factors, the time series factors, and the meteorological driving variables in the target area meteorological station, and form a high-dimensional input feature vector; a model training module, configured to train a pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variables, the reconstruction model being established based on a nonlinear fitting algorithm; a precipitation reconstruction module, configured to perform spatial reconstruction on the monthly scale precipitation of a to-be-measured area meteorological station by using the trained reconstruction model, and output grid monthly scale precipitation.

[0014] In a third aspect, the present application provides an electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the precipitation spatial reconstruction method as described above when executing the computer program.

[0015] In a fourth aspect, the present application provides a computer readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the precipitation spatial reconstruction method described above.

[0016] The present application has the following beneficial effects: 1. The present application fuses terrain factors and multi-source meteorological driving factors to construct a unified high-dimensional input feature system, integrates spatial terrain properties, time series information and meteorological driving variables, forms a 14-dimensional feature vector, and can more comprehensively capture the multi-scale driving mechanism of precipitation formation. At the same time, the present application adapts to various nonlinear modeling algorithms such as XGBoost, BP neural network and LSTM, effectively solves the problems of poor adaptability and unnatural spatial transition of traditional methods in complex terrain areas through the powerful feature learning ability of deep learning. The method is particularly suitable for sparse areas such as plateau and mountainous areas, and can stably generate high-resolution precipitation grid data, providing reliable data support for hydrological simulation, climate assessment and other applications.

[0017] 2. The present application effectively solves the modeling problem caused by missing meteorological observation data through a hierarchical and progressive BP neural network interpolation method and a missing data elimination rule based on wet period / dry period. The systematic preprocessing process ensures the quality and consistency of the input data, laying a solid foundation for model training. In the model deployment stage, the present application supports 1km x 1km high-resolution grid precipitation reconstruction, and the output results can be converted into multiple standard formats such as GeoTIFF and NetCDF, facilitating subsequent business system integration and application. Compared with traditional interpolation techniques, the present scheme significantly improves the prediction stability and accuracy in complex terrain areas while maintaining computational efficiency, and is particularly suitable for large-scale, multi-year monthly precipitation reconstruction tasks, providing a powerful technical tool for regional climate research and water resource management. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 FIG. 1 is a flowchart of a precipitation spatial reconstruction method based on terrain and meteorological multi-factor fusion driving according to an embodiment of the present application; Figure 2 FIG. 2 is another flowchart of a precipitation spatial reconstruction method based on terrain and meteorological multi-factor fusion driving according to an embodiment of the present application; Figure 3 FIG. 3 is a technical roadmap of a data preprocessing step according to an embodiment of the present application; Figure 4 FIG. 4 is a technical roadmap of a high-dimensional feature construction step according to an embodiment of the present application; Figure 5This is a schematic diagram of the structure of a precipitation spatial reconstruction system driven by the fusion of topography and meteorological multi-factors, according to an embodiment of this application. Figure 6 This is a schematic diagram of the structure of an electronic device according to one embodiment of this application; Figure 7 This is a schematic diagram illustrating an implementation example of this application, based on measured precipitation data from 95 independent verification meteorological stations nationwide. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0020] In one specific embodiment, such as Figure 1 , 2 As shown in Figures 3 and 4, this application proposes a precipitation spatial reconstruction method driven by the fusion of topographic and meteorological multi-factors. This method is suitable for accurate interpolation and modeling of high-resolution monthly precipitation grid data under different topographic and climatic conditions. The specific method includes the following steps: S1: Obtain the spatial topographic factors and corresponding time series factors of the meteorological station in the target area. The meteorological observation data includes meteorological factors and precipitation. The meteorological factors include daily / monthly scale temperature, atmospheric pressure, relative humidity and sunshine duration.

[0021] In step S1, the preparation and preprocessing of meteorological observation data includes: The original meteorological data used in this application comes from the Meteorological Bureau's shared meteorological station network, covering a specified time period, such as 1961-2019, and spatially encompassing multiple meteorological stations, demonstrating broad geographical representativeness and climatic diversity. The data includes daily-scale conventional meteorological elements such as temperature, air pressure, humidity, sunshine duration, and precipitation.

[0022] In practice, during long-term meteorological observations, a certain proportion of missing data and outlier records exist due to objective reasons such as equipment failure, extreme meteorological events, and abnormal data transmission. This application systematically preprocesses the data before modeling, employing differentiated strategies for different types of variables, as detailed in the following steps.

[0023] S2: Determine meteorological driving variables based on the data after hierarchical progressive interpolation of the meteorological factors; and determine the target variable after processing the precipitation according to the preset missing data removal rules, wherein the missing data removal rules are formulated based on the correspondence between the wet season and the dry season.

[0024] In a preferred embodiment, step S2 determines the meteorological driving variable according to the meteorological factor layer-by-layer progressive interpolation of the data after interpolation, including: for the missing data of temperature, air pressure, relative humidity and sunshine duration in the meteorological factor, using the layer-by-layer progressive reconstruction method of BP neural network to interpolate, and through the multi-layer perception structure to sequentially reconstruct the missing values of air pressure, relative humidity and sunshine duration, to form the processed data after recovery of the missing data. The processed data is monthly summarized to generate the monthly scale meteorological driving variable, including: the average temperature, average air pressure and average relative humidity determined based on the daily average value of the effective observation day in the month; the monthly maximum temperature, monthly minimum temperature, maximum air pressure, minimum air pressure and maximum relative humidity determined based on the maximum value or minimum value of the effective observation day in the month; and the monthly cumulative sunshine duration determined based on the summation of the sunshine duration of the effective observation day in the month.

[0025] In a preferred embodiment, step S2 determines the target variable after processing the precipitation according to the preset missing data elimination rule, including: determining the monthly precipitation missing days of the regional meteorological station in the wet period or dry period; if the monthly precipitation missing days in the wet period are greater than or equal to a first threshold, or if the monthly precipitation missing days in the dry period are greater than or equal to a second threshold, the monthly precipitation sample is marked as missing; and the precipitation of the effective observation day whose missing days in the month do not exceed the corresponding threshold is accumulated to obtain the monthly cumulative precipitation and serve as the target variable.

[0026] It should be noted that the precipitation is the target variable of the present application, and if it is artificially filled before model training, it may introduce unrealistic data noise or pseudo signal, interfere with the learning process of the model on the precipitation generation mechanism, and affect the reliability of the prediction performance. Therefore, the present application only determines whether the sample participates in modeling according to the preset missing data elimination rule in the process of converting daily data to monthly cumulative data. Regarding the preset missing data elimination rule, an example is as follows: Wet period (May to October): Precipitation is frequent and strong. If the missing days in the month are greater than or equal to 10 days, the precipitation record is considered incomplete, and the sample in the month is marked as missing; Dry period (November to next April): Precipitation is sparse. If the missing days in the month are greater than or equal to 15 days, the precipitation data in the month is considered unavailable; If the missing days do not exceed the corresponding threshold, the precipitation values of all effective observation days in the month are accumulated to calculate the monthly precipitation.

[0027] Only the samples with available target variables are retained to participate in model training to construct a supervised learning data set. The samples marked as missing do not participate in model training, but after the model is constructed, the present application can call the trained reconstruction model to combine the corresponding input features to perform post-prediction on the precipitation in the missing months, to realize high-precision filling of the historical observation records.

[0028] S3: Establish a grid input feature system based on the spatial terrain factors, time series factors and meteorological driving variables of the target area weather station, and form a high-dimensional input feature vector.

[0029] To achieve fine modeling of the spatiotemporal distribution of precipitation, the present application systematically integrates spatial terrain attributes, time series information and meteorological driving variables to establish a unified high-dimensional input feature system. For example, let the weather station number be j∈{1, 2, …, M}, the time range cover y∈{1961, …, 2019}, and each year contain m∈{1, …, 12} monthly records.

[0030] In a preferred embodiment, step S3 establishes a grid input feature system based on the spatial terrain factors, time series factors and meteorological driving variables of the target area weather station, and forms a high-dimensional input feature vector, including: A grid system is constructed at a set spatial resolution, covering the target area weather station. For each grid cell i in the grid system, the observation data of each weather station in any year and month are extracted, defined as the input sample as a high-dimensional input feature vector : : The spatial terrain factors are determined based on the static attributes of the target area weather station within a set period , wherein represents the latitude of the weather station j, represents the longitude of the weather station j, represents the altitude of the weather station j; The time series factors are determined based on the acquisition time of the meteorological observation data , wherein represents the year; represents the month; The meteorological driving factors are determined based on the monthly scale meteorological driving variables of the meteorological observation data , wherein represents the monthly average temperature, , represent the maximum / minimum value of the daily maximum / minimum temperature in the month, represents the monthly average air pressure, , represent the maximum / minimum value of the daily air pressure in the month, represents the monthly average relative humidity, represents the maximum value of the daily average relative humidity in the month, represents the monthly cumulative sunshine duration.

[0031] It can be understood that the set of factors is the main meteorological driving variable affecting the formation and evolution of precipitation, reflecting the temperature, humidity, pressure conditions and light time characteristics on a monthly scale.

[0032] In a preferred embodiment, before training the pre-constructed reconstruction model, the method further comprises: determining the cumulative precipitation of the target area weather station ; cumulative precipitation and the high-dimensional input feature vector are constructed into a sample set , , respectively, the start and end years, m units are months; based on the set division rule, the stations in the sample set are divided into training weather stations and verification weather stations , as follows: .

[0033] In a preferred embodiment, the set division rule comprises: the verification weather stations are distributed within the preset elevation range and latitude and longitude gradient; the proportion of the number of training weather stations to the total number of stations is a first proportion value, and the training weather stations are evenly distributed in different climate zones; the proportion of the number of verification weather stations to the total number of stations is a second proportion value, and the second proportion value is less than the first proportion value.

[0034] Correspondingly, the selection of the verification weather station follows the following principles in specific implementation: covering as wide an elevation range and latitude and longitude gradient as possible to ensure the adaptability of the model under different terrain and climate conditions. Considering that the weather stations in high-altitude areas are relatively sparse, in order to ensure sufficient sample support during model training, the present application preferentially uses most of the available measured data for training, and the number of verification stations can be appropriately controlled. Under the premise of not affecting the representativeness of the evaluation, the balance between data utilization efficiency and verification effect is considered.

[0035] S4: training the pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variable, the reconstruction model being established based on a nonlinear fitting algorithm.

[0036] More specifically, the model is trained based on the training set. In order to improve the adaptability of the model to non-normal distribution, zero value and negative value data, the Yeo-Johnson method is uniformly used to standardize the features and target variables: wherein and is a standardization transformation function fitted according to sample data in the training phase. The transformation can process sample data containing zero or negative values, and is suitable for the feature distribution characteristics of meteorological variables in complex terrain areas.

[0037] In the training phase, the above input standardizer and the target variable reverse standardizer are saved as independent files to ensure that the corresponding standardizer files can be directly called in the subsequent model prediction and inference phase, and any grid location or station data is consistently preprocessed, thereby improving the physical comparability and model reproducibility of the reconstruction results.

[0038] In a preferred embodiment, the nonlinear fitting algorithm includes an XGBoost model, a BP neural network model, and an LSTM model.

[0039] After completing the training sample construction and standardization processing, based on the unified terrain-meteorological-time input feature system, a variety of models with nonlinear modeling capability are used to train and spatially reconstruct monthly precipitation. The present application adapts to multiple regression modeling paths, including the ensemble learning model XGBoost, the BP neural network (Back Propagation neural network), and the long short-term memory network (LSTM), to enhance the adaptability and prediction accuracy of the model under complex terrain and multiple climate conditions.

[0040] In the model training phase, all training samples are divided into a training set and a validation set in a 95%:5% ratio. The training set is used for model parameter fitting, and the validation set is used to monitor the generalization ability of the model to ensure that the structure and hyperparameter combination have good stability and migration performance. After division, the training and validation sets are saved respectively for subsequent tuning and accuracy evaluation.

[0041] In the XGBoost model, first, the training set and validation set data are converted into DMatrix format, and the model hyperparameters are set, including the maximum tree depth (such as 12), the learning rate (such as 0.06), the minimum subsample weight (such as 7), the subsample ratio and the column sample ratio (such as 0.9), the L1 regularization coefficient (such as 0.8), and the L2 regularization coefficient (such as 8.0). In the training process, an early stopping mechanism is used to control overfitting, and the training is automatically terminated when the validation set error does not significantly decrease for a certain number of consecutive rounds (such as 200 rounds). At the same time, the validation error and the optimal number of iterations in the training process are recorded to obtain the model version with the optimal validation performance.

[0042] In the BP neural network model, a multi-layer feedforward network structure is constructed, and the typical structure includes 3-5 hidden layers, and the number of single-layer neurons can be set to 256, 128, 64, etc. Each hidden layer uses a LeakyReLU or ReLU activation function, and the output layer is a linear node. The model uses mean square error (MSE) as the loss function, uses the Adam optimizer for backpropagation training, and introduces an early stopping strategy and a learning rate adaptive mechanism during the training process to avoid overfitting and oscillation training phenomena.

[0043] In the LSTM model, the normalized input features are reshaped into a three-dimensional tensor structure (sample number × time step × feature number) to capture the precipitation time series features. The model structure is a multi-layer stacked LSTM network, and the number of hidden units is set to 128, 64, 32, and the output layer is a fully connected node. During the training process, the performance monitoring of the validation set, the learning rate adjustment, and the optimal model weight saving strategy are also used to improve the robustness and generalization ability of the training.

[0044] In implementation, during the training process of all models, correlation coefficient (CC), mean error (ME), mean absolute error (MAE), and root mean square error (RMSE) are used as evaluation indicators to quantitatively evaluate the prediction effect of the training set and the validation set. The evaluation results include: CC represents the linear correlation between the predicted value and the measured value; ME reflects the systematic deviation of the prediction error; MAE measures the average absolute error amplitude; RMSE comprehensively reflects the overall error level of the model.

[0045] After training each group of models, the system automatically compares the evaluation results corresponding to different hyperparameter combinations, and retains the model parameter configuration and training weight with the best validation set performance. The final output includes: (1) the trained model file (such as.model,.h5); (2) the standardizer model file (features and target variables are saved separately); (3) the optimal hyperparameter configuration file (such as.json or.txt); (4) the training-validation result log (including error evaluation indicators).

[0046] S5: Use the trained reconstruction model to reconstruct the monthly precipitation of the meteorological station in the test area, and output the gridded monthly precipitation.

[0047] More specifically, after the model training is completed, the performance-optimal model is deployed in the precipitation spatial reconstruction system, combined with the saved input standardizer and target variable denormalizer , the study area is reconstructed by grid precipitation.

[0048] S5.1 Grid construction and feature extraction A spatial grid system with a spatial resolution of 1km x 1km resolution is constructed, and the total number of grids is .

[0049] 1) Spatial terrain factors (extracted from digital elevation model DEM): : Latitude of the i-th grid cell : Longitude of the i-th grid cell : Altitude of the i-th grid cell (m).

[0050] 2) Construction of time factors: 3) Meteorological driving factors (provided by the reconstructed monthly scale grid weather product): : Monthly average temperature of the i-th grid cell (℃); • , : Maximum / minimum value of daily maximum / minimum temperature in the i-th grid cell (℃) in the month; : Monthly average air pressure of the i-th grid cell (hPa); • , : Maximum / minimum value of daily air pressure in the i-th grid cell (hPa) in the month; • : Monthly average relative humidity of the i-th grid cell (%); • : Maximum value of daily average relative humidity in the i-th grid cell (%) in the month; : Monthly cumulative sunshine duration of the i-th grid cell (h).

[0051] The complete input feature vector of the i-th grid, year y, and month m is constructed as: S5.2 Standardization and model prediction The above grid feature input vector is standardized by the standardizer saved in the training stage Convert: The standardized features are input into the prediction model, and the optimal model parameter file is called to calculate the precipitation estimate (standardized form): The model output results are de-standardized to restore to the original physical quantity scale (unit: mm): S5.3 Product Output and Format Conversion The final output results are organized into a grid format product, supporting export in multiple standard formats: raster image format (GeoTIFF); scientific computing format (NetCDF); table format (CSV).

[0052] In specific implementation, the application further includes step S6: verification of the accuracy of the grid precipitation product based on the verification meteorological data.

[0053] To evaluate the accuracy and physical consistency of the high-resolution monthly precipitation grid data product generated by the system in spatial prediction, independent meteorological station measured precipitation data (i.e. verification dataset ) is used for verification analysis.

[0054] The specific method is: from the reconstructed precipitation product, the grid point prediction value corresponding to the geographical location of each verification meteorological station is extracted, and the time series pairing is carried out with the measured monthly precipitation data of the station, and statistical indicators are used for accuracy analysis. The evaluation indicators selected include: correlation coefficient (CC), mean absolute error (MAE), mean error (ME) and root mean square error (RMSE).

[0055] In a specific embodiment, as Figure 5 shown, the application proposes a precipitation spatial reconstruction system based on the fusion of terrain and meteorological multi-factors, for implementing the precipitation spatial reconstruction method proposed in the above embodiments, the system includes: A data acquisition module 10 is configured to acquire spatial terrain factors and time series factors corresponding to meteorological observation data in a target area meteorological station, the meteorological observation data including meteorological factors and precipitation, and the meteorological factors including daily / montly scale air temperature, atmospheric pressure, relative humidity and sunshine duration; A data processing module 11 is configured to determine meteorological driving variables according to the data after stratified progressive interpolation of the meteorological factors; and determine target variables after processing the precipitation according to a preset missing data removal rule, the missing data removal rule being formulated based on the corresponding wet period and dry period; a feature construction module 12, configured to establish a grid input feature system based on spatial terrain factors, time series factors and meteorological driving variables of the target area weather station, and form a high-dimensional input feature vector; a model training module 13, configured to train a pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variable, the reconstruction model being established based on a nonlinear fitting algorithm; a precipitation reconstruction module 14, configured to use the trained reconstruction model to perform spatial reconstruction on the monthly scale precipitation of the weather station in the to-be-tested area, and output the grid monthly scale precipitation.

[0056] The specific limitations of the precipitation spatial reconstruction system based on the fusion driving of terrain and meteorological multi-factors can refer to the limitations of the precipitation spatial reconstruction method based on the fusion driving of terrain and meteorological multi-factors in the foregoing, and will not be repeated here. It should be noted that the modules in the above precipitation spatial reconstruction system correspond to steps S1 to S5 in the implementation of the above precipitation spatial reconstruction method, and the instances and application scenarios realized by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above embodiments.

[0057] The following will combine an embodiment to perform precipitation reconstruction modeling and verification application based on the above precipitation spatial reconstruction method.

[0058] To verify the applicability and precision performance of the method of the present application at the national scale, the daily observation data of a certain meteorological bureau shared weather station network from 1961 to 2019 was selected, and a monthly precipitation sample set was constructed by summarizing. The data covers a total of 2386 weather stations in the national range, and the terrain types cover plain, hill, basin, mountain and plateau regions, which are widely representative.

[0059] In the experimental design, all the stations are divided into a training set and a verification set, wherein 95 meteorological stations with wide spatial distribution, sufficient altitude and latitude coverage are selected as independent verification stations, and the rest are used as model training samples. The model input features include: altitude, high-precision longitude and latitude, month, year, monthly average temperature, maximum / minimum daily temperature, monthly average pressure, maximum / minimum daily pressure, relative humidity and sunshine duration, etc. A total of 11 space-meteorological driving factors, and the output target is the cumulative precipitation (unit: mm) of the corresponding month.

[0060] The method of the present application is respectively adapted to three types of nonlinear modeling paths for precipitation reconstruction training: XGBoost, BP neural network (feedforward multilayer perceptron) and LSTM (stacked recurrent neural network), and a unified input feature system is used, and the Yeo-Johnson method is used for standardization processing.

[0061] The optimal parameter results of each model training stage are as follows: 1. XGBoost model training and validation results: Training set: CC = 0.85, ME = 74mm, MAE = 74mm, RMSE = 115mm Test set: CC = 0.83, ME = 72mm, MAE = 72mm, RMSE = 109mm 2. LSTM model training and validation results: Training set: CC = 0.83, ME = 70mm, MAE = 70mm, RMSE = 104mm Validation set: CC = 0.83, ME = 70mm, MAE = 70mm, RMSE = 105mm 3. BP neural network model training and validation results: Training set: CC = 0.82, ME = 72mm, MAE = 73mm, RMSE = 113mm Validation set: CC = 0.82, ME = 73mm, MAE = 72mm, RMSE = 112mm Figure 7 The comprehensive evaluation results of the three types of precipitation spatial reconstruction models (XGBoost, BP neural network, and LSTM) used in the invention are shown based on the measured precipitation data of 95 independent verification meteorological stations nationwide. From the evaluation results, the XGBoost model has the lowest median error in both RMSE and MAE, showing better precipitation reconstruction accuracy; the BP neural network and LSTM models have similar overall performance in the three indicators, achieving high correlation and small error amplitude. Overall, the three models have good prediction results under the support of a unified input feature system, effectively adapting to the monthly precipitation reconstruction requirements under complex terrain and multiple climate conditions, fully verifying the practical feasibility and promotional value of the invention in multi-model adaptability and nonlinear relationship modeling.

[0062] The error indicators obtained during the model training and validation process, as well as the spatial precision evaluation results based on the measured data of 95 independent meteorological stations, all indicate that the unified input feature system constructed in the invention has good fitting precision and spatial generalization ability under various modeling frameworks such as XGBoost, BP neural network, and LSTM. This method can stably output precipitation prediction results in national scale, areas with steep terrain and sparse meteorological observations, fully verifying the universality, stability, and promotional application value of the invention in high-resolution precipitation spatial reconstruction tasks.

[0063] In one embodiment, as shown in Figure 6 A computer device 21 is provided, which can be a terminal or a server, and the computer device 21 comprises a processor 23, a memory 24 and a network interface 25 connected by a system bus 22. The processor 23 of the computer device 21 is configured to provide computing and control capabilities. The memory of the computer device 24 comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface 26 of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement the steps of the precipitation spatial reconstruction method in any of the above embodiments.

[0064] In one embodiment, a computer readable storage medium is provided, and when the instructions in the computer readable storage medium are executed by the processor 23 of the computer device 21, the computer device 21 is enabled to perform the steps of the precipitation spatial reconstruction method in any of the above embodiments.

[0065] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. A precipitation spatial reconstruction method based on the fusion of topographic and meteorological multi-factors, characterized in that, The method comprises: acquiring meteorological observation data corresponding to spatial terrain factors and time series factors in a target area meteorological station, the meteorological observation data including meteorological factors and precipitation, and the meteorological factors including daily / mensual scale air temperature, atmospheric pressure, relative humidity and sunshine duration; determining meteorological driving variables according to the meteorological factors after layered progressive interpolation, and determining target variables after processing the precipitation according to a preset missing data removal rule, the missing data removal rule being formulated based on the corresponding wet period and dry period; establishing a grid input feature system based on the spatial terrain factors, time series factors and meteorological driving variables of the target area meteorological station, and forming a high-dimensional input feature vector; training a pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variables, the reconstruction model being established based on a nonlinear fitting algorithm; using the trained reconstruction model to spatially reconstruct mensual scale precipitation of a to-be-measured area meteorological station, and outputting grid mensual scale precipitation.

2. The method according to claim 1, wherein the method is characterized by The method further comprises: using a layered progressive reconstruction method of BP neural network to interpolate missing data of temperature, atmospheric pressure, relative humidity and sunshine duration in daily meteorological factors, sequentially reconstructing missing values of atmospheric pressure, relative humidity and sunshine duration through a multi-layer perception structure to form to-be-processed data after missing data is recovered; monthly summarizing the to-be-processed data to generate mensual scale meteorological driving variables, including: determining average air temperature, average atmospheric pressure and average relative humidity based on daily average values of effective observation days in the month; determining monthly maximum air temperature, monthly minimum air temperature, atmospheric pressure maximum value, atmospheric pressure minimum value and relative humidity maximum value based on maximum values or minimum values of effective observation days in the month; determining monthly cumulative sunshine duration based on summation of sunshine durations of effective observation days in the month.

3. The method according to claim 2, wherein the method is characterized by The method further comprises: determining missing data days of monthly precipitation when the meteorological station is in the wet period or the dry period; if the missing data days of monthly precipitation in the wet period are greater than or equal to a first threshold value, or if the missing data days of monthly precipitation in the dry period are greater than or equal to a second threshold value, marking the monthly precipitation sample as missing; accumulating precipitation of effective observation days in the month whose missing data days are not greater than the corresponding threshold value to obtain monthly cumulative precipitation as the target variable.

4. The method according to claim 3, wherein the method is characterized by The method further comprises: A grid system is constructed with a set spatial resolution, covering the target area weather stations, and the observation data of each weather station in each grid cell i is extracted for any year and month as the input sample is defined as a high-dimensional input feature vector : : Determining spatial terrain factors based on static attributes of target area weather stations within a set period , wherein, denotes the latitude of weather station j, denotes the longitude of weather station j, denotes the altitude of weather station j; Determining time series factors based on acquisition time of meteorological observation data , wherein, represents the year; represents the month; Determining weather driving factors based on meteorological observation data of monthly scale weather driving variables , wherein, represents the monthly mean air temperature, , represents the maximum / minimum value of the daily maximum / minimum air temperature within the month, represents the monthly mean air pressure, , represents the maximum / minimum value of the daily maximum / minimum air pressure within the month, represents the monthly mean relative humidity, represents the maximum value of the daily mean relative humidity within the month, represents the monthly cumulative sunshine duration.

5. The method according to claim 4, wherein, before training the pre-constructed reconstruction model, the method further comprises: Determining cumulative precipitation at a target area weather station ; accumulated precipitation and the high-dimensional input feature vector constructed as a sample set , , respectively are the start and end years, m units are months The sample set is divided into training weather stations and verification weather stations based on a set division rule as follows:​​ 。 6. The method according to claim 5, wherein, the set division rule includes: verifying that the meteorological stations are distributed in a preset elevation range and latitude / longitude gradient; training the proportion of the number of meteorological stations to the total number of stations to be a first proportion value, and the meteorological stations being uniformly distributed in different climate zones; verifying that the proportion of the number of meteorological stations to the total number of stations is a second proportion value, the second proportion value being less than the first proportion value.

7. The method according to claim 1, wherein, The nonlinear fitting algorithm includes an XGBoost model, a BP neural network model and an LSTM model.

8. A precipitation spatial reconstruction system based on the fusion of topographic and meteorological multi-factors driving, characterized in that, A system for implementing the precipitation spatial reconstruction method of any one of claims 1-7, the system comprising: a data acquisition module configured to acquire spatial terrain factors of meteorological stations in a target region, meteorological observation data corresponding to time series factors, the meteorological observation data including meteorological factors and precipitation, and the meteorological factors including daily / monthly scale air temperature, atmospheric pressure, relative humidity, and sunshine duration; a data processing module configured to determine meteorological driving variables based on the meteorological factors and the data after layered progressive interpolation, and to determine target variables after processing the precipitation according to a preset missing data removal rule, the missing data removal rule being formulated based on corresponding wet period and dry period; a feature construction module configured to establish a grid input feature system based on the spatial terrain factors of the meteorological stations in the target region, the time series factors, and the meteorological driving variables, and to form a high-dimensional input feature vector; a model training module configured to train a pre-constructed reconstruction model in combination with the high-dimensional input feature vector and the target variables, the reconstruction model being established based on a nonlinear fitting algorithm; a precipitation reconstruction module configured to perform spatial reconstruction on monthly scale precipitation of meteorological stations in a to-be-measured region using the trained reconstruction model, and to output grid monthly scale precipitation.

9. An electronic device, comprising: A computer readable storage medium having stored therein a computer program, the computer program being executed by a processor to implement the steps of the precipitation spatial reconstruction method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program stored in a computer readable storage medium, the computer program being executed by a processor to implement the steps of the precipitation spatial reconstruction method of any one of claims 1-7.

Images