Search

CN-121997269-A - Rainfall forecasting method and system based on improved GRU multisource information fusion model

CN121997269ACN 121997269 ACN121997269 ACN 121997269ACN-121997269-A

Abstract

The invention provides a rainfall forecasting method and system based on an improved GRU multisource information fusion model, which comprises the steps of obtaining radar reflectivity data, carrying out dynamic grading Z-I relation conversion on the radar data to generate a rainfall intensity sequence, extracting multisource information including weather auxiliary data sequences such as a rainfall station actually measured rainfall sequence and temperature and humidity, carrying out self-adaptive fusion on the multisource information through a trainable characteristic weight mechanism to construct a sample set, dividing the sample set into a training set and a testing set, processing and enhancing training set data, constructing an improved GRU neural network model fused with an attention mechanism, training the neural network model through the training set, verifying model performance through the testing set, inputting the actually measured multisource data into the trained model, and outputting a short-cut rainfall forecasting value of a future 5-minute time scale. The invention effectively solves the problems of data missing of the rainfall station and radar data noise interference.

Inventors

  • ZHANG TING
  • WANG HONGNA
  • Yang Dingying
  • ZHAN CHANGXUAN

Assignees

  • 福州大学

Dates

Publication Date
20260508
Application Date
20260131

Claims (10)

  1. 1. The rainfall forecasting method based on the improved GRU multisource information fusion model is characterized by comprising the following steps of: s1, acquiring radar reflectivity factor data; s2, carrying out dynamic Z-I relation conversion on radar reflectivity factor data to generate a precipitation intensity sequence; step S3, multi-source information is obtained, wherein the multi-source information comprises a radar precipitation sequence, a rainfall station history precipitation sequence and an air temperature and humidity auxiliary sequence; S4, carrying out self-adaptive fusion on the multi-source information through a characteristic weight mechanism to construct a sample set; s5, dividing the sample set into a training set and a testing set; s6, carrying out data enhancement processing on the training sample set; s7, constructing an improved GRU neural network model; s8, training the improved GRU neural network model by using a training set, and testing by using a testing set to obtain a final trained model; and S9, inputting the input data to be predicted into the trained model to obtain a final rainfall forecast value.
  2. 2. The rainfall forecasting method based on the improved GRU multisource information fusion model of claim 1, wherein step S2 comprises the following steps: Step S21, taking a station to be predicted as a center, converting the acquired original radar reflectivity factor data expressed in decibel units into linear data Z according to the following formula: (1) ; step S22, adopting a dynamic grading Z-I relation method, and calculating precipitation intensity I t of converted data Z according to intensity grade k of the converted data Z and selecting coefficients a k and b k preset by the grade, wherein the precipitation intensity I t is calculated by the following formula (2): (2) ; Wherein t represents radar reflectivity and precipitation intensity corresponding to the t moment; And S23, repeating the calculation of the formula (2) on all pixels in the radar detection range, and finally generating a quantitative rainfall estimation sequence I= [ I 1 ,I 2 ,... ,I t ] consistent with the radar observation range and the resolution.
  3. 3. The rainfall forecasting method based on the improved GRU multisource information fusion model of claim 2, wherein step S4 comprises the following steps: In step S4, the calculation process of the adaptive weight of the multi-source feature allocation is as follows: step S41, defining an original input vector containing F features as follows: , Simultaneously initializing a trainable weight vector of the same dimension ; Step S42, restraining each original weight parameter in a (0, 1) interval by using a Sigmoid function, wherein a normalized weight vector calculation formula is as follows: (3); Wherein σ () represents a Sigmoid function, wherein w i represents a single weight vector; Step S43, normalizing the weight vector And the original input characteristic vector And multiplying the elements to obtain a weighted eigenvector calculation formula as follows: (4); Wherein the weighted feature vector is used as an actual input for subsequently improving the GRU network; Step S44, splicing the radar precipitation sequence and the rainfall station historical precipitation sequence to be used as an input matrix X t for improving the GRU neural network model, wherein X t is as follows: (5); (6); Wherein, the For the actual measurement of the precipitation data, Radar precipitation data for the rainfall station, For the ambient temperature of the basin, Is the ambient humidity; for an input vector, M is the time step of the input matrix and F is the number of variables in the input vector.
  4. 4. A rainfall forecasting method based on an improved GRU multisource information fusion model according to claim 3, characterized in that step S6 comprises the following: Step S61, adopting a softmax method to normalize and enhance data of the training set to obtain a normalized training set And an enhanced training set The normalization process is expressed by the following formula: (7); Wherein x is the normalized data sequence, x is the original data sequence, and x min and x max are the minimum and maximum values in the original data sequence; step S62, sampling the normalized sequence by adopting a sliding window, and performing overlapping sliding window sampling to generate a training sub-sample, wherein the process of generating the training sub-sample is represented by the following formula: (8); Wherein L is the window length, S is the sliding step length; Step S63, injecting tiny random noise into each sample point, simulating data acquisition errors, and enhancing model robustness, wherein the enhancement process is expressed by the following formula: (9); Wherein x is the normalized data sequence, x is the original data sequence, x min and x max are the minimum and maximum values in the original data sequence, wherein L is the window length and S is the sliding step length.
  5. 5. The rainfall forecasting method based on the improved GRU multisource information fusion model of claim 4, wherein step S7 comprises the following steps: step S71, constructing an improved GRU neural network model, which comprises the following steps: when the window length is L, the sliding step number is set to be 1 each time, and the window is selected from Extracting continuous data with length L, taking the continuous data as a training set X train of an improved GRU network, extracting the (L+1) th data of a rainfall site historical precipitation sequence as a tag Y train of X train , simultaneously extracting the continuous data with length L from X text , taking the continuous data as a test set X text of the improved GRU network, extracting the (L+1) th data of the rainfall site historical precipitation sequence as a tag Y test of X text , and taking the training set and the test set as normalized historical data.
  6. 6. The rainfall forecasting method based on the improved GRU multisource information fusion model of claim 5, wherein step S8 comprises the following steps: step S81, setting training iteration times of an improved GRU network as epochs, taking a mean square error mse as a loss function loss, and judging that training is finished when loss gradually decreases and tends to be stable in the iteration process; Step S82, training set And the initialized hidden layer state h 0 is input into the modified GRU network as a whole, wherein the process of input into the modified GRU network comprises the following steps: Step S821. The process of inputting to the improved GRU network comprises equations (10) through (17), wherein equations (10) through (17) comprise the following: (10); (11); (12); (13); (14); (15); (16); (17); In formulas (10) to (17), W z and b z represent the weight matrix and bias vector of the update gate, W r and b r represent the weight matrix and bias vector of the reset gate, W and b represent the weight matrix and bias vector of the candidate state, W o and b o represent the weight matrix bias vector of the output layer, h t-1 represents the network hidden state at time t-1, x t represents the input feature vector of the network at time t, σ () is a sigmoid function transform, and as indicated by the element-wise multiplication; Step S822. The process of inputting into the improved GRU network includes executing equations (10) through (17), wherein executing equations (10) through (17) includes the following: Firstly, according to formula (10), calculating the update gate to obtain output value z t of the update gate, secondly, according to formula (11), calculating the reset gate to obtain output value r t of the reset gate, and then according to formula (12), calculating the candidate hidden state Updating the candidate hidden state according to the formula (13), introducing an attention mechanism on the basis of the updated candidate hidden state, calculating an attention weight a t of each time step according to the formulas (14) and (15), carrying out weighted summation on the hidden state, obtaining a context vector c according to the formula (16), and finally obtaining a final output result y t through an output layer according to the context vector c of the formula (17).
  7. 7. The method for rainfall forecast based on the improved GRU multisource information fusion model of claim 6, wherein step S8 further comprises the following: Step S83, adopting a time-based back propagation algorithm, carrying out gradient calculation and optimization updating on all parameters in the improved GRU network by calculating an error between a model predicted output value Y * train and a real label Y train , and obtaining the trained improved GRU network through continuous iteration.
  8. 8. The rainfall forecasting method based on the improved GRU multisource information fusion model of claim 7, wherein step S8 comprises the following steps: Step S91, input data to be predicted is input into a trained feature weight module as an input matrix X per , and a final weight vector obtained by training is applied Performing feature scaling to obtain weighted real-time input matrix ; Then inputting the data into a trained improved GRU neural network model for forward propagation calculation to obtain a normalized output value And is opposite to Inverse normalization is carried out to obtain rainfall station fusion rainfall estimation values in a plurality of future times 。
  9. 9. A rainfall forecasting system based on an improved GRU multisource information fusion model, comprising an electronic device, wherein the electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements a rainfall forecasting method based on an improved GRU multisource information fusion model as claimed in any one of claims 1 to 8 when executing the computer program.
  10. 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, wherein the computer program when executed by a processor implements a rainfall forecasting method based on an improved GRU multi-source information fusion model as claimed in any one of claims 1 to 8.

Description

Rainfall forecasting method and system based on improved GRU multisource information fusion model Technical Field The invention provides a rainfall forecasting method and a rainfall forecasting system based on an improved GRU multisource information fusion model, and relates to the technical field of weather forecasting and hydrologic information. Background Precipitation is a core element in meteorological and hydrologic simulation, and the rainfall information with high precision and high space-time resolution has important significance for mountain torrent early warning in small watershed, accurate management of water resources and hydraulic engineering scheduling. Conventionally, ground rainfall stations are regarded as "true values" of precipitation monitoring, but in mountainous areas with complex terrain, the stations are sparsely distributed, and equipment faults or data loss are easy to occur in extreme weather. Although satellite remote sensing precipitation products (such as GPM and TRMM) can provide large-scale and high-frequency observation, the spatial resolution is coarse, significant errors exist in complex terrain areas, and the accurate capture requirement of small watershed on precipitation space variability is difficult to meet. In recent years, weather radar is a key data source for short-term rainfall prediction, and can provide reflectivity information with high space-time resolution. However, radar data is vulnerable to clutter, beam shielding and atmospheric interference, and traditional Z-I relationship conversion methods perform poorly in strongly varying precipitation scenarios. In addition, existing data fusion studies are focused on large-scale, long-time sequences (such as month and year scales), and studies aiming at small watershed, short-time, especially minute-scale (such as 5 minutes) forecast are relatively lacking. The reason for this is as follows: The fusion capability of multi-source heterogeneous data (radar, rainfall station and meteorological element) is insufficient, and the importance difference of different element characteristics cannot be fully considered; Most of the existing researches fail to fully consider the specificity of small mountainous areas, namely short confluence time and rapid response, and have extremely high requirements on timeliness and accuracy of forecasting. The forecasting scale of 1-3 hours is difficult to meet the actual demand of mountain small watershed flood early warning, and the 5-minute-level short-term forecasting with higher resolution faces the challenges of large data noise and strong time-space variability. The model is easy to fit on a limited sample, has poor generalization capability, and is difficult to adapt to the sudden and local characteristics of rainfall in mountainous areas. ‌ A Disclosure of Invention In view of the above, in order to make up for the blank and the deficiency of the prior art, the invention provides a rainfall forecasting method and a system based on an improved GRU multisource information fusion model, which are used for solving the problems of insufficient multisource heterogeneous data fusion, insufficient model generalization capability, failure to fully consider the particularity of short converging time and rapid response of a small drainage basin in the conventional rainfall forecasting in a small drainage basin in a mountain area, The invention provides a rainfall forecasting method and a rainfall forecasting system based on an improved GRU multisource information fusion model, wherein the rainfall forecasting method comprises the following steps: The invention provides a rainfall forecasting method based on an improved GRU multisource information fusion model, which is characterized by comprising the following steps of: s1, acquiring radar reflectivity factor data; s2, carrying out dynamic Z-I relation conversion on radar reflectivity factor data to generate a precipitation intensity sequence; step S3, multi-source information is obtained, wherein the multi-source information comprises a radar precipitation sequence, a rainfall station history precipitation sequence and an air temperature and humidity auxiliary sequence; S4, carrying out self-adaptive fusion on the multi-source information through a characteristic weight mechanism to construct a sample set; s5, dividing the sample set into a training set and a testing set; s6, carrying out data enhancement processing on the training sample set; s7, constructing an improved GRU neural network model; s8, training the improved GRU neural network model by using a training set, and testing by using a testing set to obtain a final trained model; and S9, inputting the input data to be predicted into the trained model to obtain a final rainfall forecast value. Further, step S2 includes the following: Step S21, taking a station to be predicted as a center, converting the acquired original radar reflectivity factor data expressed in decibel units