Daily rainfall downscaling simulation method based on deep residual network

Abstract

The invention provides a daily rainfall downscaling simulation method based on a deep residual network, and aims to overcome the defects of the traditional rainfall simulation technology in the aspects of time and space precision, in particular to cope with challenges brought by climate changes and extreme weather. By introducing the deep learning technology and utilizing the strong feature learning capability and the nonlinear mapping capability, the accuracy and the meticulous degree of rainfall simulation are remarkably improved. The method comprises the steps of data preprocessing, network model construction and training and high-resolution daily rainfall data prediction simulation. According to the method, the space-time precision of rainfall simulation under the climate change condition can be effectively improved, and a scientific basis and practical support are provided for water resource management, urban drainage design and ecological environment evaluation.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross - application of hydrological climate change and artificial intelligence, and particularly relates to a daily rainfall downscaling simulation method based on a deep residual network. Background Art

[0002] In current climate change research, global climate models (GCMs) provide important large - scale climate information. However, their spatial resolution is usually between 0.5° and 4°, which limits their application at the urban or regional level. To meet the needs of regional climate assessment, researchers have developed various downscaling techniques, including dynamical downscaling and statistical downscaling. Dynamical downscaling relies on regional climate models for numerical simulation, while statistical downscaling is based on the statistical relationship between global climate model predictors and regional climate variables.

[0003] In recent years, the development of deep learning techniques and high - performance computing has brought new opportunities to downscaling research. Neural network models, especially convolutional neural networks (CNNs) and deep residual networks (DRNs), have shown good potential in downscaling research. These models can capture complex non - linear relationships in data and overcome the limitations of traditional methods to a certain extent. Existing technical methods mainly include: network algorithms that transform low - spatial - resolution climate variables (such as rainfall, temperature, etc.) into high - resolution climate variables; improving climate variables with the same spatial resolution according to climate predictors (atmospheric pressure, geopotential, wind field, humidity, etc.); research from low - spatial - resolution predictors to high - resolution climate variables, and extrapolation prediction research from low - resolution predictors to the ground observation station or regional average. However, existing deep - learning models still face challenges in dealing with daily - scale rainfall downscaling, especially in future prediction scenarios lacking daily - scale, high - resolution environmental variable data.

[0004] Therefore, in the context of frequent climate change and extreme weather, it is necessary to propose and establish a daily rainfall downscaling simulation method based on a deep residual network. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a daily rainfall downscaling simulation method based on a deep residual network, which can effectively improve the spatial resolution and prediction accuracy of daily - scale rainfall under climate change scenarios.

[0006] To achieve the above - mentioned purpose and technical effects, the present invention is implemented by the following technical solutions:

[0007] In the first aspect, the present invention provides a daily rainfall downscaling simulation method based on a deep residual network, including the following steps:

[0008] Obtain the real-time rainfall influencing factors of the area to be predicted; the rainfall influencing factors include meteorological prediction factors, underlying surface data, and regional boundaries;

[0009] Preprocess the rainfall influencing factors to obtain the preprocessed rainfall influencing factors;

[0010] Input the preprocessed rainfall influencing factors into the trained rainfall deep learning network to obtain the grid-type rainfall data of the area to be predicted.

[0011] Further, the meteorological prediction factors include sea level pressure, near-surface air temperature, relative humidity at 500 hPa pressure, geopotential, and relative humidity and geopotential at 800 hPa pressure;

[0012] The underlying surface data includes elevation models and land use and cover.

[0013] Further, preprocessing the rainfall influencing factors to obtain the preprocessed rainfall influencing factors includes:

[0014] Resample the meteorological prediction factors and convert them into data in a projected coordinate system with the same area;

[0015] For the data of meteorological prediction factors, underlying surface data, and regional boundaries, uniformly use the standard score (Z-score) normalization method to scale the data to the same magnitude to obtain the preprocessed rainfall influencing factors.

[0016] Further, the method for the rainfall deep learning network to perform rainfall prediction includes:

[0017] Input the data of meteorological prediction factors into the upsampling sub-module for upsampling, adjust it to be consistent with the resolution of the underlying surface data and regional boundary data, and connect it with the underlying surface data and regional boundary data to generate preliminary sampling data;

[0018] Input the preliminary sampling data into multiple residual sub-modules, and perform skip connection and superposition operations with the outputs of the multiple residual sub-modules to obtain the predicted rainfall data.

[0019] The residual sub-module includes multiple convolutional sub-modules, and the input of the residual sub-module performs skip connection and superposition operations with the outputs of the multiple convolutional sub-modules to obtain the output of the residual sub-module.

[0020] The convolutional sub-module is used to perform convolution operations, non-linearization, and batch normalization.

[0021] Further, the training method of the rainfall deep learning network includes:

[0022] Obtain a training data set; the training data set includes grid rainfall data for each region and corresponding rainfall influencing factors;

[0023] Preprocess the training data set;

[0024] Based on the preprocessed training data set, train in the form of mini-batch data, select Adaptive Moment Estimation (Adam) as the training optimization algorithm, select Mean Squared Error (MSE) as the loss function, and train the rainfall deep learning network to obtain a trained rainfall deep learning network.

[0025] Further, obtaining the training data set includes:

[0026] Use the rainfall fusion data set (CHIRPS) based on ground rain gauges and satellite observations as the data source for grid rainfall data.

[0027] Use the climate system model V2 - medium resolution (BCC-CSM2-MR) climate change data developed by the National Climate Center of the China Meteorological Administration as the data source for meteorological predictors, match the geographical coordinates of the meteorological predictor data with the local projection coordinates, and resample to the projection coordinate system using the nearest distance method.

[0028] Organize a set of meteorological predictors, underlying surface data, regional boundaries, and corresponding rainfall data into a data sample, and each data sample is the smallest unit in the machine learning training process;

[0029] Organize the obtained various data into multiple data samples to obtain the training data set.

[0030] Further, the method for preprocessing the training data set includes global normalization and data augmentation.

[0031] Further, the method of data augmentation includes:

[0032] Sort the original data by total rainfall, select the top 20% of the data for data augmentation, and select 5 commonly used data augmentation methods, namely left - right mirroring, up - down mirroring, rotating 90°, rotating 180°, and rotating 270°. After performing 5 - time data augmentation on the heavy rainfall data, a data set of the same size as the original data is formed.

[0033] Further, during the training process of the deep learning network, the proportion of the training subset and the validation subset is set to 0.8. In network training, the original data and the augmented data in the training subset are used in a ratio of 1:1, and only the original data is used in the validation subset.

[0034] Compared with the prior art, the beneficial effects achieved by the present invention:

[0035] 1. The deep learning network for rainfall downscaling constructed by the present invention takes into account land use changes. Compared with the traditional method that directly performs spatial downscaling on rainfall data in global climate model data using a bias correction function, it insufficiently considers the high bias in the spatial distribution of rainfall and the influence of underlying surface conditions. The present invention performs downscaling on meteorological predictors related to rainfall in global climate model data through resampling, combines it with land use data, and uses it as the input of the residual network to form a downscaling deep residual network framework, so as to output the spatial distribution of medium and small scale rainfall with high resolution, overall improving the adaptability of the downscaled rainfall spatial distribution data and being able to simulate the impacts of climate change and urbanization on regional rainfall changes.

[0036] 2. The present invention designs to introduce a deep residual structure into the deep learning network for rainfall downscaling. Compared with the existing neural network models for rainfall downscaling, the present invention comprehensively uses neural network technologies such as residual network structure, convolutional neural network, transposed convolution, and ReLU-based non-linear activation function to better reflect the continuity in the spatial distribution of rainfall. ReLU (Rectified Linear Unit) non-linearity is used in both the convolutional sub-module and the upsampling sub-module. Transposed convolution is the core operation in the upsampling sub-module of this method.

[0037] 3. In the TensorFlow machine learning platform, the present invention uses the Keras framework to design three directly callable neural network sub-modules: convolutional sub-network, upsampling sub-network, and residual sub-network. These three sub-modules are written as independent functions to form a reusable function library, which not only increases the convenience of adjusting the network depth in the actual application process of the constructed deep residual network framework, but also is used for the development of other deep learning models.

[0038] 4. The present invention sorts the original training data by total rainfall and selects the top 20% of the data to form strong rainfall processes, and uses image reconstruction technologies such as mirroring and rotation to perform data augmentation on the strong rainfall processes, improving the sensitivity of the deep learning network to rainfall data.

[0039] 5. The present invention designs a usage strategy for the original training dataset and the augmented dataset during the training process of the rainfall deep residual network. It uses 80% of the original data and 80% of the augmented data for optimal model training, then uses 20% of the original data for verification. After verification, it uses all the original data and augmented data for enhanced training, forming a "training-verification-training" mode, which not only ensures the effective verification of network parameters but also improves the utilization rate of all training data.

[0040] 6. This method not only has good practical value but also high universality. The present invention aims to propose a daily rainfall downscaling simulation method based on a deep residual network to address the deficiencies of the prior art. This method utilizes the powerful feature extraction ability and complex relationship modeling ability of the deep residual network to improve the accuracy and reliability of daily rainfall downscaling. Compared with traditional statistical downscaling methods, this method has obvious advantages in simulating extreme precipitation events and capturing local climate characteristics. In addition, this method can also adapt to different climate conditions and topographic features, providing more refined precipitation predictions for regional climate assessment and decision-making. Through this method, the impact of climate change on regional weather patterns can be better understood and predicted, thus providing a scientific basis for coping with climate change. Description of the Drawings

[0041] Figure 1 Schematic diagram of the deep residual network framework for rainfall downscaling;

[0042] Figure 2 Schematic diagram of the convolutional sub-module structure;

[0043] Figure 3 Schematic diagram of the upsampling sub-module structure;

[0044] Figure 4 Schematic diagram of the residual sub-module structure;

[0045] Figure 5 Schematic diagram of the framework for using the rainfall downscaling residual network model dataset of the Yangtze River Delta Urban Agglomeration;

[0046] Figure 6 Schematic diagram of the learning curve of the rainfall downscaling residual network model of the Yangtze River Delta Urban Agglomeration (a, b, c, d are spring, summer, autumn, and winter respectively). Detailed Embodiments

[0047] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the protection scope of the present invention.

[0048] Embodiment 1:

[0049] This embodiment provides a daily rainfall downscaling simulation method based on a deep residual network, which specifically includes the following steps:

[0050] Obtain real-time rainfall influencing factors (including meteorological prediction factors, underlying surface data, and boundary data) of the area to be predicted;

[0051] Preprocess the rainfall influencing factors to obtain preprocessed rainfall influencing factors; The preprocessing method mainly includes two types of processing:

[0052] 1. Appropriate resampling is required for the six meteorological prediction factors. Since the global climate data (GCM) data is generally in the form of latitude and longitude grids, the low-latitude grid area is large and the high-latitude grid area is small. Therefore, it is necessary to resample to the projection coordinate system data with the same area, and the grid edge unit is changed from the original ° to km. In the Yangtze River Delta example, 1.125°×1.125° is resampled to 90km×90km.

[0053] 2. Standardization (scale the data to the same order of magnitude to remove the impact of dimension). All 9 types of input data (6 meteorological prediction factors, 2 underlying surface data and boundary data) must be standardized, and the standard score (Z-score) standardization method is uniformly adopted, that is, each data is subtracted from the mean of the data set and then divided by the standard deviation of the data set.

[0054] The preprocessed meteorological prediction factors are input into the trained rainfall deep learning network to obtain grid-type rainfall data for the area to be predicted.

[0055] The trained deep learning network model takes raster tiff data as input (which can be understood as pictures), including:

[0056] 6 meteorological prediction factors (sea level pressure (PSL), near surface air temperature (TAS), relative humidity (RH) at 500hPa (hPa, unit of air pressure) and 800hPa, gravity potential (GPH)),

[0057] 2 underlying surface data (elevation model (DEM), land use and land cover (LULC)),

[0058] 1 area boundary.

[0059] The resolution of meteorological prediction factors is relatively low, 90km×90km in the Yangtze River Delta example; the underlying surface and boundary data are of high resolution, 5km×5km.

[0060] The output is a 5km×5km grid of daily rainfall in the area (which can be understood as a spatial distribution map of daily rainfall).

[0061] When training a rainfall deep learning network for prediction, the main steps are:

[0062] Step 1: Prepare rainfall influencing factors;

[0063] Step 2: Preprocess the data (standardize and scale the data to the same magnitude to remove the impact of different magnitudes of variables);

[0064] Step 3, Python reads data;

[0065] Step 4: Assign the data to the parameters of the network model (trained);

[0066] Step 5: The network model executes the prediction command;

[0067] Step 6: Save the output data (normalized daily rainfall data);

[0068] Step 7: Post-process the output (inverse normalization, restoring rainfall to the original magnitude scale);

[0069] Step 8: Convert the output data to a common format (tiff)

[0070] As Figure 1 shown, the rainfall deep learning network is constructed based on the Keras architecture and includes:

[0071] A convolutional sub-module, an upsampling sub-module, and a residual sub-module,

[0072] As Figure 1 shown, the processing process of the deep learning grid model is mainly as follows:

[0073] Step 1: Six meteorological prediction factors (input_1, i.e., Figure 1 input A in

[0074] are passed through the upsampling module (transpose convolution operation, non-linearization; continuously executed 3 times in this example) to convert the resolution from 90km × 90km to 5km × 5km;

[0075] Step 2: Two types of underlying surface data and boundary data (input_2) are preliminarily processed through the convolutional sub-module (convolution operation, non-linearization, batch normalization), with the resolution remaining unchanged, 5km × 5km;

[0076] Step 3: Combine the results of Step 1 and Step 2 and process them again using the convolutional sub-module;

[0077] Step 4: Input the result of Step 3 into the nested residual sub-module (three series-connected residual networks nested in an overall residual network). The residual sub-module adds a skip connection structure outside the convolutional sub-module, thus transforming the direct prediction of data into the prediction of the residual between it and the true value. This structure increases the depth of the network (reflecting the depth of deep learning) on the one hand and improves the optimization efficiency on the other hand (mathematically proven that it is easier for the target (here the residual) to tend to 0 than for two numbers to continuously approach each other).

[0078] Step 5: Perform a final convolution operation on the result of Step 4 as the output of the predicted variable rainfall, with a resolution of 5km × 5km.

[0079] During the training process, each set of data, namely the nine predictors and the predicted variable rainfall, constitutes a data sample (in one-to-one correspondence); each data sample is the smallest unit in the machine learning training process and cannot be shuffled during the training process.

[0080] The subsequent data augmentation and splitting of the training subset / validation subset are all for the sample set. Data augmentation is the increase in strong rainfall events (corresponding samples); the splitting of the training subset / validation subset is a random splitting of the entire sample set. During this process, each sample itself is not affected in any way.

[0081] Example 2:

[0082] Based on the same technical concept as Example 1, this example provides a daily rainfall downscaling simulation method based on a deep residual network, which can effectively improve the spatial resolution and prediction accuracy of daily-scale rainfall under climate change scenarios.

[0083] 1. Implementation Environment

[0084] This method is implemented by deploying on a computer workstation.

[0085] Hardware Requirements: The central processing unit (CPU) is equipped with at least an Intel Xeon series processor (main frequency ≥ 2.30 GHz) or a computing platform of the same level. For the graphics processing unit (GPU) during the training phase, it is recommended to select NVIDIA's Tesla P100 / K80 / T4 or Google's TensorFlow dedicated processor (TPU); during the inference phase, the performance metrics of the graphics processing unit can be appropriately reduced. The system memory configuration is recommended to be ≥ 36G DDR4 memory. The hard disk should use a solid-state drive (SSD) of more than 1TB.

[0086] System Software Requirements: (1) For the training environment, it is recommended that the operating system be Ubuntu 20.04 LTS, the deep learning framework be TensorFlow 2.10+, and the Python environment be Python 3.8 or above + Conda. (2) For the inference environment, the system can be Windows 10 or above or ubuntu 20.04 or above.

[0087] 2. Applicable Scenarios

[0088] This method is mainly applicable to the analysis and research of the spatio-temporal characteristics of regional rainfall under global climate change. Based on the spatio-temporal distribution data of regional high-resolution rainfall simulated by this method under climate change scenarios, it can be applied to regional water resource prediction and management, urban drainage and flood prevention planning and design, and ecological environment change assessment.

[0089] The method idea proposed by the present invention can be used for downscaling research of other meteorological predictors, such as temperature.

[0090] 3. Overall Concept of the Present Invention

[0091] Using the deep residual neural network algorithm, a non - linear mapping relationship is established between 9 kinds of input data in rainfall influencing factors [6 kinds of low - resolution meteorological data (sea - level pressure (PSL), near - surface air temperature (TAS), relative humidity (RH) and geopotential (GPH) at 500 hPa (hectopascal, air pressure unit) and 800 hPa air pressure), 2 kinds of high - resolution underlying surface data (digital elevation model (DEM), land use and land cover (LULC)), and regional boundaries] and regional high - resolution rainfall data. By using the method of the present invention, it is possible to convert the large - scale climate information (low - resolution meteorological data) provided by the global climate model (GCM) into high - resolution regional rainfall information.

[0092] I. Determining the Training Data of the Rainfall Downscaling Model

[0093] 1. Select raster - type rainfall data as the prediction variable, and use the rainfall fusion dataset (CHIRPS) jointly developed by the Climate Hazards Center of the University of California, Santa Barbara, USA and the United States Agency for International Development as the training data source. The rainfall estimation algorithm of this dataset is advanced, integrating the high accuracy of ground rain gauge observation data and the wide coverage of satellite remote sensing images; the dataset has comprehensive spatial coverage, covering all land areas from 50°N to 50°S; it has a long time range, providing more than 40 years of data from 1981 to the present; and it has a high spatial resolution, with a single raster being 5 km×5 km.

[0094] 2. Select rainfall influencing factors (predictors).

[0095] (1) Select six physical variables closely related to surface rainfall, namely sea - level pressure (PSL), near - surface air temperature (TAS), relative humidity (RH) and geopotential (GPH) at 500 hPa (hectopascal, air pressure unit) and 800 hPa air pressure, as low - precision meteorological predictors, and use the climate system model V2 - medium resolution (BCC - CSM2 - MR) climate change data developed by the National Climate Center of the China Meteorological Administration. This dataset covers historical meteorological data from 1851 to 2014 and predicted meteorological data from 2015 to 2100; the spatial resolution of its ground data is 1.125°×1.125°. Here, the geographical coordinates should be matched with the local projection coordinates, and resampled to the projection coordinate system using the nearest - neighbor method. Without loss of accuracy, it is necessary to ensure that the remaining prediction variables are integer multiples. For example, in the Yangtze River Delta region, it can be resampled to 90 km×90 km.

[0096] (2) To explore the interaction between urbanization and regional climate change, in addition to atmospheric physical factors, surface factors are added, including the Digital Elevation Model (DEM) and Land Use and Land Cover (LULC).

[0097] The surface digital elevation data is an inert indicator that hardly changes over time. The ALOS World 3D – 30m (AW3D30) global digital surface elevation model dataset released by the Earth Observation Research Center of the Japan Aerospace Exploration Agency (JAXA) is selected.

[0098] The surface land use data uses the European Space Agency (ESA) Global Land Cover dataset (ESA-CCI-LC) as the basic data. This dataset is a global land use data product released by the Catholic University of Louvain (UCL) in France relying on the European Space Agency's Climate Change Initiative (CCI) project. This dataset follows the land use classification framework of the Food and Agriculture Organization of the United Nations (FAO), with a horizontal resolution of 300 m and an annual time resolution. It is a long-term global high-resolution land cover classification dataset from 1992 to the present. According to the "Classification of Current Land Use GB / T21010-2017" and the "Standard for Classification of Urban Land Use and Planning Construction Land GB50137-2011", the land cover of this dataset is simplified to land use, and the reclassification relationship is shown in Table 1 and Table 2 to produce local land use classification data.

[0099] Table 1 Conversion Table of European Space Agency Land Cover Data (ESA-CCI-LC) to Land Use

[0100]

[0101] Table 2 Conversion Table of European Space Agency Land Cover Data (ESA-CCI-LC) to Land Use - Continued

[0102]

[0103] (3) Considering that the input of the convolutional neural network model is structured data, that is, it presents a rectangular shape when projected onto a two-dimensional Cartesian coordinate system, and the boundary of a specific administrative region is an irregular shape. Therefore, a research area marker variable is added, with a value of 1 within the area and 0 outside the area.

[0104] II. Design of the Deep Learning Model Architecture

[0105] 1. Overall network structure design. The TensorFlow machine learning system is adopted, and the Keras architecture is selected. In the Python programming language environment, a rainfall deep learning network is built. The modular modeling idea is used to build the network structure, and sub-networks / sub-functions such as convolutional blocks, resampling blocks, and residual blocks are constructed respectively. By calling the sub-networks, a depth-adjustable overall network structure is built, as Figure 1 shown.

[0106] In the above framework diagram, input A (i.e., meteorological prediction factors) and input B (i.e., underlying surface data and regional boundary data) represent prediction factors with different resolutions respectively. They are resampled so that input A, input B, and the output have the same spatial resolution. The resampled feature data are concatenated and used as the input of the residual network. After the depth simulation operation of this network, the output is obtained.

[0107] 2. Construction of sub-modules.

[0108] (1) Convolutional sub-module. In a deep network, generally, convolution operation, non-linearity, and batch normalization (BN) are jointly used. Therefore, in the present invention, a convolutional layer, an activation layer, and a BN layer are connected in sequence to construct a convolutional sub-network, which can be directly used for the call of other sub-networks or the overall network. Based on the convolutional sub-network structure as Figure 2 shown, a convolutional sub-module is constructed. This module adds the "sub-network iteration times" based on the default parameters of APIs such as Conv2D, ReLU, and BatchNormalization in tf.keras. By modifying this parameter, the depth control of the convolutional sub-network can be achieved.

[0109] (2) Upsampling sub-module. A transposed convolutional layer and an activation layer are connected in sequence to construct an upsampling sub-network, and the structure is shown in Figure 3 . Based on the upsampling sub-network structure as Figure 3 shown, an upsampling sub-module is constructed. This module adds the "upsampling factor list" based on the default parameters of APIs such as Conv2DTranspose and ReLU in tf.keras. The upsampling factor list contains information such as the upsampling factor value, the number of upsampling executions, and the order. In specific operations, the upsampling factor value is assigned to the moving step (the moving step must be greater than 1) as the scaling ratio for a single upsampling. By increasing the values in this list, multiple upsamplings can be achieved.

[0110] (3) Residual sub-module. By calling the convolutional sub-network to construct a residual mapping network, and then adding a skip connection structure to form a residual sub-network structure, as Figure 4 shown. Based on the structure as Figure 4Based on the residual sub-network structure shown, an upsampling residual sub-module is constructed. For the residual mapping therein, it is implemented through a convolutional sub-network, that is, a convolutional sub-module with an iteration number of n is called. Similarly, a parameter of "iteration number" is also designed in the function module of this sub-network to control the repetition times of the residual sub-network.

[0111] III. Formulating Deep Learning Training Experiments

[0112] 1. The downscaling of rainfall data has the characteristics of typical large-sample data and it is difficult to directly train the entire data set. Therefore, the mini-batch method is adopted for training, which can effectively reduce the demand for computing resources. In deep learning, the size of the mini-batch is a hyperparameter. Generally, pre-experiments can be carried out on common values such as 16, 32, 64, 128, etc., and the selection can be judged according to the running time and effect. When the size of the mini-batch is set to 32, it means that in one training process, 32 groups of data records participate in the training.

[0113] 2. Select the Adaptive Moment Estimation (Adam) as the training optimization algorithm. The Adaptive Moment Estimation algorithm is an improvement of the gradient descent method, which has the advantages of high numerical calculation efficiency, less memory occupation, and being applicable to large data sets or high-dimensional parameter spaces. The learning rates of the optimization algorithm are respectively selected at five levels of 0.001, 0.0001, 0.00001, 0.000001, and 0.0000001 for hierarchical optimization.

[0114] 3. For the regression simulation problem of rainfall downscaling, according to the maximum likelihood estimation theory, select the Mean Squared Error (MSE) as the loss function / objective function, and the Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) as the evaluation indicators for network learning.

[0115] 4. Preprocess the training data set. The data from 1992 to 2014 is used to train the deep learning network. Before training, the original data needs to be preprocessed, including global normalization and data augmentation.

[0116] (1) Global normalization is to eliminate the influence of the difference in the order of magnitude of different variables. This invention realizes it through normalization, that is, the operation of subtracting the mean value from the data set and then dividing by the mean square error. The mean value and mean square error here will also be used in the inverse normalization post-processing process of the subsequent simulation results.

[0117] (2)Compared with the rainless time, the rainfall time is significantly less, that is, rainfall events are significantly sparse. In other words, the effective data for training is very sparse. In view of this, the present invention introduces a data augmentation technique to expand the dataset, and only performs data augmentation on strong rainfall events. The specific approach is to sort the original data by total rainfall, and select the top 20% of the data for data augmentation. This top 20% of the rainfall data is the strong rainfall event in this study. Five commonly used data augmentation methods are selected, namely left-right mirroring, up-down mirroring, rotating 90°, rotating 180°, and rotating 270°. After performing data augmentation on the strong rainfall data 5 times, a dataset of the same size as the original data is formed.

[0118] 5. During the training process of the deep learning network, the ratio of the training subset to the validation subset is set to 0.8. In the network training, the original data and the augmented data in the training subset are used in a 1:1 ratio, and only the original data is used in the validation subset. The original dataset and the augmented dataset are used in the network training as shown in Figure 5 the figure. The original dataset is divided into a training sub-dataset A and a validation sub-dataset B, and the augmented dataset is divided into a training sub-dataset C and a validation sub-dataset D. Then, validation, training, and validation are performed respectively, followed by retraining.

[0119] 6. Selection of the training machine. The deep learning training process has high requirements for both memory and computing power. When training the above deep learning network, the graphics processing unit (GPU) of the device can select Tesla P100 / K80 / T4 of NVIDIA, and the TensorFlow dedicated processor (TPU) of Google; the GPU video memory should be selected in combination with the simulation area and network depth, generally not less than 36G. When validating or predicting, the performance indicators of the graphics processing unit can be appropriately reduced.

[0120] This embodiment provides a Python function code that can reuse the convolutional sub-module, residual sub-module, and upsampling sub-module, including:

[0121] 1. Convolutional sub-module function. Using the Python programming language, by using TensorFlow system APIs such as Conv2D, ReLU, and BatchNormalization in Keras, the convolutional operation, non-linearization, and batch normalization are used in combination. Based on this, a parameter for increasing the number of num_iteration iterations is added to control the depth of the convolutional network.

[0122] 2. Residual sub-module function. Using the Python programming language, by calling the convolutional sub-network function with 2 iterations, and additionally supplementing with the Add system API in Keras, a residual network structure is constructed. On this basis, a parameter for the number of num_iteration iterations is added to form a cascaded residual network, and the number of iterations can be effectively controlled.

[0123] 3. Upsampling sub-module function. Using the Python programming language, two types of upsampling networks based on transposed convolution and the nearest neighbor method are constructed by using TensorFlow system APIs such as Conv2DTranspose, Conv2D, UpSampling2D, and ReLU in Keras, so as to achieve resampling of images from low resolution to high resolution.

[0124] Example 2:

[0125] In this embodiment of the present invention, the construction of a deep learning network model for daily rainfall downscaling in the Yangtze River Delta urban agglomeration region is taken as an example for illustration.

[0126] Step 1: Determine the model training dataset. Nine predictor factors and one predicted variable are selected for the Yangtze River Delta urban agglomeration. The predicted quantity uses a 5km×5km grid-type rainfall dataset. The predictor factors include 6 atmospheric physical factors, 2 underlying surface factors, and 1 marker for the research scope. The summary of the data attribute characteristics of the training dataset is shown in Table 3.

[0127] Table 3 Data attribute characteristics of the rainfall downscaling residual network model training in the Yangtze River Delta urban agglomeration

[0128]

[0129] Step 2: Determine the network parameters of the deep learning model for daily rainfall downscaling in the Yangtze River Delta urban agglomeration region. According to the spatial resolution of the predictors, they are generalized into two types of input tensors. The meteorological physical variables of 500 hPa RH, 850 hPa RH, 500 hPa GPH, 850 hPa GPH, PSL, and TAS at 90 km×90 km are denoted as input_1, and LULC, DEM, and Domain_Mark are denoted as input_2. Perform three upsampling operations on input_1, with the upsampling multiples being 3×3, 2×2, and 3×3 in sequence, so that the feature information data structures of the two types of input data are the same. Then, splice the feature information of the two types of data, and construct the mapping relationship between the input and output through a residual network, which includes a global residual structure and 3 nested residual networks. Finally, convert the feature information into an output through a convolution operation. This network structure consists of 7 types of tf.keras neural network structure layers: an input layer, a convolutional layer, a ReLU activation layer, a BN layer, a transposed convolutional layer, a splicing layer, and an addition layer. The deep residual network structure for daily rainfall downscaling in the Yangtze River Delta urban agglomeration has a total of 41 layers, including 2 input layers, 10 convolutional layers, 3 transposed convolutional layers, 12 ReLU activation layers, 9 BN layers, 1 splicing layer, and 4 addition layers.

[0130] Step 3: Determine the hyperparameters of the deep learning network. For the residual network model of rainfall downscaling in the Yangtze River Delta urban agglomeration, a smaller convolutional kernel is used, and its size is uniformly set to 3×3. The moving step size is set according to the situation, and is set to 1 in the convolution operation, that is, the convolution operation is performed one by one. In the transposed convolution operation, 2 or 3 are set respectively according to the requirements of upsampling. In order to ensure that the output of the convolution operation has the same size as the input tensor, the edge padding method is adopted, and the padding value is uniformly set to 0.

[0131] After multiple simulation experiments, the size of the mini-batch data is uniformly set to 64, that is, in one training process, 64 data records participate in the training. In one generation of training process, for all the data from 1992 to 2014, it can be calculated according to the following formula that it can be divided into approximately 105 groups of mini-batch data.

[0132]

[0133] For a specific single training, the two parameters of the number of training epochs and the learning rate are fixed. For different training sessions, they are adjusted according to the training results, and the number of training epochs should be optimal for the current training. In this case, the number of training epochs is set between 20 and 40, and the learning rate starts from 0.001 for experimentation. If the model is properly set, for a model that has completed training, a second training can be implemented. Modify the number of training epochs and the learning rate to adapt to the new training. In this case, 0.001, 0.0001, 0.00001, 0.000001, and 0.0000001 will be successively used to train the network multiple times to improve the accuracy of the model.

[0134] Step 4: Based on the global model training, considering that the Yangtze River Delta region has typical climate characteristics with distinct seasons. In view of this, to improve the accuracy of the model, according to the traditional four-season division method in China, the training experiments are divided into 4 groups, that is, the network is fine-tuned separately for different seasons to generate the optimal model for each season. The errors in the training process of the optimal models for the four seasons are plotted as learning curves, as Figure 6 shown. Figure 6 It is a schematic diagram of the learning curve of the rainfall downscaling residual network model for the Yangtze River Delta urban agglomeration. a, b, c, and d are respectively the learning curves of the rainfall downscaling residual network models for spring, summer, autumn, and winter. It can be seen that between the 100th and 110th epochs, the root mean square error (RMSE) and mean absolute error (MAE) of the training subset and validation subset have stabilized.

[0135] Step 5: Model verification. For the case of the rainfall downscaling residual network model in the Yangtze River Delta urban agglomeration area, two groups of experiments are designed to verify the trained optimal model. The test design is as follows:

[0136] (1) Overall error level. Randomly select 20% of the data from the original data, apply the optimal model to predict the rainfall data, and calculate RMSE and MAE respectively. Repeat the experiment 5 times and calculate the arithmetic mean as an index reflecting the overall error level of the model.

[0137] (2) Heavy rainfall error level. Sort the original data by total rainfall, select the top 20% of the data as heavy rainfall event data, apply the optimal model to predict the rainfall data, and calculate RMSE and MAE respectively as an index reflecting the error level of the model in heavy rainfall events.

[0138] Conduct the verification experiment of the optimal model according to the above test design, and the errors are shown in Table 4.

[0139] Table 4 Verification errors of the optimal model of the rainfall downscaling residual network for the Yangtze River Delta urban agglomeration

[0140]

[0141] As can be seen from Table 4, the overall error is less than that of heavy rainfall events; the error in the dry season (such as winter) is less than that in the season with concentrated rainfall (such as summer). Since the unit of the error index is mm, the numerical size of the error is consistent with the input data, that is, the greater the rainfall, for this large dataset, the corresponding larger the RMSE and MAE will be.

[0142] Compared with the existing reports, the daily rainfall deep learning downscaling model constructed in the present invention takes 8 physical environmental variables as inputs, and the model architecture is better; from the perspective of the accuracy of the Yangtze River Delta urban agglomeration region case, in the high-magnification sampling environment of 18 times, the model architecture constructed in the present invention can still show a relatively low error level.

[0143] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0145] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable apparatus to generate a computer-implemented process, thereby providing instructions for implementing the steps of the process Figure 1 in one process or a plurality of processes and / or boxes Figure 1 or steps for implementing the functions specified in one box or a plurality of boxes.

[0147] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A daily rainfall downscaling simulation method based on a deep residual network, characterized in that, It includes the following steps: Obtain the real-time rainfall influencing factors of the area to be predicted; the rainfall influencing factors include meteorological prediction factors, underlying surface data, and regional boundaries; Preprocess the rainfall influencing factors to obtain the preprocessed rainfall influencing factors; Input the preprocessed rainfall influencing factors into the trained rainfall deep learning network to obtain the grid-type rainfall data of the area to be predicted.

2. The daily rainfall downscaling simulation method based on a deep residual network according to claim 1, characterized in that The meteorological prediction factors include sea level pressure, near-surface air temperature, relative humidity at 500 hPa pressure, geopotential, and relative humidity and geopotential at 800 hPa pressure; The underlying surface data includes elevation models and land use and cover data.

3. The daily rainfall downscaling simulation method based on a deep residual network according to claim 1, characterized in that Preprocessing the rainfall influencing factors to obtain the preprocessed rainfall influencing factors includes: Resample the meteorological prediction factors and convert them into data in a projected coordinate system with the same area; For the data of meteorological prediction factors, underlying surface data, and regional boundaries, uniformly adopt the Z-score normalization method to scale the data to the same magnitude to obtain the preprocessed rainfall influencing factors.

4. The daily rainfall downscaling simulation method based on a deep residual network according to claim 1, characterized in that The method for the rainfall deep learning network to perform rainfall prediction includes: Input the data of meteorological prediction factors into the upsampling sub-module for upsampling, adjust it to be consistent with the resolution of the underlying surface data and regional boundary data, and connect it with the underlying surface data and regional boundary data to generate preliminary sampling data; Input the preliminary sampling data into multiple residual sub-modules, and perform skip connection and superposition operations with the outputs of multiple residual sub-modules to obtain the predicted rainfall data; The residual sub-module includes multiple convolutional sub-modules, and the input of the residual sub-module performs skip connection and superposition operations with the outputs of multiple convolutional sub-modules to obtain the output of the residual sub-module; The convolutional sub-module is used to perform convolution operations, non-linearization, and batch normalization.

5. The daily rainfall downscaling simulation method based on a deep residual network according to claim 1, wherein The training method of the rainfall deep learning network includes: Obtain a training data set; the training data set includes the grid-type rainfall data of each region and the corresponding rainfall influencing factors; Preprocess the training data set; Based on the preprocessed training data set, train in the form of mini-batch data, select the Adaptive Moment Estimation Adam as the training optimization algorithm, select the Mean Squared Error MSE as the loss function, and train the rainfall deep learning network to obtain the trained rainfall deep learning network.

6. The daily rainfall downscaling simulation method based on a deep residual network according to claim 5, wherein Obtaining a training data set includes: Use the rainfall fusion data set based on ground rain gauges and satellite observations as the data source of the grid-type rainfall data; Use the climate system model V2 - medium resolution BCC-CSM2-MR climate change data as the data source of the meteorological prediction factors, match the geographical coordinates of the meteorological prediction factor data with the local projection coordinates, and resample it to the projected coordinate system using the nearest distance method; Organize a set of meteorological prediction factors, underlying surface data, regional boundaries, and the corresponding set of rainfall data into a data sample; Organize the obtained various data into multiple data samples to obtain a training data set.

7. The daily rainfall downscaling simulation method based on the deep residual network according to claim 5, wherein The method for preprocessing the training data set includes global normalization and data augmentation.

8. The daily rainfall downscaling simulation method based on a deep residual network according to claim 7, characterized in that, The method of data augmentation includes: Sort the original data by total rainfall, and select the top 20% of the data with the total rainfall for data augmentation. Five commonly used data augmentation methods are selected, namely left-right mirroring, up-down mirroring, rotation by 90°, rotation by 180°, and rotation by 270°. After implementing data augmentation 5 times, a dataset of the same size as the original data is formed.

9. The daily rainfall downscaling simulation method based on the deep residual network according to claim 7, wherein During the training process of the deep learning network, the ratio of the training subset to the validation subset is set to 0.8; in the network training, the original data and the augmented data in the training subset are used in a 1:1 ratio, and only the original data is used in the validation subset.

Images