Short-time approaching rainfall prediction method based on multi-source fusion data and MIM network

Abstract

The invention discloses a short-time approaching rainfall prediction method based on multi-source fusion data and an MIM network, and the method comprises the steps: carrying out the space-time matching of radar observation data, Chinese land surface data assimilation system data and site observation data, carrying out the preprocessing of the matched data, including feature factor extraction, missing value processing and the like, building a deep learning model, and carrying out the prediction of the short-time approaching rainfall. The model is trained, and a multi-source fusion rainfall inversion result is output based on real-time radar data; furthermore, errors of a falling area and rainfall intensity of intelligent grid rainfall forecast are identified, a phase correction technology is utilized to correct a position error of an intelligent grid forecast rainfall zone, meanwhile, precision evaluation is performed on a prediction effect of the deep learning model according to the rainfall amount of short-time nowcasting, and intelligent prediction of short-time nowcasting is realized.

Description

Technical Field

[0001] The present invention provides a deep learning model for extrapolating and predicting short-term imminent precipitation based on multi-source fusion data, which relates to the fields of meteorological monitoring, multi-source data processing, computer network algorithm application, and belongs to the field of model algorithm development. Background Art

[0002] As global warming intensifies, the amount of water vapor in the atmosphere has significantly changed. Numerical simulations and satellite observations indicate that for every 1°C increase in Earth's surface temperature, the total atmospheric water content increases by approximately 7%. Simultaneously, the intensity and frequency of extreme precipitation have significantly increased, posing a serious threat to human life, property, and ecological stability. Precipitation forecasting, particularly short-term precipitation forecasting, plays a crucial role in meteorology. Accurate precipitation forecasts provide crucial information to governments and relevant departments, enabling them to promptly activate emergency response plans and organize evacuations, thereby minimizing casualties and property losses caused by precipitation-induced secondary disasters such as landslides, mudslides, and urban flooding. In construction, precipitation can cause waterlogging at construction sites, impacting quality. Accurate precipitation forecasts can help construction companies adjust their plans and schedule work processes in advance, avoiding weather-related delays. Precipitation forecasts can also provide early warning information, enabling them to implement safety measures and ensure the safety of construction workers.

[0003] In the research of precipitation forecasting technology, traditional numerical model research uses high-resolution numerical models such as CMA-MESO, CMA-GD and CMA-SH to forecast precipitation. These models can simulate atmospheric physical processes and predict the temporal and spatial distribution of precipitation. Due to the existence of model errors, including defects in physical parameterization schemes, calculation errors, defects in dynamic equations, etc., the results of numerical predictions have great uncertainty; while statistical forecasting research is based on historical precipitation data and related meteorological elements to establish statistical models to predict future precipitation conditions. With the in-depth study of the physical mechanisms affecting climate anomalies and the interactions between various ocean and atmospheric systems, the selection of factors with physical significance has become an important means of statistical prediction. Based on the analysis of the statistical and physical relationship between each factor and the predicted quantity, various mathematical and statistical methods are used to establish prediction models, which is still the main means of climate prediction.

[0004] With the development of technology in recent years, artificial intelligence forecasting methods based on deep learning, such as deep convolutional neural networks, have received increasing attention. Compared with traditional linear regression models, intelligent forecasting models based on deep learning often achieve better forecasting results, accuracy, and stability. Convolutional neural networks have powerful image recognition and nonlinear simulation capabilities, and are well-suited for predicting the occurrence of persistent and significant weather anomalies. Deep neural networks not only provide modeling for complex nonlinear systems, but also provide a higher level of abstraction, thereby improving the model's feature extraction capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to solve the problem of prediction accuracy. Therefore, the present invention provides a short-term precipitation prediction method based on multi-source fusion data and MIM network. By fusing and preprocessing multi-source data, a deep learning model is constructed using the MIM network. Site matching is performed based on the model prediction results and site observation data, the model's prediction accuracy is evaluated and the model prediction effect is optimized.

[0006] The technical solution of the present invention is specifically as follows:

[0007] A short-term precipitation prediction method based on multi-source fusion data and MIM network includes the following steps:

[0008] Step 1: Collect observation data from ground stations near the target area;

[0009] Step 2: Use the neighborhood method and bilinear interpolation method to preprocess the collected data. Based on the annual climate values ​​of the station, some missing values ​​in the station are supplemented. Then, the observation data are gridded using the bilinear interpolation method and the station data are interpolated onto the grid field.

[0010] Step 3: Use the preprocessed data to construct a precipitation observation dataset from input to output according to the modeling mapping method;

[0011] Step 4: Based on data updates, regularly acquire multi-source data through the data acquisition interface and process it to obtain product data with the same temporal and spatial resolution. Based on the acquired multi-source data, normalize the data by setting a threshold to normalize the data to between 0 and 1.

[0012] Step 5: Based on radar and satellite data, the MIM network is used to extract and map spatiotemporal features, a deep convolutional network is constructed to extract fusion features, and quantitative precipitation fusion inversion results are output;

[0013] Step 6: Take the test set data, perform inversion based on the model, and evaluate the model based on ground-based precipitation data to select the optimal model;

[0014] Step seven: the radar and satellite provided precipitation products are revised by the method of probability density function matching to generate the optimal multi-source fusion precipitation retrieval product.

[0015] In step one, the observation data includes precipitation, temperature, humidity, air pressure, wind direction and wind speed data parsed from the original message, and the data time interval is 1 minute.

[0016] In step two, the bilinear interpolation method is: in two directions, once linear interpolation is performed, and the value of the known function f at Q 11 =(x 1, y1), Q 12 =(x1,y2), Q 21 =(x2,y1) and Q 22 =(x2,y2) four points,

[0017] First, calculate the linear interpolation in the X direction as follows:

[0018]

[0019] The linear interpolation R1 coordinate is (x, y1), and the linear interpolation R2 coordinate is (x, y2), wherein x1

[0020] Then, linear interpolation is performed in the Y direction, that is, the interpolated grid element value f(P) is obtained:

[0021]

[0022] In step four, the formula for normalizing the data is

[0023]

[0024] Wherein, X' represents the value of the multi-source data, Min represents the minimum value of the data, and Max represents the maximum value of the data.

[0025] Step seven: when correcting the precipitation position of the intelligent grid forecast, the Fourier method and the objective diagnostic evaluation method are combined, the simultaneous short-term nowcasting precipitation forecast field and the intelligent grid forecast field are processed by fast Fourier transform, the spectral space deviation of the simultaneous precipitation observation and forecast is calculated, and the precipitation distribution of the numerical forecast is corrected according to the spectral space deviation.

[0026] The main process of correcting the precipitation position of the intelligent grid forecast in step seven is as follows:

[0027] 7.1. The overall displacement of the intelligent grid precipitation phase relative to the short-term nowcasting forecast is calculated by the fast Fourier transform method, and the precipitation position is corrected;

[0028] The fast Fourier transform method assumes that the numerical intelligent grid forecast precipitation field at time T is R f (x,y), and the short-term forecast precipitation field at the same time is R a (x,y); using discrete Fourier transform technology, R f (x,y) and R a (x,y) is converted to frequency domain space:

[0029]

[0030] Where u=0,1,2,…N-1, assuming R a (u,v) is determined only by R a (u, v) is obtained by simple translation, while ignoring the intensity change of the precipitation field, then according to the properties of Fourier transform, we can get:

[0031] R a (u,v)=R f (u,v)exp (-j(ux0+vy0)) (7)

[0032] Where R a (u,v) and R f (u,v) are R a (x,y) and R f The Fourier transform of (x,y), x0 and y0 are the required movement parameters, and their cross power spectrum is:

[0033]

[0034] In the above formula, R f (u,v) * R f The complex conjugate of (u, v) is used to inversely transform exp(-j(ux0+vy0)) to obtain the two-dimensional impulse function l(x-x0,y-y0). By finding the peak position of its phase correlation coefficient, the translation parameters x0 and y0 are finally determined;

[0035] 7.2. Identify target areas through objective diagnostic assessment methods and then characterize them, including the intensity and shape of both the short-term and impending precipitation forecast fields and the smart grid forecast fields.

[0036] The objective diagnosis and evaluation method includes two steps: target identification and target feature description and matching. First, a convolution operation is performed on the precipitation field predicted by the smart grid. The convolution function is used in the convolution process:

[0037]

[0038] Where f is the original precipitation field of the smart grid, is the filter function, the variables (x, y) and (u, v) are the grid coordinates, the filter function It is a circular filter determined by the influence radius R. Then, the precipitation threshold T is used to binarize the precipitation area. After the binarization, sporadic precipitation and weak precipitation in the precipitation field will be filtered out. The boundary of the "target" is outlined by the binarized coefficient field. These outlined areas are the "identification targets".

[0039] 7.3. After the short-term and impending precipitation forecast fields are identified, the attributes of the objects in the two fields are calculated, including precipitation center, area, and intensity. The objects in the two fields are then matched. After the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term and impending precipitation forecast using a dynamic weighting method.

[0040] The feature quantities used in matching include the center-of-gravity distance deviation of the two targets, the shortest distance between the target boundaries, the tilt angle deviation, the target overlap area ratio and the target area ratio; different weights are assigned to the above five components to obtain the value function I j ,

[0041]

[0042] C i is the confidence distribution of each feature quantity, reflecting the credibility of each feature quantity, F i,j is the membership function of each feature quantity, w i is the weight corresponding to each feature quantity. After each pair of targets is matched, the total value function I is obtained. j , I j The larger the value of , the greater the similarity between matching targets; when I j When the score is greater than 0.7, the pairing is considered successful, and the distance x between the pairing targets is calculated. i , and the total displacement of the precipitation field x0 is obtained

[0043]

[0044] Among them, w i The weight of each target area is equal to the ratio of the target area to the sum of the M target areas. The displacement is used to correct the smart grid precipitation field to obtain the final smart grid precipitation field.

[0045] The main technical method used in the correction of precipitation intensity is the Weibull distribution function method. The Weibull distribution function is:

[0046]

[0047] The probability density function is:

[0048]

[0049] Where α (> 0) is the shape parameter; β (> 0) is the scale parameter; and α0 is the location parameter. When α0 = 0, the solution to the above expression becomes a two-parameter Weibull distribution. Assuming that the short-term forecast precipitation field and the smart grid precipitation field follow the Weibull distribution, their distribution functions are obtained according to the above. Then, the smart grid precipitation field can be adjusted according to the radar estimation field. The specific relationship is:

[0050] QPF modified =CDF| QPE -1 CDF| model QPF| model (14)

[0051] Where, QPF modified is the adjusted precipitation distribution field, CDF| QPE and CDF| model are the Weibull probability distribution functions of the short-term forecast precipitation field and the smart grid forecast field respectively;

[0052] 7.4. After the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term precipitation forecast using a dynamic weighting method. During the fusion process, the weights of the predicted values ​​of the short-term precipitation forecast and the weight of the smart grid forecast need to be adjusted as they change over time. Within a shorter forecast period, the weight of the short-term precipitation forecast is greater. As the forecast period increases, the weight of the smart grid forecast results increases. The specific formula is:

[0053] R blending (t) = (1 - w(t)) * R Radar (t)+w(t) * R model (t) (15)

[0054] Where, t = 1, 2, ..., 6h, R blending (t) represents the fused rainfall forecast at time t, R Radar (t) represents the short-term forecast precipitation at time t, R model represents the forecast rainfall of the smart grid at time t, w(t) represents the weight coefficient of the smart grid, and the weight calculation adopts the empirical equation of the hyperbolic tangent function:

[0055]

[0056] Where α and β represent the endpoint weights of the smart grid at t = 1 and t = 6, respectively. To make the weight curve change smoothly, the value of γ is set to 1.

[0057] This paper proposes a short-term precipitation forecasting method based on multi-source fusion data and MIM (Memory in Memory) network. By fusing radar, CLDAS and station observation data, a multi-source input feature with spatiotemporal matching is constructed. The memory module of the MIM network is used to enhance the modeling capability of the nonlinear precipitation process, and the prediction model is dynamically evaluated and optimized, thereby improving the accuracy of short-term precipitation forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of the present invention;

[0059] Figure 2 is a bilinear difference map;

[0060] Figure 3 To output the training data set;

[0061] Figure 4 For ST-LSTM blocks and MIM blocks;

[0062] Figure 5 Graphs for non-stationary modules (MIM-N) and stationary modules (MIM-S);

[0063] Figure 6 The MIM network diagram has three MIMs and one ST-LSTM;

[0064] Figure 7 Revised processing flow chart for nowcasting. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the implementation described is only a part of the implementation of the present invention, not all embodiments. Based on the implementation of the present invention, all other implementations obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The present invention proposes a short-term near-fall precipitation prediction method based on multi-source fusion data and MIM network. This prediction algorithm can predict short-term near-fall precipitation and correct the near-fall prediction results based on blending extrapolation and numerical forecast fusion. The method first performs spatiotemporal matching of model data, CLDAS data, and station observation data. The matched data is then preprocessed, including feature factor extraction and missing value processing. A deep learning model is then constructed based on the extracted feature factors and CLDAS data, and the model is trained. The model's prediction results are then cleared and the final prediction results are output. Finally, station matching is performed based on the model prediction results and station observation data, the corresponding grid data is extracted, and the accuracy of the model is evaluated.

[0067] To achieve the above objectives, the specific steps of the present invention are as follows:

[0068] Step 1: Collect observation data from ground stations near the target area (including CLDAS data from national stations and radar network data from regional stations), including six elements of precipitation, temperature, humidity, air pressure, wind direction, and wind speed parsed from the original messages. For example: the time period is from 20:00 on December 31, 2012 to 20:00 on December 31, 2021 (the observation meteorological day is from 20:00 the previous day to 20:00 on the current day), with a total length of 9 years and an interval of 1 minute.

[0069] Step 2: Use the neighborhood method and bilinear interpolation method to preprocess the collected data. Based on the annual climate values ​​of the station, supplement some missing values ​​in the station (the time series before and after the missing values ​​and the data of nearby stations are relatively complete, and the anomalies in the data set are removed. The values ​​are less than or equal to 0 or exceed the historical extreme values). Then use the bilinear interpolation method to grid the observation data and interpolate the station data to the grid field to form the cumulative precipitation every 6 minutes. The core of the bilinear interpolation method is to perform a linear interpolation in each of the two directions. If we want to get the value of the unknown function f at point P = (x, y), assuming that we know the function f at Q 11 =(x 1, y1), Q 12 =(x1,y2),Q 21 =(x2,y1) and Q 22 =(x2,y2) the values ​​of the four points, as shown in the attached Figure 2 shown.

[0070] First calculate the linear interpolation in the X direction:

[0071]

[0072] The coordinates of linear interpolation R1 are (x, y1), and the coordinates of linear interpolation R2 are (x, y2), where x1<x<x2;

[0073] Then perform linear interpolation in the Y direction to obtain the interpolated grid feature value f(P):

[0074]

[0075] Step 3: After the pre-processing step, the data is mapped according to the modeling method to establish a precipitation observation data set from input to output "20 frames → 20 frames". The output training data set is as shown in the attached figure. Figure 3 shown.

[0076] Step 4: Based on data updates, periodically acquire radar, satellite, and automatic station data from multiple sources within two hours through the data acquisition interface. Use spatial range cropping and spatial resolution scaling to obtain product data with the same spatiotemporal resolution. Based on the acquired multi-source data, perform data normalization, using a threshold to normalize the data to a value between 0 and 1. Filling in missing values ​​makes the already high-precision observational data more complete in the time series, and the samples in the dataset are closer to actual observations. Furthermore, data normalization effectively reduces the range of precipitation variability during extreme precipitation events, allowing deep learning models to better capture the changing characteristics of precipitation during training. If the input data is too large for deep learning model inference, radar data can be compressed through downsampling to accommodate the precipitation prediction model.

[0077] The formula for normalizing the data is:

[0078]

[0079] Among them, x′ represents the value of multi-source data, Min represents the minimum value of the data, and Max represents the maximum value of the data.

[0080] Step 5: Based on radar and satellite data, the MIM network is used to extract and map spatiotemporal features, and a deep convolutional network is constructed to extract fusion features for precipitation extrapolation prediction, and output quantitative precipitation fusion inversion results. Natural spatiotemporal processes can be very unstable in many aspects, such as the accumulation, deformation or dissipation of radar echoes in precipitation forecasts. According to the decomposition of Cramer's rule, any non-stationary process can be decomposed into a deterministic time-varying polynomial plus a zero-mean random term. The MIM module utilizes the differential signal between adjacent recurring states and two cascaded, self-updating memory modules to simulate the non-stationary and approximately stationary characteristics in spatiotemporal dynamics. By stacking multiple MIM blocks, the model has the opportunity to capture higher-order non-stationarities, gradually making the spatiotemporal process stable and making future sequences more predictable. The structural diagram of the MIM network is shown in the attached figure. Figure 4 To the attached Figure 6 shown.

[0081] Figure 4 In the figure, the left image shows the ST-LSTM block, and the right image shows the proposed memory-in-memory (MIM) block. MIM aims to introduce two recurrent modules (yellow squares) to replace the forget gate (dashed box) in ST-LSTM. MIM-N is a non-stationary module, and MIM-S is a stationary module. Note that the MIM block cannot be used in the first layer because the input Xt is replaced by the Hl-1t block.

[0082] Figure 5In this paper, non-stationary modules (MIM-N) and stationary modules (MIM-S) are interconnected in a cascade structure in the MIM block, and non-stationarity is modeled by differences.

[0083] Step 6: Use the test set data to perform inversion based on the constructed model, and evaluate the model effect based on ground-based precipitation data to select the optimal model. Test it using traditional test indicators, which mainly include TS score, false alarm rate, missed alarm rate, and hit rate.

[0084] Step 7: When correcting the precipitation position of the smart grid forecast, a combination of the Fourier method and the objective diagnostic evaluation method is used. By performing fast Fourier transform processing on the short-term precipitation forecast field and the smart grid forecast field at the same moment, the spectral space deviation between the precipitation observation and forecast at the same moment is calculated. Based on the spectral space deviation, the numerically predicted precipitation distribution is corrected.

[0085] In step seven, the precipitation position predicted by the smart grid is corrected. The main process is as follows:

[0086] 7.1. Use the Fast Fourier Transform method to calculate the overall displacement of the smart grid precipitation level relative to the short-term nowcast and correct the precipitation position;

[0087] The fast Fourier transform method assumes that the numerical intelligent grid forecast precipitation field at time T is R f (x,y), and the short-term forecast precipitation field at the same time is R a (x,y). Using discrete Fourier transform technology, R f (x,y) and R a (x,y) is converted to frequency domain space.

[0088]

[0089] Where u=0,1,2,…N-1, assuming R a (u,v) is determined only by R a (u, v) is obtained by simple translation, while ignoring the intensity change of the precipitation field, then according to the properties of Fourier transform, we can get:

[0090] R a (u,v)=R f (u,v)exp (-j(ux0+vy0)) (7)

[0091] Where R a (u,v) and R f (u,v) are R a (x,y) and R fThe Fourier transform of (x,y), x0 and y0 are the required movement parameters. Their cross power spectrum is:

[0092]

[0093] In the above formula, R f (u,v) * R f The complex conjugate of (u, v). Performing the inverse transform of exp(-j(ux0+vy0)) yields the two-dimensional impulse function l(x-x0, y-y0). By finding the peak position of its phase correlation coefficient, the translation parameters x0 and y0 are ultimately determined.

[0094] 7.2. Identify target areas through objective diagnostic assessment methods and then characterize them, including the intensity and shape of both the short-term and impending precipitation forecast fields and the smart grid forecast fields.

[0095] The objective diagnosis and evaluation method includes two steps: target identification and target feature description and matching. First, a convolution operation is performed on the precipitation field predicted by the smart grid. The convolution function is used in the convolution process:

[0096]

[0097] Where f is the original precipitation field of the smart grid, is the filter function. The variables (x,y) and (u,v) are the grid coordinates. The filter function is a circular filter determined by the influence radius R. The purpose of convolution is to smooth the original precipitation field, making it more continuous and filtering out unpredictable smaller-scale systems. A precipitation threshold T is then used to binarize the precipitation area. This binarized precipitation field removes sporadic and weak precipitation. The binarized coefficient field can be used to delineate the boundaries of the "target," and these delineated areas are the "identified targets."

[0098] 7.3. After the identification of the short-term and impending precipitation forecast fields and the smart grid forecast precipitation fields is completed, the attributes of the targets in the two fields are calculated, including the precipitation center (the target's precipitation location), area, intensity, etc., and then the targets in the two fields are matched; after the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term and impending precipitation forecast using the dynamic weighting method.

[0099] The features used in the matching include the center of gravity distance deviation of the two targets, the shortest distance between the target boundaries, the tilt angle deviation, the target overlap area ratio and the target area ratio. Different weights are assigned to the above five components to obtain the value function I j .

[0100]

[0101] C i It is the confidence distribution of each feature quantity, reflecting the credibility of each feature quantity. i,j is the membership function of each feature quantity. i is the weight corresponding to each feature. After each pair of targets is matched, the total value function I j I j The larger the value of , the greater the similarity between matching targets. j When the score is greater than 0.7, the pairing is considered successful. i , we can get the total displacement x0 of the precipitation field

[0102]

[0103] where w i is the weight of each target area, which is equal to the ratio of the target area to the sum of the M target areas. The displacement is used to correct the smart grid precipitation field to obtain the final smart grid precipitation field.

[0104] The main technical method used in the correction of precipitation intensity is the Weibull distribution function method. The Weibull distribution function is:

[0105]

[0106] The probability density function is:

[0107]

[0108] In the formula, α (>0) is the shape parameter; β (>0) is the scale parameter; and α0 is the location parameter. When α0 = 0, the solution to the above expression becomes a two-parameter Weibull distribution. In this method, least squares estimation is used to obtain the parameters. Assuming that the short-term forecast precipitation field and the smart grid precipitation field follow a Weibull distribution, their distribution functions are obtained as described above. The smart grid precipitation field can then be adjusted based on the radar estimated field. The specific relationship is:

[0109] QPF modified =CDF| QPE -1 CDF| model QPF| model (14)

[0110] Where, QPF modified is the adjusted precipitation distribution field, CDF| QPE and CDF| model They are the Weibull probability distribution functions of the short-term forecast precipitation field and the smart grid forecast field, respectively.

[0111] 7.4. After the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term forecast precipitation using a dynamic weighting method. During the fusion process, the weight of the predicted value of the short-term forecast precipitation and the weight of the smart grid forecast need to be adjusted as they change over time. Within a shorter forecast period, the weight of the short-term forecast precipitation is greater. As the forecast period increases, the weight of the smart grid forecast results increases. The specific formula is:

[0112] R blending (t) = (1 - w(t)) * R Radar (t)+w(t) * R model (t) (15)

[0113] Where, t = 1, 2, ..., 6h, R blending (t) represents the fused rainfall forecast at time t, R Radar (t) represents the short-term forecast precipitation at time t, R model represents the forecast rainfall of the smart grid at time t, and w(t) represents the weight coefficient of the smart grid. The weight calculation adopts the empirical equation of the hyperbolic tangent function:

[0114]

[0115] Where α and β represent the endpoint weights of the smart grid at t = 1 and t = 6, respectively. To make the weight curve change smoothly, the value of γ is set to 1.

Claims

1. A short-term precipitation prediction method based on multi-source fusion data and MIM network, characterized by: The following steps are involved: Step 1: Collect observation data from ground stations near the target area; Step 2: Use the neighborhood method and bilinear interpolation method to preprocess the collected data. Based on the annual climate values ​​of the station, some missing values ​​in the station are supplemented. Then, the observation data are gridded using the bilinear interpolation method and the station data are interpolated onto the grid field. Step 3: Use the preprocessed data to construct a precipitation observation dataset from input to output according to the modeling mapping method; Step 4: Based on data updates, regularly acquire multi-source data through the data acquisition interface and process it to obtain product data with the same temporal and spatial resolution. Based on the acquired multi-source data, normalize the data by setting a threshold to normalize the data to between 0 and 1. Step 5: Based on radar and satellite data, the MIM network is used to extract and map spatiotemporal features, a deep convolutional network is constructed to extract fusion features, and quantitative precipitation fusion inversion results are output; Step 6: Take the test set data, perform inversion based on the model, and evaluate the model based on ground-based precipitation data to select the optimal model; Step 7: Use the probability density function matching method to correct the precipitation products provided by radar and satellite to generate the optimal multi-source fusion precipitation inversion product.

2. The method for short-term precipitation prediction based on multi-source fusion data and MIM network according to claim 1 is characterized by: In step 1, the observation data includes precipitation, temperature, humidity, air pressure, wind direction and wind speed data parsed from the original message, and the data interval is once every 1 minute.

3. The method for short-term precipitation prediction based on multi-source fusion data and MIM network according to claim 1 is characterized by: In step 2, the bilinear interpolation method is: perform a linear interpolation in each of the two directions. Given a function f in Q 11 =(x 1, y1), Q 12 =(x1,y2),Q 21 =(x2,y1) and Q 22 =(x2,y2) the values ​​of the four points, First calculate the linear interpolation in the X direction: The coordinates of linear interpolation R1 are (x, y1), and the coordinates of linear interpolation R2 are (x, y2), where x1<x<x2; Then perform linear interpolation in the Y direction to obtain the interpolated grid feature value f(P):

4. The method for short-term precipitation prediction based on multi-source fusion data and MIM network according to claim 1 is characterized by: In step 4, the formula for normalizing the data is: Among them, X′ represents the value of multi-source data, Min represents the minimum value of the data, and Max represents the maximum value of the data.

5. The method for short-term precipitation prediction based on multi-source fusion data and MIM network according to claim 1 is characterized in that: Step 7: When correcting the precipitation position of the smart grid forecast, a combination of the Fourier method and the objective diagnostic evaluation method is used. By performing fast Fourier transform processing on the short-term precipitation forecast field and the smart grid forecast field at the same moment, the spectral space deviation between the precipitation observation and forecast at the same moment is calculated. Based on the spectral space deviation, the numerically predicted precipitation distribution is corrected.

6. The method for short-term precipitation prediction based on multi-source fusion data and MIM network according to claim 5, characterized in that: In step seven, the precipitation position predicted by the smart grid is corrected. The main process is as follows: 7.

1. Use the Fast Fourier Transform method to calculate the overall displacement of the smart grid precipitation level relative to the short-term nowcast and correct the precipitation position; The fast Fourier transform method assumes that the numerical intelligent grid forecast precipitation field at time T is R f (x,y), and the short-term forecast precipitation field at the same time is R a (x,y); using discrete Fourier transform technology, R f (x,y) and R a (x,y) is converted to frequency domain space: Where u=0,1,2,…N-1, assuming R a (u,v) is determined only by R a (u, v) is obtained by simple translation, while ignoring the intensity change of the precipitation field, then according to the properties of Fourier transform, we can get: R a (u,v)=R f (u,v)exp (-j(ux0+vy0)) (7) Where R a (u,v) and R f (u,v) are R a (x,y) and R f The Fourier transform of (x,y), x0 and y0 are the required movement parameters, and their cross power spectrum is: In the above formula, R f (u,v) * R f The complex conjugate of (u, v) is used to inversely transform exp(-j(ux0+vy0)) to obtain the two-dimensional impulse function l(x-x0,y-y0). By finding the peak position of its phase correlation coefficient, the translation parameters x0 and y0 are finally determined; 7.

2. Identify target areas through objective diagnostic assessment methods and then characterize them, including the intensity and shape of both the short-term and impending precipitation forecast fields and the smart grid forecast fields. The objective diagnosis and evaluation method includes two steps: target identification and target feature description and matching. First, a convolution operation is performed on the precipitation field predicted by the smart grid. The convolution function is used in the convolution process: Where f is the original precipitation field of the smart grid, is the filter function, the variables (x, y) and (u, v) are the grid coordinates, the filter function It is a circular filter determined by the impact radius R. Then, the precipitation threshold T is used to binarize the precipitation area. After the binarization, sporadic precipitation and weak precipitation in the precipitation field will be filtered out. The boundaries of the "target" are outlined by the binarized coefficient field. These outlined areas are the "identified targets". 7.

3. After the short-term and impending precipitation forecast fields are identified, the attributes of the objects in the two fields are calculated, including precipitation center, area, and intensity. The objects in the two fields are then matched. After the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term and impending precipitation forecast using a dynamic weighting method. The feature quantities used in matching include the center-of-gravity distance deviation of the two targets, the shortest distance between the target boundaries, the tilt angle deviation, the target overlap area ratio and the target area ratio; different weights are assigned to the above five components to obtain the value function I j , C i is the confidence distribution of each feature quantity, reflecting the credibility of each feature quantity, F i,j is the membership function of each feature quantity, w i is the weight corresponding to each feature quantity. After each pair of targets is matched, the total value function I is obtained. j , I j The larger the value of , the greater the similarity between matching targets; when I j When the score is greater than 0.7, the pairing is considered successful, and the distance x between the pairing targets is calculated. i , and the total displacement of the precipitation field x0 is obtained in, w i The weight of each target area is equal to the ratio of the target area to the sum of the M target areas. The displacement is used to correct the smart grid precipitation field to obtain the final smart grid precipitation field. The main technical method used in the correction of precipitation intensity is the Weibull distribution function method. The Weibull distribution function is: The probability density function is: Where α (> 0) is the shape parameter; β (> 0) is the scale parameter; and α0 is the location parameter. When α0 = 0, the solution to the above expression becomes a two-parameter Weibull distribution. Assuming that the short-term forecast precipitation field and the smart grid precipitation field follow the Weibull distribution, their distribution functions are obtained according to the above. Then, the smart grid precipitation field can be adjusted according to the radar estimation field. The specific relationship is: QPF modified =CDF| QPE -1 CDF| model QPF| model (14) Where, QPF modified is the adjusted precipitation distribution field, CDF| QPE and CDF| model are the Weibull probability distribution functions of the short-term forecast precipitation field and the smart grid forecast field respectively; 7.

4. After the phase and intensity of the smart grid forecast field are corrected, it is fused with the short-term precipitation forecast using a dynamic weighting method. During the fusion process, the weights of the predicted values ​​of the short-term precipitation forecast and the weight of the smart grid forecast need to be adjusted as they change over time. Within a shorter forecast period, the weight of the short-term precipitation forecast is greater. As the forecast period increases, the weight of the smart grid forecast results increases. The specific formula is: R blending (t)=(1-w(t)) * R Radar (t)+w(t) * R model (t) (15) Where, t = 1, 2, ..., 6h, R blending (t) represents the fused rainfall forecast at time t, R Radar (t) represents the short-term forecast precipitation at time t, R model represents the forecast rainfall of the smart grid at time t, w(t) represents the weight coefficient of the smart grid, and the weight calculation adopts the empirical equation of the hyperbolic tangent function: Where α and β represent the endpoint weights of the smart grid at t = 1 and t = 6, respectively. To make the weight curve change smoothly, the value of γ is set to 1.

Images