CN116449460B - Regional month precipitation prediction method and system based on convolution UNet and transfer learning - Google Patents

Abstract

The invention discloses a regional monthly precipitation prediction method and system based on convolution UNet and transfer learning. The method selects forecast factors by performing correlation analysis on regional monthly precipitation and precipitation influencing factors, and establishes a regional monthly precipitation based on convolution UNet network. The monthly precipitation prediction model uses daily scale meteorological data and uses a 30-day sliding window to generate simulated monthly scale meteorological data, pre-trains the regional monthly precipitation prediction model, and then trains the regional monthly precipitation prediction model with the actual observed monthly scale data. , perform parameter fine-tuning; the trained model is used to predict regional monthly precipitation in the future; this invention fully considers factors such as meteorological conditions, topographic features, and underlying surface conditions that affect precipitation, and at the same time, based on daily scale meteorological data A sliding time window is used to obtain simulated monthly-scale meteorological data to expand pre-training samples, enhance the application value of the prediction model in small-sample learning scenarios, and improve the accuracy of regional monthly precipitation.

Description

Regional monthly precipitation prediction method and system based on convolution UNet and transfer learning

Technical field

The invention relates to a regional monthly precipitation prediction method and system, especially a regional monthly precipitation prediction method and system based on convolution UNet and transfer learning.

Background technique

Precipitation is an important factor in the global material and energy cycle and has an important impact on people's daily life and economic and social development. Accurate and timely knowledge of precipitation time, location, and intensity is critical to understanding the operation of the Earth system and improving forecasts of weather, climate, freshwater resources, and natural disaster events. Monthly precipitation forecasting methods can be divided into the following two types: (1) Dynamic model, which applies a set of mathematical and physical equations to simulate different climate conditions for physical processes at different spatial and temporal scales to reveal the interrelationships between the ocean, atmosphere and land. The main advantage of the dynamic prediction method is that the physical mechanism is clear. The disadvantage is that the spatial resolution of the model is coarse, the detailed regional climate information is lacking, the prediction of regional climate elements is relatively limited, and the precipitation forecasting skills of the dynamic model are not high and cannot reach the level of operational application. ; (2) Statistical method, directly applying the statistical relationship between various prediction variables (such as sea surface temperature) and hydrological and climate variables (precipitation, etc.) for modeling and forecasting. Statistical methods are data-driven models with smaller computational burdens and simpler modeling. However, existing statistical prediction methods are mostly limited to single-point and single-station predictions, do not consider the spatial information of precipitation, and do not involve regional-scale precipitation predictions. At the same time, the current observation length of hydrometeorological data is less than a hundred years, which means that less than tens of thousands of data samples can be used for model learning and training. For general deep learning models, the number of studyable samples is too small. Too few samples can easily lead to overfitting of the model, making it difficult to achieve the expected prediction effect.

Contents of the invention

Purpose of the invention: The purpose of the invention is to provide a regional monthly precipitation prediction method and system based on convolution UNet and transfer learning that achieves high-precision prediction, so as to solve the deficiencies in the existing technology.

Technical solution: The regional monthly precipitation prediction method based on convolution UNet and transfer learning according to the present invention includes the following steps:

Conduct correlation analysis on regional monthly precipitation and precipitation influencing factors, and screen forecast factors;

Use the forecast factors as the input of the convolutional UNet network and the regional monthly precipitation as the output in the future to build a regional monthly precipitation prediction model;

Use daily-scale meteorological data to generate simulated monthly-scale meteorological data using a 30-day sliding window, pre-train the regional monthly precipitation prediction model, and then use the observed actual monthly-scale meteorological data to train the regional monthly precipitation prediction model to fine-tune parameters;

Based on the measured regional monthly meteorological data at the current time, the regional monthly precipitation prediction model after pre-training and parameter fine-tuning is used to predict regional monthly precipitation in the future.

Further, the factors affecting precipitation include meteorological elements, topographic and geomorphological elements and underlying surface elements.

Further, the correlation analysis of regional monthly precipitation and precipitation influencing factors is performed, and the screening of forecast factors includes using classification and regression trees to score the importance of the precipitation influencing factors, and using a correlation evaluation method to calculate the precipitation influencing factors. Correlation coefficient with regional monthly precipitation, and select the predictor factors in descending order according to the importance score and correlation coefficient.

Further, the daily-scale meteorological data uses a 30-day sliding window to generate simulated monthly-scale meteorological data. Pre-training the regional monthly precipitation prediction model includes: each set of daily-scale meteorological data uses a 30-day sliding window to generate 30 The simulated monthly scale meteorological data composed of daily scale meteorological data is used to pre-train the regional monthly precipitation prediction model using the Adam algorithm.

Further, training the regional monthly precipitation prediction model with the observed actual monthly precipitation data to obtain the regional monthly precipitation prediction model includes using the model parameters obtained by the pre-training as initial parameters for training through transfer learning, Perform parameter fine-tuning and super-parameter selection on the regional monthly precipitation prediction model.

Further, the correlation analysis of regional monthly precipitation and precipitation influencing factors, after screening the forecast factors, also includes unifying the spatial resolution of the forecast factor data, integrating the forecast factor data to the same spatial scale based on the spatial interpolation method, and improving the forecast. Factor data are standardized.

The regional monthly precipitation prediction system based on convolution UNet and transfer learning according to the present invention includes:

The forecast factor screening module is used to conduct correlation analysis on regional monthly precipitation and precipitation influencing factors, and screen forecast factors;

A regional monthly precipitation prediction model building module is used to use the forecast factors as inputs to the convolutional UNet network and the regional monthly precipitation as the output in the future to build a regional monthly precipitation prediction model;

The model training module is used to generate simulated monthly-scale meteorological data using daily-scale meteorological data using a 30-day sliding window, pre-train the monthly precipitation prediction model in the region, and then use the observed actual monthly-scale meteorological data to predict the monthly precipitation in the region. Predictive model parameters are fine-tuned;

The regional monthly precipitation prediction module is used to predict regional monthly precipitation in the future based on the measured monthly scale data of the forecast factors at the current time, using the regional monthly precipitation prediction model after pre-training and fine-tuning parameters.

The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements the convolution-based UNet and transfer learning. Regional monthly precipitation prediction method.

The computer-readable storage medium of the present invention stores a computer program. When the computer program is executed by a processor, the regional monthly precipitation prediction method based on convolution UNet and transfer learning is implemented.

Beneficial effects: Compared with the existing technology, the advantages of the present invention are: (1) The convolutional UNet network is used to realize regional prediction of precipitation; unlike the previous single-point and single-station prediction methods, the present invention uses a convolutional neural network to consider precipitation. , temperature and other meteorological elements, depict the spatial structure of precipitation, reflect the spatial differences of precipitation elements, enable the spatial information characteristics of meteorological elements to be more fully utilized in the precipitation prediction process, and output high-precision regional monthly precipitation Prediction results; (2) The present invention uses transfer learning, uses daily scale meteorological data to generate a large amount of monthly scale meteorological data through the sliding window method, pre-trains the prediction model, realizes deep learning prediction of small sample scenarios, and improves its prediction performance ; (3) The present invention simultaneously considers the impact of meteorological conditions, topographic and geomorphological characteristics, underlying surface conditions, etc. on precipitation prediction in the construction of the prediction model, based on the importance score, mutual information, autocorrelation/partial autocorrelation of classification and regression trees and other methods, comprehensively select the main forecast factors that affect regional monthly precipitation forecast, and improve the accuracy of monthly precipitation forecast.

Description of the drawings

Figure 1 is a flow chart of the regional monthly precipitation prediction method of the present invention.

Figure 2 is a schematic diagram of the sliding time window method of the present invention.

Figure 3 is an architecture diagram of the regional monthly precipitation prediction model in the embodiment of the present invention.

Detailed ways

The technical solution of the present invention will be further described below with reference to the accompanying drawings.

Deep learning:

Definition 1: Deep learning usually refers to deep neural network, which is a type of model with multi-layer network structure developed on the basis of neural network, including long short-term memory network (Long short-term memory, LSTM), convolutional neural network (Convolutional neural network, CNN) and other common structures.

UNet network:

Definition 2: UNet is improved based on FCM (fully convolutional network) and has a symmetrical U-shaped structure, including decoder, bottleneck module, decoder and other parts.

Transfer learning:

Definition 3: Transfer learning is a machine learning method that uses existing knowledge to solve problems in different but related fields.

Classification and regression trees:

Definition 4: Classification and regression tree (CART) is a type of decision tree algorithm. CART selects the split point and split variable for each node to minimize the overall node impurity. In regression tasks, impurity is represented by mean square error, which is used as the basis for dividing nodes. In CART, the feature (variable) importance score is provided by calculating the reduction of each feature to the splitting criterion (usually the Gini coefficient in classification tasks and the mean square error in regression tasks).

As shown in Figure 1, the regional monthly precipitation prediction method of convolution UNet and transfer learning includes the following steps:

Step 1: Selection of predictors.

1) Determine the set of potential influencing factors, including the comprehensive effects of precipitation on atmospheric circulation, topography and underlying surface conditions. Meteorological factors such as early precipitation, temperature, wind speed, humidity, evapotranspiration, topographic factors such as terrain elevation, and underlying surface factors such as vegetation coverage and soil moisture are used as potential influencing factors for regional monthly precipitation.

2) Importance and correlation evaluation. First, apply classification and regression trees to evaluate the importance of each feature. Taking the set of potential influencing factors as input and the regional monthly precipitation to be predicted as output, a regional monthly precipitation simulation prediction model based on CART is constructed, the reduction amount of each potential influencing factor that can be used as the classification standard is calculated, and the importance of each feature is obtained. Score. Secondly, correlation evaluation methods such as mutual information, autocorrelation, and partial autocorrelation are applied to calculate the correlation coefficient between each potential influencing factor and regional monthly precipitation.

Taking Gini importance, a measure of variable importance commonly used in CART, as an example, the variable importance evaluation calculation process is as follows:

Assume that the proportion of the kth category samples at node τ in the current sample set is as follows:

_pk ＝ _nk /n(1)

In the formula, p _k represents the proportion of the k-th sample, n _k is the number of k-th category samples, and n is the total number of samples.

Then, Gini impurity can be expressed in the following form,

i(τ)＝1-∑ _k p _k ² (2)

Among them, i(τ) is the Gini impurity.

When the sample is divided into sub-nodes τ ₁ and τ ₂ respectively, the Gini impurity will change. Use formula (3) to define the reduction of impurity i(τ),

Δi(τ)＝i(τ)-p _l i(τ ₁ )-p _r i(τ ₂ ) (3)

Among them, p _l and p _r respectively represent the sample subsets at the left and right of the τ node.

Then, the variable importance can be obtained by calculating the reduction in impurity i(τ) of all nodes (for feature θ),

I _G (θ)＝∑ _τ Δi _θ (τ) (4)

Among them, I _G (θ) is the importance of the variable, and Δi _θ (τ) is the change in Gini impurity.

The mutual information I(X;Y) between variables is calculated as follows:

Among them, X and Y are the joint probability density of X and Y of influencing factors and monthly precipitation respectively, p(x,y), p(x) and p(y) are the marginal probability densities of X and Y respectively.

The variable autocorrelation coefficient ACF is calculated as follows:

_In the formula _,

3) Selection of model forecast factors, sorting them from large to small according to the importance score and correlation coefficient of each potential influencing factor, and selecting the model forecast factors. In practical applications, the analysis results of known rainfall causes can be combined to comprehensively screen model forecast factors. For example, when the importance scores and correlation coefficient ranking results of potential influencing factors are inconsistent, the model can be selected based on the comprehensive comparison and selection of known rainfall cause analysis results. Predictor factors.

Step 2, data preprocessing.

Based on the selected model predictors, various data are collected and preprocessed.

First, unify the spatial resolution of model predictor data and integrate the data to the same spatial scale based on spatial interpolation methods; second, standardize all model predictor data.

In this embodiment, the Z-Score normalization method is used, as shown in the following formula:

Among them, x _scaled and x _obs represent the standardized data and original observation data respectively, and μ _x and σ _x represent the mean and standard deviation of the observation data respectively.

Step 3, design the model architecture.

Using convolution UNet as the basic structure, the selected model forecast factors as input, and the regional monthly precipitation in the future as the output, a regional monthly precipitation prediction model is constructed. Convolutional UNet adopts an encoder-decoder construction, where the encoder includes multiple sequences of convolution and downsampling operations, and after each downsampling, the number of filters is doubled in the convolutional layer. The decoder is similar, including the same sequence of convolution and upsampling operations, with the feature map reduced by half after each upsampling. The last layer uses a convolution operation to output the result. Skip connections are used to connect the features of the decoder to the features of the encoder. In the convolution operation, ReLU is used for activation; in the last convolution layer, the linear function is used for activation.

The UNet network adopts a fully convolutional structure and is very suitable for regional model construction. Geospatial data naturally has time and spatial information, that is, it can be represented by multi-dimensional arrays with local correlations. Convolutional neural networks are designed to process data represented in the form of multi-dimensional arrays such as images and videos. CNN can extract information features hidden in the data through convolution and pooling operations. Therefore, it has good applicability to use convolution UNet to construct a regional precipitation forecast model with spatiotemporal distribution characteristics. Secondly, UNet adopts the classic encoding-decoding structure. First, the network performs downsampling through the encoder; secondly, it performs upsampling through the decoder; in the middle, the connection between low-level features and high-level features is achieved through skip connections similar to deep residual networks. This structure allows the UNet network to obtain better model effects at a smaller cost (including parameters and sample number).

Step 4, model pre-training.

1) Generate simulated monthly scale meteorological data, using daily precipitation, temperature and other data, using a 30-day sliding window method to generate a pre-training sample of simulated monthly precipitation and monthly temperature consisting of 30 daily precipitation and temperature. The sliding time window is calculated as shown in the following formula, and the schematic diagram for generating simulated monthly scale meteorological data is shown in Figure 2.

Among them, t is the time, t=1, 2, 3..., to the sample length; n is the length of the sliding window, which is taken as 30 in the present invention.

2) Pre-train the model, apply the generated pre-training sample set, use the Adam algorithm, set the learning rate to 1e-3, pre-train the model, and save the model pre-training parameters.

Step 5, model training.

1) Divide the actual sample set: Use hydrometeorological observation monthly scale data (precipitation, temperature, etc.) to form an actual sample set, and divide the actual sample into three subsets: training, verification, and testing. Among them, the training set is used for model parameter fitting, the validation set is used to evaluate the prediction performance of the trained model to optimize hyperparameters, and the test set is used to evaluate the model fitting performance after optimization.

2) Parameter fine-tuning: Use a sample set composed of actual measured data, apply transfer learning, use the pre-training parameters saved in the previous steps as the initial parameters of this round of model training, set the learning rate to 1e-5, and use the root mean square error, average Relative error, Nash efficiency coefficient and other evaluation indicators are used, and Adam algorithm, regularization, early stopping and other strategies are used to fine-tune the parameters of the model and select super parameters.

Step 6: Based on the constructed model, use the measured precipitation, temperature and other data at the current moment to predict the regional monthly precipitation at the next moment.

The method of the present invention is verified below through specific experiments.

Taking China's monthly precipitation data as an example, the regional monthly precipitation prediction method based on convolution UNet and transfer learning is used to predict my country's monthly precipitation.

(1) Actual measurement data

Monthly precipitation and monthly temperature observation data come from the China surface precipitation and temperature monthly value 0.5°×0.5° grid data set (V2.0) released by the National Meteorological Information Center. Daily precipitation and daily temperature observation data come from China's surface precipitation. , daily temperature value 0.5°×0.5° grid data set (V2.0). The precipitation and temperature observation length is from January 1961 to December 2019. The 1km digital elevation model data comes from the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences. The above data were re-scaled into a 1° × 1° grid data set through bilinear interpolation.

(2) Selection of model predictors

Apply CART's importance score, variable mutual information, and autocorrelation/partial autocorrelation analysis to calculate the importance and correlation between precipitation and potential influencing factors such as temperature, humidity, wind speed, terrain elevation data, and soil moisture. According to the importance score and correlation coefficient, elevation data, previous precipitation, and temperature were selected as model forecast factors. Then the model input is the previous precipitation, temperature elements and elevation data at t-1, t-6, t-11 and t-12, and the model output is the precipitation at time t, where t represents the current month and t-1 represents the previous month. Month, t-12 represents the previous year.

(3)Data preprocessing

The data in step (1) above were re-scaled into a 1° × 1° grid data set through bilinear interpolation, and all data were standardized.

(4)Model architecture design

The architecture of the regional monthly precipitation prediction model constructed in this embodiment is shown in Figure 3. Among them, the encoder includes two sequences of convolution and downsampling operations. A single sequence contains two 3×3 convolution kernels, followed by a 2×2 maximum pooling operation with a step size of 2. Each time After sampling, the number of filters is doubled in the convolutional layer (16->32->64). The decoder is similar, including 2 sequences of convolution and upsampling operations. A single sequence contains a 2×2 upsampling operation, which reduces the feature map by half (64->32->16), and then performs 2 3 ×3 convolution operation. The last layer uses a 1×1 convolution operation to output the result. Skip connections are used to connect the features of the decoder to the features of the encoder. In the convolution operation, ReLU is used for activation; in the last convolution layer, the linear function is used for activation.

(5)Model pre-training

Apply daily precipitation and daily temperature data, use a 30-day sliding window method to generate pre-training samples, form a pre-training sample set, use the Adam algorithm, set the learning rate to 1e-3, pre-train the model, and save the pre-training model parameters.

(6)Model training

Apply transfer learning, use measured monthly precipitation and monthly temperature data to form an actual sample set, use the pre-training model parameters as the initial values of this round of training parameters, fine-tune the parameters of the pre-training model, set the learning rate to 1e-5, and add the samples The set is divided into training set, verification set, and test set, corresponding to the model training period from 1961 to 2000, the verification period from 2001 to 2010, and the model testing period from 2011 to 2019.

(7) Model application and evaluation

The root mean square error, relative error, Nash coefficient, and correlation coefficient are used to compare and evaluate the model effect. The specific results are shown in Table 1. Table 1 is the regional monthly precipitation prediction effect table.

Table 1

It can be seen from the above experimental results that the present invention uses convolution UNet and transfer learning to predict regional monthly precipitation, which can well describe the spatiotemporal distribution characteristics of regional monthly precipitation, and also shows good prediction performance in small sample learning scenarios, and can Provide accurate and reliable regional-scale monthly precipitation prediction results.

The regional monthly precipitation prediction system based on convolution UNet and transfer learning according to the present invention includes:

The forecast factor screening module is used to conduct correlation analysis on regional monthly precipitation and precipitation influencing factors, and screen forecast factors;

A regional monthly precipitation prediction model building module is used to use the forecast factors as inputs to the convolutional UNet network and the regional monthly precipitation as the output in the future to build a regional monthly precipitation prediction model;

The model training module is used to generate simulated monthly-scale meteorological data using daily-scale meteorological data using a 30-day sliding window, pre-train the monthly precipitation prediction model in the region, and then train the monthly precipitation in the region with the actual monthly-scale meteorological data observed. Predictive model parameters are fine-tuned;

The regional monthly precipitation prediction module is used to predict regional monthly precipitation in the future based on the measured monthly scale data of the forecast factors at the current time, using the regional monthly precipitation prediction model after pre-training and fine-tuning parameters.

The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements the convolution-based UNet and transfer learning. Regional monthly precipitation prediction method.

The computer-readable storage medium of the present invention stores a computer program. When the computer program is executed by a processor, the regional monthly precipitation prediction method based on convolution UNet and transfer learning is implemented.

The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage, flash memory, or the desired program code in a form that may be used to store instructions or data structures and any other media that can be accessed by a computer.

The processor is used to execute the computer program stored in the memory to implement various steps in the methods involved in the above embodiments.

Claims (10)

1. The regional month precipitation prediction method based on convolution UNet and transfer learning is characterized by comprising the following steps of:

carrying out correlation analysis on regional month precipitation and precipitation influence factors, and screening forecasting factors;

taking the forecasting factors as input of a convolution UNet network, taking regional month precipitation at the future moment as output, and constructing a regional month precipitation prediction model;

generating simulated month scale weather data by adopting a 30-day sliding window according to the day scale weather data, pre-training the regional month precipitation prediction model, and training the regional month precipitation prediction model according to the observed actual month scale weather data to carry out parameter fine adjustment;

and predicting regional month precipitation at a future moment by using the regional month precipitation prediction model after pretraining and parameter fine tuning according to the measured month scale data of the current moment predictor.

2. The regional moon precipitation prediction method based on convolution UNet and transfer learning according to claim 1, wherein the precipitation influencing factors comprise meteorological elements, topography elements and underlying elements.

3. The regional month precipitation prediction method based on convolution UNet and transfer learning according to claim 1, wherein the step of performing correlation analysis on regional month precipitation and precipitation influence factors comprises the steps of grading importance of the precipitation influence factors by using classification and regression trees, calculating correlation coefficients of the precipitation influence factors and regional month precipitation by using a correlation evaluation method, and selecting the forecasting factors from large to small according to the importance grading and the correlation coefficients.

4. The regional month precipitation prediction method based on convolution UNet and transfer learning according to claim 1, wherein the generating of the simulated month scale weather data by adopting a 30-day sliding window with daily scale weather data includes the pre-training of the regional month precipitation prediction model by adopting a 30-day sliding window with each group of daily scale weather data to generate the simulated month scale weather data consisting of 30 daily scale weather data, and the pre-training of the regional month precipitation prediction model by adopting an Adam algorithm.

5. The regional month precipitation prediction method based on convolution UNet and transfer learning according to claim 1, wherein training the regional month precipitation prediction model according to observed actual month scale meteorological data to obtain a regional month precipitation prediction model comprises performing parameter tuning and super-parameter selection on the regional month precipitation prediction model by transfer learning by taking model parameters obtained by the pre-training as initial parameters of training.

6. The regional month precipitation prediction method based on convolution UNet and transfer learning according to claim 1, wherein the correlation analysis is performed on regional month precipitation and precipitation influence factors, the spatial resolution of the predictor data is unified after the predictor is screened, the predictor data is integrated to the same spatial scale based on a spatial interpolation method, and standardized processing is performed on the predictor data.

7. Regional month precipitation prediction system based on convolution UNet and transfer learning, characterized by comprising:

the forecasting factor screening module is used for carrying out correlation analysis on regional month precipitation and precipitation influence factors and screening forecasting factors;

the regional month precipitation prediction model building module is used for taking the predictor as the input of the convolution UNet network, taking regional month precipitation at the future moment as the output, and building a regional month precipitation prediction model;

the model training module is used for generating simulated month scale weather data by adopting a 30-day sliding window according to the day scale weather data, pre-training the regional month precipitation prediction model, and then performing parameter fine adjustment on the regional month precipitation prediction model according to the observed actual month scale weather data;

and the regional month precipitation prediction module is used for predicting regional month precipitation at a future moment by using the regional month precipitation prediction model after pretraining and parameter fine tuning according to the month scale data actually measured by the current moment forecasting factors.

8. The regional month precipitation prediction system based on convolutional UNet and transfer learning according to claim 7, wherein the precipitation influencing factors comprise meteorological elements, topographical elements and underlying elements.

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the computer program when loaded to the processor implements the regional moon precipitation prediction method based on convolution UNet and transfer learning according to any of claims 1-6.

10. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements a regional moon precipitation prediction method based on convolution UNet and transfer learning according to any one of claims 1-6.