Medium- and long-term precipitation forecast method in the Yangtze River Basin based on GRU-Attention

Abstract

The present invention belongs to the technical field of precipitation prediction, and specifically provides a medium- and long-term precipitation forecasting method for the Yangtze River basin based on GRU-Attention, including: data preprocessing: preprocessing the historical precipitation daily data of the upper Yangtze River basin within a set time period, and obtaining the time series corresponding to the spatial field that meets the set conditions as the prediction object; clarifying the key prediction factor combination of precipitation in the upper Yangtze River; building a GRU-Attention deep learning model; setting a training set and a test set, training the key prediction factor combination of each prediction object, selecting the optimal hyperparameter combination, and obtaining the optimal model for each prediction object; providing real-time prediction results based on the key prediction factor combination and the optimal prediction model. This method searches for key prediction factors based on the evolution law of precipitation in the upper Yangtze River, constructs an effective artificial intelligence model for medium- and long-term prediction of precipitation in the upper Yangtze River, and improves the current sub-seasonal prediction skills of precipitation in the upper Yangtze River.

Description

Technical Field

[0001] The present invention belongs to the technical field of precipitation prediction, and in particular, relates to a medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention. Background Art

[0002] Sub-seasonal / medium- and long-term forecasts aim to predict weather anomalies in the next 2 weeks to 3 months, and are a key link in the seamless forecasting of weather and climate. Droughts and floods frequently occur in the upper reaches of the Yangtze River. The medium- and long-term forecast of precipitation in the upper reaches of the Yangtze River is of great significance to disaster prevention and mitigation and the sustainable development of the economy and society in the region. Due to the complex terrain of the upper reaches of the Yangtze River and the many factors that affect precipitation, the current sub-seasonal forecast level of precipitation in the region based on the initialization dynamic model is very limited. In the extended period to the sub-seasonal scale, the initial contribution of the atmosphere has been greatly attenuated, and the underlying surface effect of external forcing has just begun to appear. The sources of predictability within the sub-seasonal scale are very scarce. The current sub-seasonal forecast is an internationally difficult issue, and there is an urgent need to seek more potential sources of predictability and new sub-seasonal forecast methods to improve and improve the current sub-seasonal forecast level. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention, to find key prediction factors according to the evolution law of precipitation in the upper reaches of the Yangtze River, to construct an effective artificial intelligence model for medium- and long-term precipitation forecasting in the upper reaches of the Yangtze River, and to improve the current sub-seasonal forecasting skills for precipitation in the upper reaches of the Yangtze River.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention, comprising the following steps:

[0005] Step 1: Data preprocessing: Preprocess the historical precipitation daily data of the upper reaches of the Yangtze River within the set time period, cluster the data according to the set conditions, and obtain the time series corresponding to the spatial field that meets the set conditions as the prediction object;

[0006] Step 2: Identify the key prediction factor combination of precipitation in the upper reaches of the Yangtze River;

[0007] Step 3: Build the GRU-Attention deep learning model;

[0008] Step 4: Set the training set and test set, train the key prediction factor combination of each prediction object, select the optimal hyperparameter combination, and obtain the optimal prediction model for each prediction object as the final artificial intelligence model for medium- and long-term prediction of precipitation in the upper reaches of the Yangtze River;

[0009] Step 5: Provide real-time prediction results based on the combination of key prediction factors and the optimal prediction model.

[0010] In a preferred solution, in step 1, data preprocessing is performed using sliding average and cluster analysis, including the following steps:

[0011] S101. Calculate the daily data as the ten-day average using a sliding average, and calculate the ten-day deviation data;

[0012] S102. K-means clustering technology is used to reveal the spatial distribution characteristics of the temporal variation law of precipitation on a ten-day scale in the upper reaches of the Yangtze River. Areas with consistent temporal evolution laws in space are identified as clustering key areas. For each clustering key area, the average area of ​​each clustering key area is taken as the prediction object.

[0013] In a preferred embodiment, in step 2, one or more of the decadal-scale data of outward longwave radiation, global sea surface temperature, geopotential height of sea level pressure at 50 hPa and 500 hPa, and horizontal wind field at 700 hPa are selected as prediction factors.

[0014] In the preferred scheme, in the step 2, for each spatial field of precipitation in the upper reaches of the Yangtze River, the correlation coefficient between each cluster key area and the prediction factor of more than 10 to 60 days ahead is calculated respectively, the key areas of different prediction factors with different prediction time effectiveness are clarified, and the regional average value of the significant area on the correlation coefficient graph is calculated as the prediction factor time series, that is, the prediction factor index series, and each prediction factor time series is combined into a potential prediction factor time series combination, and the prediction factor time series with the highest correlation with the prediction object in each type of prediction factor is analyzed and screened, and the screened prediction factor time series are combined into the key prediction factor combination of the prediction object of the cluster key area, and the above operation is repeated to obtain the key prediction factor combination of all cluster key areas.

[0015] In the preferred scheme, in step three, a GRU gated recurrent neural network is selected as the prediction model, and an attention mechanism is added to form a GRU-Attention deep learning model.

[0016] In the preferred solution, the GRU-Attention deep learning model is divided into three layers: input layer, hidden layer and output layer;

[0017] At the input layer, the input value at each moment is the key predictor index sequence at the moment in the key predictor combination;

[0018] The hidden layer consists of three parts. The first layer is the GRU layer. The first dimension of the input and output tensors is the batch size. It is responsible for processing the input sequence and capturing the dependencies in the time series data. The second layer is a multi-head attention mechanism layer. It uses the specified number of heads and dropout ratio to operate on the output of the GRU layer to further extract features and enhance the model's ability to learn the relationship between different parts of the sequence. The third layer is the LeakyReLU activation function layer.

[0019] The final output layer is a fully connected layer, which maps the features processed by the LeakyReLU activation function to the final output. It contains only one neuron and outputs a value at each moment, which is the predicted object of each cluster key area.

[0020] In the preferred scheme, in the step four, a training period and a test period are constructed, and the data of the training period and the test period correspond to the training set and the test set respectively. The key prediction factor combination time series of each prediction object is used as the characteristic value, and the prediction object time series is fitted by the model to construct a hyperparameter space. After the hyperparameter selection training, a suitable optimal model is obtained, and the optimal prediction model for each cluster key area with a fixed prediction time is obtained.

[0021] In a preferred embodiment, the hyperparameter space is constructed by setting different values ​​for the delta value of the Huber loss function, the number of hidden units of the GRU layer, the learning rate, the number of times the entire training set is traversed during training, the number of heads in the multi-head attention mechanism, the dropout ratio used in the multi-head attention mechanism, the negative slope parameter of the LeakyReLU activation function, and the learning rate attenuation factor.

[0022] In a preferred solution, the hyperparameters are selected as the optimal hyperparameter combination formed by evaluating the mean square error as an indicator in the test set.

[0023] In the preferred scheme, in the step five, based on the key prediction factor combination of the target ten-day prediction time of 10 to 60 days, the stored key area and lead time information are used to calculate the exponential combination of the real-time prediction year key prediction factors, and input it as a characteristic value into the trained optimal prediction model to calculate the prediction value of each type of prediction object in the target ten-day prediction time of 10 to 60 days.

[0024] The present invention provides a medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention, which has the following beneficial effects:

[0025] 1. The present invention selects a factor index combination with a clear physical mechanism as a characteristic value for the prediction time of precipitation in different decades in the upper reaches of the Yangtze River, and inputs these characteristic values ​​into a deep learning model for training. This can improve the prediction skills while also making up for the lack of explanatory power of the deep learning model.

[0026] 2. This invention uses K-means clustering technology to classify the historical distribution characteristics of precipitation in the upper reaches of the Yangtze River in different periods to form 6 spatial field categories. For these spatial field categories, the present invention performs regional average processing on the prediction factors and extracts the time series data of the corresponding category as the prediction target. Compared with the traditional point-by-point prediction method, the present invention only needs to predict a single value for the same spatial field category, which significantly improves the calculation efficiency and reduces the consumption of computing resources.

[0027] 3. The present invention adopts the GRU-Attention model for prediction. As a variant of the recurrent neural network (RNN), GRU solves the problem of gradient vanishing or gradient exploding that is prone to occur in traditional RNN when processing long sequences, and is particularly suitable for modeling time series data. The attention mechanism can automatically focus on the important parts of the input data and improve the model's ability to capture key information. Compared with the traditional multivariate linear regression method, the GRU-Attention model can capture nonlinear patterns and complex relationships in time series data, and reasonably allocate feature weights through the attention mechanism, strengthen the characterization of the changing trend of mutation data, and enable the model to more accurately capture information at key moments, thereby improving prediction accuracy. At the same time, through the hyperparameter optimization library Optuna, hyperparameters such as learning rate, hidden layer, number of iterations, and activation function are tuned to improve the model's adaptability to the prediction object, and through the prediction effect.

[0028] 4. After the GRU-Attention model training is saved, you only need to enter the key prediction factor index combination corresponding to the prediction target for the year to get the prediction result. There is no need to retrain, which greatly reduces the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0030] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;

[0031] Figure 2 This is a comparison chart of the ACC score prediction skills of the GRU-Attention model and the ECWMF power model from 2013 to 2023 in Example 2. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] Embodiment 1:

[0034] like Figure 1 As shown in FIG. 1 , a medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention includes the following steps:

[0035] Step 1: Data preprocessing:

[0036] The historical daily precipitation data of the upper reaches of the Yangtze River within a set time period are preprocessed, the data are clustered according to the set conditions, and the time series corresponding to the spatial field that meets the set conditions is recorded as the prediction object.

[0037] Specifically, data preprocessing is performed using sliding average and cluster analysis, including the following steps:

[0038] S101. Use sliding average to calculate daily data into ten-day average, select 1981-2010 as the climate state, and calculate the ten-day deviation data.

[0039] S102. K-means clustering technology is used to reveal the spatial distribution characteristics of the temporal variation law of precipitation on a ten-day scale in the upper reaches of the Yangtze River. Areas with consistent temporal evolution laws in space are identified as clustering key areas. For each clustering key area, the average area of ​​each clustering key area is taken as the prediction object to reduce the consumption of computing resources.

[0040] In this embodiment, six coordinated change regions in the upper reaches of the Yangtze River are identified based on the silhouette coefficient index.

[0041] Step 2: Identify the key prediction factors of precipitation in the upper reaches of the Yangtze River.

[0042] One or more of the decadal-scale data of outward longwave radiation, global sea surface temperature, sea level pressure at 50hPa and 500hPa potential heights, and 700hPa horizontal wind field are selected as prediction factors.

[0043] For the six clustering key areas of precipitation in the upper reaches of the Yangtze River, the correlation coefficients of each category with the prediction factors of more than 10 to 60 days ahead were calculated, the key areas of different prediction factors with different prediction time effectiveness were identified, and the regional average values ​​of the significant areas on the correlation coefficient graph were calculated as the potential prediction factor time series combination. The prediction factor time series with the highest correlation with the prediction object in each category of prediction factors was analyzed and selected as the key prediction factor. All key prediction factor time series combinations become the key prediction factor combination of this type of prediction object.

[0044] Repeat the above steps to obtain the key predictor combinations of all cluster key areas.

[0045] Step 3: Build the GRU-Attention deep learning model.

[0046] The GRU gated recurrent neural network is selected as the prediction model, and the attention mechanism is added to form the GRU-Attention deep learning model.

[0047] GRU effectively controls the flow of information through the gating mechanism and can capture long-term dependencies in time series. The attention mechanism enables the model to automatically focus on the important parts of the input data and improve the model's ability to capture key information. This is especially important for time series prediction because the model needs to identify and utilize the features most relevant to the prediction target.

[0048] The GRU-Attention deep learning model is divided into three layers: input layer, hidden layer and output layer.

[0049] At the input layer, the input value at each moment is the key predictor index at the moment in the key predictor combination, that is, the regional average of the significant area on the correlation coefficient graph.

[0050] The hidden layer consists of three parts. The first layer is the GRU layer, which is set to batch_first=True. The first dimension of the input and output tensors is the batch size. It is responsible for processing the input sequence and capturing the dependencies in the time series data. The second layer is a multi-head attention mechanism layer, which uses the specified number of heads and dropout ratio to operate on the output of the GRU layer to further extract features and enhance the model's ability to learn the relationship between different parts of the sequence. The third layer is the LeakyReLU activation function layer. LeakyReLU is used to increase the nonlinearity of the model and help the model learn more complex feature representations.

[0051] The final output layer is a fully connected layer, which maps the features processed by the LeakyReLU activation function to the final output. It contains only one neuron and outputs a value at each moment, namely the predicted object of each category.

[0052] Step 4: Set the training set and test set, train the key prediction factor combination for each prediction object, select the optimal hyperparameter combination, and obtain the optimal model for each prediction object as the final artificial intelligence model for medium- and long-term prediction of precipitation in the upper reaches of the Yangtze River.

[0053] In this embodiment, artificial intelligence model training is performed based on GPU.

[0054] Artificial intelligence models typically process data in batches to optimize learning results. Compared to CPUs, GPUs can greatly speed up training time. CPUs are designed to handle a variety of computing tasks and are optimized for sequential processing, while GPUs are designed for parallel processing and are equipped with thousands of smaller cores that work together to perform multiple operations at the same time. This makes GPUs particularly suitable for matrix and vector operations common in deep learning. GPUs have thousands of CUDA cores that can perform a large number of simple computing tasks at the same time. CUDA significantly accelerates a large number of matrix operations in deep learning models, such as convolution, pooling, and other operations through parallel computing, thereby improving the training speed and inference efficiency of the model and significantly shortening the training time of the model.

[0055] 1982-2012 and 2013-2023 are selected as the training period and test period respectively. The data of the training period and the test period correspond to the training set and the test set respectively. For each category of the target decade clustering, that is, each prediction object, the key prediction factor combination time series of each prediction object is used as the feature value. The model is fitted to the prediction object time series, and a hyperparameter space is constructed. After hyperparameter selection and training, a suitable optimal model is obtained, and the optimal prediction model for each category with a fixed prediction time is obtained.

[0056] After 6 trainings, 6 optimal prediction models for the precipitation in the upper reaches of the Yangtze River were obtained. Repeating steps 2 to 5, each key prediction factor combination and the corresponding optimal prediction model after training can be obtained for the prediction time of 10 to 60 days in advance.

[0057] Hyperparameters are optimized through the Optuna library. They are parameters that need to be set before deep learning models are trained, which are different from the weights and biases learned during model training.

[0058] In this embodiment, different values ​​are set for the delta value of the Huber loss function, the number of hidden units (hidden_size) of the GRU layer, the learning rate (learning_rate), the number of times the entire training set is traversed during training (num_epochs), the number of heads (num_heads) in the multi-head attention mechanism, the dropout ratio used in the multi-head attention mechanism, the negative slope parameter (negative_slope) of the LeakyReLU activation function, and the learning rate attenuation factor (gamma) to construct a hyperparameter space. Then, the optimal hyperparameter combination is formed by using the mean square error (MSE) as an indicator for evaluation in the test set.

[0059] Step 5: Provide real-time prediction results based on the combination of key prediction factors and the optimal prediction model.

[0060] According to the key prediction factor combination of the target ten-day prediction time of 10 to 60 days, the exponential combination of the key prediction factors of the real-time prediction year is calculated by using the stored key area and lead time information, and it is input as the characteristic value into the trained optimal prediction model to calculate the prediction value of various prediction objects in the target ten-day prediction time of 10 to 60 days.

[0061] These predicted values ​​are mapped to the corresponding category spatial fields obtained in step 1 to achieve the prediction of the average precipitation anomaly of the target decade 10 to 60 days in advance.

[0062] Embodiment 2:

[0063] Taking the forecast from early June 2023, the forecast of precipitation anomaly in the upper reaches of the Yangtze River in early July 2023 as an example, the forecast is made 30 days in advance. The specific process is as follows:

[0064] Step 1: Perform preprocessing steps on the daily data of historical weather in the upper reaches of the Yangtze River, cluster the data into 6 spatial field categories according to the established standards, and calculate the corresponding time series data. Among these spatial fields, the time series data that meet the established standards are selected as prediction objects. The detailed steps are as follows:

[0065] 1) Use CN 05.1 gridded observation data from January 1981 to May 2023 with a resolution of 0.25°X 0.25° to calculate the average daily precipitation in early July of each year;

[0066] 2) Calculate the precipitation anomaly, specifically the average of early July of the current year minus the climatological state, where the climatological state is the daily average of early July from 1991 to 2020;

[0067] 3) K-means clustering was performed on the precipitation anomaly data in early July in the upper reaches of the Yangtze River Basin to obtain 6 types of precipitation anomaly spatial fields with high interpretability. The corresponding precipitation anomalies in early July of each year were superimposed with latitude weights and the regional average was calculated to obtain the time series of each type of spatial field, and each type of time series was recorded as the prediction object.

[0068] Step 2: For different types of prediction objects, select different key area calculation indexes of each element as the key prediction factor combination based on statistics and historical research. The specific process is as follows:

[0069] 1) The prediction factor data used for training and testing comes from the global daily average data from January 1981 to May 2022 of the ERA5 reanalysis, with a resolution of 1°x1°. The variables include: outward longwave radiation (OLR), global sea surface temperature (SST), sea level pressure, 50 and 500hPa geopotential heights, and 700hPa horizontal wind field; and the average daily precipitation in early July of each year is calculated, and then the precipitation anomaly is calculated. The precipitation anomaly is the average of early July of the year minus the climate state to obtain the decadal anomaly data;

[0070] 2) Based on the historical study of summer precipitation in the upper reaches of the Yangtze River. The correlation coefficients of the time series of each type of spatial field with each predictor 10 to 60 days ahead of the start of the decade were calculated, and the reliability test was performed using the t-test to analyze the key areas with the highest correlation and significance. Based on the statistical results and historical research, the SST of the tropical Pacific Ocean at different lead times, the sea level pressure of the North Atlantic Ocean and the Sea of ​​Okhotsk, the 500hPa geopotential height of the North Atlantic Ocean, the North Pacific Ocean and the Sea of ​​Okhotsk, the 50hPa geopotential height of the North Atlantic Ocean, the North Pacific Ocean and the Sea of ​​Okhotsk, the 50hPa geopotential height of Europe and Canada, and the 700hPa horizontal wind field of the tropical Pacific Ocean, the tropical Atlantic Ocean and the North Atlantic Ocean were selected as the key areas of predictors. For each type of prediction object, each element selected multiple relevant and significant lead time periods and regions as the key areas of predictors, superimposed the latitude weights, calculated the regional average, and obtained the potential predictor time series combination. The predictor time series with the highest correlation coefficient between each type of predictor and the prediction object was used as the key predictor, and the key predictor time series were combined into a key predictor combination, and the key area position and lead time information of each key predictor were saved.

[0071] Step 3: Select the GRU-Attention deep learning model, adjust the hyperparameters, and train the model. The specific process is as follows:

[0072] 1) Because the precipitation anomalies in early July in the upper reaches of the Yangtze River are divided into 6 categories and there are 6 prediction objects, 6 key prediction factor combinations are selected for the forecast timeliness. It is necessary to model and train the 6 prediction objects separately and adjust the hyperparameters;

[0073] 2) In terms of time, the key predictor sequences and prediction objects from 1982 to 2012 are selected as training sets, and 2013 to 2022 are selected as test sets. A GRU-Attention deep learning model is constructed. The model is specifically divided into three layers: input layer, hidden layer and output layer. The input layer is a multi-year key predictor combination time series, the output layer is a prediction object sequence, and the hidden layer contains a GRU layer, an attention layer, an activation function layer and a fully connected layer.

[0074] 3) Using the Huber loss function, the Optuna hyperparameter optimization library was selected to define the hyperparameter search space, and the optimal parameters were automatically selected from the hyperparameters such as the delta value of the Huber loss function, the number of hidden units in the GRU layer (hidden_size), the learning rate (learning_rate), the number of times the entire training set is traversed during training (num_epochs), the number of heads in the multi-head attention mechanism (num_heads), the dropout ratio used in the multi-head attention mechanism, the negative slope parameter (negative_slope) of the LeakyReLU activation function, and the learning rate attenuation factor (gamma). After the training is completed, the model and the optimal parameters are saved to obtain the prediction models for all 6 trainings. The hyperparameter results of the six clustering tuning are shown in Table 1.

[0075]

[0076] Step 4: Real-time prediction and projection of the spatial field. The specific process is as follows:

[0077] 1) Since the prediction factors are selected 10 to 60 days in advance, the prediction factor data of ERA5 reanalysis from April to May 2023 are selected. According to the regional and lead time information of the key prediction factor combination of each cluster key area, the regional average key prediction factor index combination is calculated, with a total of six combinations;

[0078] 2) Input various combinations of key prediction factor indexes into various trained models, input the optimal parameters, and obtain the predicted precipitation anomaly index for various types in early July 2023. Project the index to the spatial field of the corresponding class obtained in step 2 to obtain the predicted precipitation anomaly in early July 2023 in the upper reaches of the Yangtze River.

[0079] In order to verify the implementation effect of this model, the GRU-Attention model of the present invention is compared with the ACC scoring technique of the ECWMF power mode prediction results. Figure 2 shown.

[0080] The anomaly correlation coefficient (ACC), a statistic that measures the quality of a forecast system, is used to evaluate the model forecast skill by calculating the correlation between the forecast and observation. The ACC value ranges from -1 to 1. The closer it is to 1, the more consistent the forecast and observation anomalies are, that is, the higher the forecast skill.

[0081] from Figure 2 It can be seen that, compared with the ECWMF dynamic model, the ACC value of the present invention is closer to 1, proving that the forecasting skill is higher than that of the ECWMF dynamic model.

[0082] The method of the present invention can train multiple prediction models for precipitation in the upper reaches of the Yangtze River within the sub-seasonal forecast time, and combine credible key prediction factors in different forecast time periods and different forecast periods. At the same time, the sub-seasonal scale precipitation forecast accuracy in the upper reaches of the Yangtze River is improved, and resource consumption is reduced.

[0083] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A medium- and long-term precipitation forecasting method for the Yangtze River Basin based on GRU-Attention, characterized in that: The following steps are involved: Step 1: Data preprocessing: Preprocess the historical precipitation daily data of the upper reaches of the Yangtze River within the set time period, cluster the data according to the set conditions, and record the time series corresponding to the spatial field that meets the set conditions as the prediction object; use sliding average and cluster analysis to preprocess the data, including the following steps: S101. Calculate the daily data as the ten-day average using a sliding average, and calculate the ten-day deviation data; S102. Use K-means clustering technology to reveal the spatial distribution characteristics of the temporal variation of precipitation in the upper reaches of the Yangtze River on a ten-day scale, identify areas with consistent temporal evolution patterns in space as clustering key areas, and for each clustering key area, take the average of the areas in each clustering key area as the prediction object; Step 2: Identify the key prediction factor combination of precipitation in the upper reaches of the Yangtze River; Step 3: Build the GRU-Attention deep learning model; Step 4: Set the training set and test set, train the key prediction factor combination of each prediction object, select the optimal hyperparameter combination, and obtain the optimal prediction model for each prediction object as the final artificial intelligence model for medium- and long-term prediction of precipitation in the upper reaches of the Yangtze River; Step 5: Provide real-time prediction results based on the combination of key prediction factors and the optimal prediction model.

2. According to claim 1, a method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention is characterized in that: In the step 2, one or more of the decadal scale data of outward longwave radiation, global sea surface temperature, geopotential height of sea level pressure at 50 hPa and 500 hPa, and horizontal wind field at 700 hPa are selected as prediction factors.

3. According to claim 2, a method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention is characterized in that: In the step 2, for each spatial field of precipitation in the upper reaches of the Yangtze River, the correlation coefficient between each cluster key area and the prediction factor of more than 10 to 60 days ahead is calculated respectively, the key areas of different prediction factors with different prediction time effectiveness are clarified, and the regional average value of the significant area on the correlation coefficient graph is calculated as the prediction factor time series, that is, the prediction factor index series, and each prediction factor time series is combined into a potential prediction factor time series combination, and the prediction factor time series with the highest correlation with the prediction object in each type of prediction factor is analyzed and screened, and the screened prediction factor time series are combined into the key prediction factor combination of the prediction object of the cluster key area, and the above operation is repeated to obtain the key prediction factor combination of all cluster key areas.

4. According to claim 1, a method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention is characterized in that: In the step three, a GRU gated recurrent neural network is selected as the prediction model, and an attention mechanism is added to form a GRU-Attention deep learning model.

5. According to claim 4, a method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention is characterized in that: The GRU-Attention deep learning model is divided into three layers: input layer, hidden layer and output layer; At the input layer, the input value at each moment is the key predictor index sequence at the moment in the key predictor combination; The hidden layer consists of three parts. The first layer is the GRU layer. The first dimension of the input and output tensors is the batch size. It is responsible for processing the input sequence and capturing the dependencies in the time series data. The second layer is a multi-head attention mechanism layer, which uses the specified number of heads and dropout ratio to operate on the output of the GRU layer to further extract features and enhance the model's ability to learn the relationship between different parts of the sequence; the third layer is the LeakyReLU activation function layer; The final output layer is a fully connected layer, which maps the features processed by the LeakyReLU activation function to the final output. It contains only one neuron and outputs a value at each moment, which is the predicted object of each cluster key area.

6. The method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention according to claim 1 is characterized in that: In the step 4, a training period and a test period are constructed, and the data of the training period and the test period correspond to the training set and the test set respectively. The key prediction factor combination time series of each prediction object is used as the characteristic value, and the prediction object time series is fitted by the model to construct a hyperparameter space. After the hyperparameter selection training, a suitable optimal model is obtained, and the optimal prediction model for each cluster key area with a fixed prediction time is obtained.

7. The method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention according to claim 6 is characterized in that: The hyperparameter space is constructed by setting different values ​​for the delta value of the Huber loss function, the number of hidden units of the GRU layer, the learning rate, the number of times the entire training set is traversed during training, the number of heads in the multi-head attention mechanism, the dropout ratio used in the multi-head attention mechanism, the negative slope parameter of the LeakyReLU activation function, and the learning rate attenuation factor.

8. The method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention according to claim 6 is characterized in that: The hyperparameters are selected to form the optimal hyperparameter combination using mean square error as an evaluation indicator in the test set.

9. The method for medium- and long-term precipitation forecasting in the Yangtze River Basin based on GRU-Attention according to claim 1 is characterized in that: In the step five, based on the key prediction factor combination of the target ten-day prediction time of 10 to 60 days, the stored key area and lead time information are used to calculate the exponential combination of the real-time prediction year key prediction factors, and the exponential combination is input as a characteristic value into the trained optimal prediction model to calculate the prediction value of each prediction object in the target ten-day prediction time of 10 to 60 days.

Images