Rainfall forecasting method and system based on improved GRU multi-source information fusion model

Abstract

The invention provides a rainfall forecasting method and system based on an improved GRU multi-source information fusion model. The method comprises the following steps: acquiring radar reflectivity data; performing dynamic grading Z-I relation conversion on the radar data to generate a rainfall intensity sequence; multi-source information is extracted, wherein the multi-source information comprises a rainfall station actually-measured rainfall sequence and meteorological auxiliary data sequences such as temperature and humidity; carrying out adaptive fusion on multi-source information through a trainable feature weight mechanism, and constructing a sample set; dividing the sample set into a training set and a test set; processing and enhancing the training set data; constructing an improved GRU neural network model fused with an attention mechanism; training the neural network model by using the training set, and verifying the performance of the model through the test set; and inputting actually measured multi-source data into the trained model, and outputting a short-term rainfall forecast value at the time scale of the next five minutes. According to the method, the problems of rainfall station data missing and radar data noise interference are effectively solved.

Description

Technical Field

[0001] This invention proposes a rainfall forecasting method and system based on an improved GRU multi-source information fusion model, which relates to the fields of meteorological forecasting and hydrological information technology. Background Technology

[0002] Precipitation is a core element in meteorological and hydrological simulation. High-precision, high spatiotemporal resolution rainfall information is of vital importance for flash flood warnings in small watersheds, precise water resource management, and water conservancy project scheduling.

[0003] Traditionally, ground-based rain gauges have been considered the "true value" for precipitation monitoring. However, in mountainous areas with complex terrain, these stations are sparsely distributed and prone to equipment failure or data loss under extreme weather conditions. Although satellite remote sensing precipitation products (such as GPM and TRMM) can provide large-scale, high-frequency observations, their spatial resolution is relatively coarse and they have significant errors in complex terrain areas, making it difficult to meet the needs of small watersheds for accurate capture of spatial variability in precipitation.

[0004] In recent years, weather radar has become a key data source for short-term precipitation forecasting, providing high spatiotemporal resolution reflectivity information. However, radar data is susceptible to ground clutter, beam obstruction, and atmospheric interference, and traditional ZI relationship transformation methods perform poorly in precipitation scenarios with drastic intensity changes. Furthermore, existing data fusion research largely focuses on large-scale, long-term time series (such as monthly and yearly scales), with a lack of research on small watershed, short-term, and especially minute-level (such as 5-minute) forecasts. The reasons for this are as follows:

[0005] The ability to fuse multi-source heterogeneous data (radar, rain gauge, meteorological elements) is insufficient, and the differences in the importance of different element characteristics are not fully considered;

[0006] Most existing studies have failed to fully consider the unique characteristics of small watersheds in mountainous areas—short confluence time and rapid response, requiring extremely high timeliness and accuracy in forecasts. Forecast scales of 1-3 hours are insufficient to meet the actual needs of flood warnings in small watersheds in mountainous areas, while higher-resolution 5-minute short-term forecasts face challenges such as high data noise and strong spatiotemporal variability.

[0007] The model is prone to overfitting on limited samples, exhibits poor generalization ability, and struggles to adapt to the sudden and localized characteristics of rainfall in mountainous areas. Summary of the Invention

[0008] In view of this, in order to fill the gaps and deficiencies in the existing technology, this invention proposes a rainfall forecasting method and system based on an improved GRU multi-source information fusion model. This method addresses the problems in existing short-term rainfall forecasting for small watersheds in mountainous areas, such as insufficient fusion of multi-source heterogeneous data, insufficient model generalization ability, and failure to fully consider the special characteristics of short runoff times and rapid response in small watersheds.

[0009] This invention proposes a rainfall forecasting method and system based on an improved GRU multi-source information fusion model, comprising the following:

[0010] This invention proposes a rainfall forecasting method based on an improved GRU multi-source information fusion model, characterized by the following:

[0011] Step S1: Obtain radar reflectivity factor data;

[0012] Step S2: Perform dynamic ZI relationship transformation on radar reflectivity factor data to generate precipitation intensity sequence;

[0013] Step S3: Obtain multi-source information, including: radar precipitation sequence, historical precipitation sequence from rain gauge stations, and auxiliary sequences for temperature and humidity;

[0014] Step S4: Adaptively fuse multi-source information through a feature weighting mechanism to construct a sample set;

[0015] Step S5: Divide the sample set into a training set and a test set;

[0016] Step S6: Perform data augmentation on the training sample set;

[0017] Step S7: Construct an improved GRU neural network model;

[0018] Step S8: Train the improved GRU neural network model using the training set and test it using the test set to obtain the final trained model;

[0019] Step S9: Input the input data to be predicted into the trained model to obtain the final rainfall forecast value.

[0020] Further, step S2 includes the following:

[0021] Step S21: Centered on the site to be predicted, convert the raw radar reflectivity factor data, expressed in decibels, into linear data Z using the following formula:

[0022] (1);

[0023] Step S22: Using the dynamic grading ZI relationship method, the transformed data Z is processed according to its intensity level k, and a preset coefficient a for that level is selected. k and b k The precipitation intensity I is calculated using the following formula (2). t for:

[0024] (2);

[0025] Where t represents the radar reflectivity and precipitation intensity at time t;

[0026] Step S23: Repeat formula (2) for all pixels within the radar detection range to finally generate a quantitative rainfall estimation sequence I=[I1, I2, ..., I...] that is consistent with the radar observation range and resolution. t ].

[0027] Further, step S4 includes the following:

[0028] In step S4, the calculation process for the adaptive weights of the multi-source feature allocation is as follows:

[0029] Step S41: Define the original input vector containing F features. At the same time, initialize a trainable weight vector with the same dimension. ;

[0030] Step S42: Use the Sigmoid function to constrain each original weight parameter to the interval (0,1). The formula for calculating the normalized weight vector is as follows:

[0031] (3);

[0032] Where σ() represents the Sigmoid function; where w i Represents a single weight vector;

[0033] Step S43: Convert the normalized weight vector Compared with the original input feature vector The weighted eigenvector is calculated by performing element-wise multiplication, as follows:

[0034] (4);

[0035] The weighted feature vectors will serve as the actual input to the subsequent improved GRU network.

[0036] Step S44: Concatenate the radar precipitation sequence and the historical precipitation sequence from rain gauge stations, and use the result as the input matrix X of the improved GRU neural network model. t , where X t As shown in the following formula:

[0037] (5);

[0038] (6);

[0039] in, For actual precipitation data, This refers to precipitation data from rain gauge radar. The ambient temperature of the watershed, For ambient humidity; Let M be the input vector, M be the time step of the input matrix, and F be the number of variables in the input vector.

[0040] Further, step S6 includes the following:

[0041] Step S61: Normalize and augment the training set using the softmax method to obtain the normalized training set. and the enhanced training set The normalization process is represented by the following formula:

[0042] (7);

[0043] Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence;

[0044] Step S62: Apply sliding window sampling to the normalized sequence, and perform overlapping sliding window sampling to generate training sub-samples. The process of generating training sub-samples is represented by the following formula:

[0045] (8);

[0046] Where L is the window length and S is the sliding step size;

[0047] Step S63: Inject small amounts of random noise into each sample point to simulate data acquisition errors and enhance the robustness of the model. The enhancement process is represented by the following formula:

[0048] (9);

[0049] Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence; where L is the window length and S is the sliding step size.

[0050] Further, step S7 includes the following:

[0051] Step S71: Construct an improved GRU neural network model, including the following:

[0052] When the window length is L, and the number of sliding steps is 1, from... Extract continuous data of length L and use it as the training set X for the improved GRU network.train Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. train The label Y train At the same time from X text Extract continuous data of length L and use it as the test set X for improving the GRU network. text Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. text The label Y test The training and test sets are used as normalized historical data.

[0053] Further, step S8 includes the following:

[0054] Step S81: Set the number of training iterations for the improved GRU network to epochs, and use the mean squared error mse as the loss function. When the loss gradually decreases and stabilizes during the iteration process, the training can be considered complete.

[0055] Step S82: Transfer the training set The initial hidden layer state h0, together with the initial state, is input into the improved GRU network; the process of inputting into the improved GRU network includes the following:

[0056] Step S821: The process of inputting into the improved GRU network includes formulas (10) to (17); wherein formulas (10) to (17) include the following:

[0057] (10);

[0058] (11);

[0059] (12);

[0060] (13);

[0061] (14);

[0062] (15);

[0063] (16);

[0064] (17);

[0065] In formulas (10) to (17), W z and b z W represents the weight matrix and bias vector of the update gate. r and b rThis represents the weight matrix and bias vector of the reset gate, where W and b represent the weight matrix and bias vector of the candidate state. o and b o h represents the weight matrix bias vector of the output layer. t-1 Let x represent the hidden state of the network at time t-1. t Let represent the input feature vector of the network at time t, σ() is the sigmoid function transformation, and ⊙ represents element-wise multiplication;

[0066] Step S822: The process of inputting into the improved GRU network includes executing formulas (10) to (17); wherein executing formulas (10) to (17) includes the following:

[0067] First, the output value z of the update gate is calculated according to formula (10). t Secondly, the output value r of the reset door is calculated according to formula (11). t Then, the candidate hidden states are calculated according to formula (12). Then, the candidate hidden state is updated according to formula (13), and an attention mechanism is introduced on this basis. The attention weight a at each time step is calculated according to formulas (14) and (15). t Next, the hidden states are weighted and summed to obtain the context vector c according to formula (16); finally, the context vector c is passed through the output layer to obtain the final output result y according to formula (17). t .

[0068] Further, step S8 includes the following:

[0069] Step S83: Using a time-based backpropagation algorithm, predict the output value Y by calculating the model. * train Compared with the real label Y train The error between the parameters is used to calculate gradients and optimize updates for all parameters in the improved GRU network. Through continuous iteration, the trained improved GRU network is obtained.

[0070] Further, step S8 includes the following:

[0071] Step S91: Use the input data to be predicted as the input matrix X per The input is fed into the pre-trained feature weight module, and the final weight vector obtained during training is applied. Feature scaling yields the weighted real-time input matrix:

[0072] ;

[0073] The input is then fed into the trained improved GRU neural network model for forward propagation calculation to obtain the normalized output value. and to Inverse normalization is performed to obtain the estimated combined rainfall from rain gauges over a future period. .

[0074] According to a second aspect of the present invention, a rainfall forecasting system based on an improved GRU multi-source information fusion model includes an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements a rainfall forecasting method based on an improved GRU multi-source information fusion model as described in any one of the present invention.

[0075] According to a third aspect of the present invention, a rainfall forecasting system based on an improved GRU multi-source information fusion model includes a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements a rainfall forecasting method based on an improved GRU multi-source information fusion model as described in any one of the present invention.

[0076] The present invention has the following advantages:

[0077] (1) A deep learning framework that integrates multi-source heterogeneous data and physical constraints is proposed. By introducing dynamic hierarchical ZI relationship, radar reflectivity data is converted into precipitation intensity sequence with clear hydrological significance, realizing a high degree of coupling between radar electromagnetic wave scattering characteristics and precipitation physical process.

[0078] (2) A trainable feature weighting mechanism was designed to adaptively balance the contributions of multiple input sources such as radar quantitative precipitation estimation, rain gauge measured data and meteorological auxiliary environmental variables, effectively overcoming the uncertainty of a single data source and the data loss caused by equipment failure.

[0079] (3) An improved GRU neural network architecture was constructed, which, combined with attention mechanism and data augmentation strategy, significantly improved the model’s ability to capture and forecast the spatiotemporal distribution characteristics of precipitation in complex terrain, especially the center of short-term heavy precipitation and rainstorm.

[0080] In summary, the beneficial effects of this invention are: it effectively overcomes the data loss problem caused by rain gauge equipment failure or sparse distribution, achieving efficient fusion and complementarity of multi-source precipitation data. Simultaneously, this invention can provide high-precision short-term rainfall forecasts on a 5-minute timescale, significantly enhancing the ability to capture and forecast localized sudden rainstorms in small mountain watersheds. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the steps of the present invention.

[0082] Figure 2 This is a structural diagram of the fusion model of the present invention.

[0083] Figure 3 The image shows contour maps of rainfall forecasts for three rainfall events at a 5-minute timescale, based on different models of this invention. Detailed Implementation

[0084] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0085] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0086] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0087] like Figures 1 to 3 As shown, this invention proposes a rainfall forecasting method and system based on an improved GRU multi-source information fusion model, characterized by the following:

[0088] like Figure 1 As shown, this invention proposes a rainfall forecasting method based on an improved GRU multi-source information fusion model, characterized by the following:

[0089] Step S1: Obtain radar reflectivity factor data;

[0090] Step S2: Perform dynamic ZI relationship transformation on radar reflectivity factor data to generate precipitation intensity sequence;

[0091] Step S3: Obtain multi-source information, including: radar precipitation sequence, historical precipitation sequence from rain gauge stations, and auxiliary sequences for temperature and humidity;

[0092] Step S4: Adaptively fuse multi-source information through a feature weighting mechanism to construct a sample set;

[0093] Step S5: Divide the sample set into a training set and a test set;

[0094] Step S6: Perform data augmentation on the training sample set;

[0095] Step S7: Construct an improved GRU neural network model;

[0096] Step S8: Train the improved GRU neural network model using the training set and test it using the test set to obtain the final trained model;

[0097] Step S9: Input the input data to be predicted into the trained model to obtain the final rainfall forecast value.

[0098] In one embodiment of the present invention, the radar reflectivity factor data obtained in step S1 includes the observed raw reflectivity factor dBZ data.

[0099] In one embodiment of the present invention, step S2 further includes the following:

[0100] In one embodiment of the present invention, step S2 involves selecting n rainfall stations within the watershed. As the estimation target of this invention.

[0101] Step S21: Centered on the site to be predicted, convert the raw radar reflectivity factor data, expressed in decibels (dBZ), into linear data Z using the following formula:

[0102] (1);

[0103] Step S22: Using the dynamic grading ZI relationship method, the transformed data Z is classified according to its intensity level k (usually 10-75dBZ is divided into 13 levels in 5dBZ intervals), and the preset coefficient a for that level is selected. k and b k The precipitation intensity I is calculated using the following formula (2). t for:

[0104] (2);

[0105] Where t represents the radar reflectivity and precipitation intensity at time t;

[0106] Step S23: Repeat formula (2) for all pixels within the radar detection range to finally generate a quantitative rainfall estimation sequence I=[I1, I2, ..., I...] that is consistent with the radar observation range and resolution. t ].

[0107] In one embodiment of the present invention, the multi-source information obtained in step S3 further includes: historical precipitation sequences of rainfall stations;

[0108] In one embodiment of the present invention, step S4 further includes the following:

[0109] In step S4, the calculation process for the adaptive weights of the multi-source feature allocation is as follows:

[0110] Step S41: Define the original input vector containing F features:

[0111] At the same time, initialize a trainable weight vector with the same dimension. ;

[0112] Step S42: Use the Sigmoid function to constrain each original weight parameter to the interval (0,1). The formula for calculating the normalized weight vector is as follows:

[0113] (3);

[0114] Where σ() represents the Sigmoid function; where w i Represents a single weight vector;

[0115] Step S43: Convert the normalized weight vector Compared with the original input feature vector The weighted eigenvector is calculated by performing element-wise multiplication, as follows:

[0116] (4);

[0117] The weighted feature vectors will serve as the actual input to the subsequent improved GRU network.

[0118] Step S44: Concatenate the radar precipitation sequence and the historical precipitation sequence from rain gauge stations, and use the result as the input matrix X of the improved GRU neural network model. t , where X t As shown in the following formula:

[0119] (5);

[0120] (6);

[0121] in, For actual precipitation data, This refers to precipitation data from rain gauge radar. The ambient temperature of the watershed, For ambient humidity; Let M be the input vector, M be the time step of the input matrix, and F be the number of variables in the input vector.

[0122] In one embodiment of the present invention, step S5 divides the sample set into a training set X. train With test set X text Of these, 85% of the data was used as the training sample set, and the remaining data was used as the test sample set.

[0123] In one embodiment of the present invention, step S6 further includes the following:

[0124] Step S61: Before inputting the training set into the improved GRU neural network model, the training set is normalized and augmented using the softmax method; the normalized training set is obtained by normalizing the training set using the softmax method. and the enhanced training set The normalization process is represented by the following formula:

[0125] (7);

[0126] Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence;

[0127] Step S62: Apply sliding window sampling to the normalized sequence, and perform overlapping sliding window sampling to generate training sub-samples. The process of generating training sub-samples is represented by the following formula:

[0128] (8);

[0129] Where L is the window length and S is the sliding step size;

[0130] Step S63: Inject small amounts of random noise into each sample point to simulate data acquisition errors and enhance the robustness of the model. The enhancement process is represented by the following formula:

[0131] (9);

[0132] Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence; where L is the window length and S is the sliding step size.

[0133] In one embodiment of the present invention, such as Figure 2 The improved GRU neural network model structure diagram is shown in Figure S7. Step S7 includes the following:

[0134] Step S71: Construct an improved GRU neural network model, including the following:

[0135] In one embodiment of the present invention, the present invention employs a sliding window method, wherein the window length is L.

[0136] When the window length is L, and the number of sliding steps is 1, from... Extract continuous data of length L and use it as the training set X for the improved GRU network. train Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. train The label Y train At the same time from X text Extract continuous data of length L and use it as the test set X for improving the GRU network. text Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. text The label Y test The training and test sets are used as normalized historical data.

[0137] Further, step S8 includes the following:

[0138] Step S81: Set the number of training iterations for the improved GRU network to epochs, and use the mean squared error mse as the loss function. When the loss gradually decreases and stabilizes during the iteration process, the training can be considered complete.

[0139] Step S82: Transfer the training set The initial hidden layer state h0, together with the initial state, is input into the improved GRU network; the process of inputting into the improved GRU network includes the following:

[0140] Step S821: The process of inputting into the improved GRU network includes formulas (10) to (17); wherein formulas (10) to (17) include the following:

[0141] (10);

[0142] (11);

[0143] (12);

[0144] (13);

[0145] (14);

[0146] (15);

[0147] (16);

[0148] (17);

[0149] In formulas (10) to (17), W z and b z W represents the weight matrix and bias vector of the update gate. r and b r This represents the weight matrix and bias vector of the reset gate, where W and b represent the weight matrix and bias vector of the candidate state. o and b o h represents the weight matrix bias vector of the output layer. t-1 Let x represent the hidden state of the network at time t-1. t Let represent the input feature vector of the network at time t, σ() is the sigmoid function transformation, and ⊙ represents element-wise multiplication;

[0150] Step S822: The process of inputting into the improved GRU network includes executing formulas (10) to (17); wherein executing formulas (10) to (17) includes the following:

[0151] First, the output value z of the update gate is calculated according to formula (10). t Secondly, the output value r of the reset door is calculated according to formula (11). t Then, the candidate hidden states are calculated according to formula (12). Then, the candidate hidden state is updated according to formula (13), and an attention mechanism is introduced on this basis. The attention weight a at each time step is calculated according to formulas (14) and (15). t Next, the hidden states are weighted and summed to obtain the context vector c according to formula (16); finally, the context vector c is passed through the output layer to obtain the final output result y according to formula (17). t .

[0152] Further, step S8 includes the following:

[0153] Step S83: Using a time-based backpropagation algorithm, predict the output value Y by calculating the model. * train Compared with the real label Y train The error between the parameters is used to calculate gradients and optimize updates for all parameters in the improved GRU network. Through continuous iteration, the trained improved GRU network is obtained.

[0154] Further, step S8 includes the following:

[0155] Step S91: Use the input data to be predicted as the input matrix X perThe input is fed into the pre-trained feature weight module, and the final weight vector obtained during training is applied. Feature scaling is performed to obtain the weighted real-time input matrix.

[0156] ;

[0157] The input is then fed into the trained improved GRU neural network model for forward propagation calculation to obtain the normalized output value. and to Inverse normalization is performed to obtain the estimated combined rainfall from rain gauges over a future period. .

[0158] In addition to the above, the present invention also has related embodiments, including the following:

[0159] In one embodiment of the present invention, the present invention uses 24 heavy precipitation events recorded by four controlled rain gauge stations within the study basin from 2024 to 2025, divided into training and test sets according to the data volume, with a data volume ratio of 8:2. All four rain gauge stations within the basin are used as estimation targets. To illustrate the superiority of the improved GRU model of the present invention, the GRU model is used as a comparative model below. The estimation results of the test set are analyzed using the two trained models to demonstrate the advantages of the present invention compared to other models.

[0160] Table 1 shows the evaluation metrics of the model. The results indicate that for radar precipitation data with significant uncertainties, the improved GRU model greatly improves radar precipitation errors. The improved GRU model shows the most significant improvement in the estimation accuracy of the three targets, as detailed below:

[0161] Table 1 Comparison of accuracy of different estimation methods

[0162] R2 NSE MSE GRU model 0.77 0.71 0.40 Improved GRU model 0.87 0.83 0.42

[0163] Accumulated typhoon precipitation distribution maps for this watershed were generated by applying different spatial downscaling methods to satellite precipitation data. It should be noted that... Figure 3 The top-middle figure shows the GRU model, and the bottom figure shows the improved GRU model. In terms of capturing complex rainfall processes and flood peak morphology, the GRU model fails to capture concentrated rainfall areas and its trend characterization is inaccurate. The improved GRU model, on the other hand, can better simulate the overall trend of rainfall processes, especially at key time points such as the onset, peak, and dissipation of rainfall, demonstrating higher prediction accuracy and significantly improving its ability to predict extreme rainfall events and its spatial detail reproducibility.

[0164] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A rainfall forecasting method based on an improved GRU multi-source information fusion model, characterized in that, Includes the following: Step S1: Obtain radar reflectivity factor data; Step S2: Perform dynamic ZI relationship transformation on radar reflectivity factor data to generate precipitation intensity sequence; Step S3: Obtain multi-source information, including: radar precipitation sequence, historical precipitation sequence from rain gauge stations, and auxiliary sequences for temperature and humidity; Step S4: Adaptively fuse multi-source information through a feature weighting mechanism to construct a sample set; Step S5: Divide the sample set into a training set and a test set; Step S6: Perform data augmentation on the training sample set; Step S7: Construct an improved GRU neural network model; Step S8: Train the improved GRU neural network model using the training set and test it using the test set to obtain the final trained model; Step S9: Input the input data to be predicted into the trained model to obtain the final rainfall forecast value.

2. The rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 1, characterized in that, Step S2 includes the following: Step S21: Centered on the site to be predicted, convert the raw radar reflectivity factor data, expressed in decibels, into linear data Z using the following formula: (1) ； Step S22: Using the dynamic grading ZI relationship method, the transformed data Z is processed according to its intensity level k, and a preset coefficient a for that level is selected. k and b k The precipitation intensity I is calculated using the following formula (2). t for: (2) ； Where t represents the radar reflectivity and precipitation intensity at time t; Step S23: Repeat formula (2) for all pixels within the radar detection range to finally generate a quantitative rainfall estimation sequence I=[I1, I2, ..., I...] that is consistent with the radar observation range and resolution. t ].

3. The rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 2, characterized in that, Step S4 includes the following: In step S4, the calculation process for the adaptive weights of the multi-source feature allocation is as follows: Step S41: Define the original input vector containing F features as follows: ， Simultaneously initialize a trainable weight vector of the same dimension. ; Step S42: Use the Sigmoid function to constrain each original weight parameter to the interval (0,1). The formula for calculating the normalized weight vector is as follows: (3)； Where σ() represents the Sigmoid function; where w i Represents a single weight vector; Step S43: Convert the normalized weight vector Compared with the original input feature vector The weighted eigenvector is calculated by performing element-wise multiplication, as follows: (4)； The weighted feature vectors will serve as the actual input to the subsequent improved GRU network. Step S44: Concatenate the radar precipitation sequence and the historical precipitation sequence from rain gauge stations, and use the result as the input matrix X of the improved GRU neural network model. t , where X t As shown in the following formula: (5)； (6)； in, For actual precipitation data, This refers to precipitation data from rain gauge radar. The ambient temperature of the watershed, For ambient humidity; Let M be the input vector, M be the time step of the input matrix, and F be the number of variables in the input vector.

4. The rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 3, characterized in that, Step S6 includes the following: Step S61: Normalize and augment the training set using the softmax method to obtain the normalized training set. and the enhanced training set The normalization process is represented by the following formula: (7)； Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence; Step S62: Apply sliding window sampling to the normalized sequence, and perform overlapping sliding window sampling to generate training sub-samples. The process of generating training sub-samples is represented by the following formula: (8)； Where L is the window length and S is the sliding step size; Step S63: Inject small amounts of random noise into each sample point to simulate data acquisition errors and enhance the robustness of the model. The enhancement process is represented by the following formula: (9)； Where x* is the normalized data sequence; x is the original data sequence; x min and x max These are the minimum and maximum values ​​in the original data sequence; where L is the window length and S is the sliding step size.

5. A rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 4, characterized in that, Step S7 includes the following: Step S71: Construct an improved GRU neural network model, including the following: When the window length is L, and the number of sliding steps is 1, from... Extract continuous data of length L and use it as the training set X for the improved GRU network. train Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. train The label Y train At the same time from X text Extract continuous data of length L and use it as the test set X for improving the GRU network. text Extract the (L+1)th data point from the historical precipitation sequence of rainfall stations as X. text The label Y test The training and test sets are used as normalized historical data.

6. A rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 5, characterized in that, Step S8 includes the following: Step S81: Set the number of training iterations for the improved GRU network to epochs, and use the mean squared error mse as the loss function. When the loss gradually decreases and stabilizes during the iteration process, the training can be considered complete. Step S82: Transfer the training set The initial hidden layer state h0, together with the initial state, is input into the improved GRU network; the process of inputting into the improved GRU network includes the following: Step S821: The process of inputting into the improved GRU network includes formulas (10) to (17); wherein formulas (10) to (17) include the following: (10)； (11)； (12)； (13)； (14)； (15)； (16)； (17)； In formulas (10) to (17), W z and b z W represents the weight matrix and bias vector of the update gate. r and b r This represents the weight matrix and bias vector of the reset gate, where W and b represent the weight matrix and bias vector of the candidate state. o and b o h represents the weight matrix bias vector of the output layer. t-1 Let x represent the hidden state of the network at time t-1. t Let represent the input feature vector of the network at time t, σ() is the sigmoid function transformation, and ⊙ represents element-wise multiplication; Step S822: The process of inputting into the improved GRU network includes executing formulas (10) to (17); wherein executing formulas (10) to (17) includes the following: First, the output value z of the update gate is calculated according to formula (10). t Secondly, the output value r of the reset door is calculated according to formula (11). t Then, the candidate hidden states are calculated according to formula (12). Then, the candidate hidden state is updated according to formula (13), and an attention mechanism is introduced on this basis. The attention weight a at each time step is calculated according to formulas (14) and (15). t Next, the hidden states are weighted and summed to obtain the context vector c according to formula (16); finally, the context vector c is passed through the output layer to obtain the final output result y according to formula (17). t .

7. A rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 6, characterized in that, Step S8 also includes the following: Step S83: Using a time-based backpropagation algorithm, predict the output value Y by calculating the model. * train Compared with the real label Y train The error between the parameters is used to calculate gradients and optimize updates for all parameters in the improved GRU network. Through continuous iteration, the trained improved GRU network is obtained.

8. A rainfall forecasting method based on an improved GRU multi-source information fusion model according to claim 7, characterized in that, Step S8 includes the following: Step S91: Use the input data to be predicted as the input matrix X per The input is fed into the pre-trained feature weight module, and the final weight vector obtained during training is applied. Feature scaling is performed to obtain the weighted real-time input matrix. ； The input is then fed into the trained improved GRU neural network model for forward propagation calculation to obtain the normalized output value. and to Inverse normalization is performed to obtain the estimated combined rainfall from rain gauges over a future period. .

9. A rainfall forecasting system based on an improved GRU multi-source information fusion model, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a rainfall forecasting method based on an improved GRU multi-source information fusion model as described in any one of claims 1 to 8.

10. A rainfall forecasting system based on an improved GRU multi-source information fusion model, comprising a computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a rainfall forecasting method based on an improved GRU multi-source information fusion model as described in any one of claims 1 to 8.

Images