A Short-Term Precipitation Forecasting Method Based on Multi-Source Fusion Data and MIM Network

Abstract

A short-term nowcasting precipitation prediction method based on multi-source fusion data and MIM networks is proposed. This method involves spatiotemporal matching of radar observation data with data from the China Land Surface Data Assimilation System and station observation data. The matched data undergoes preprocessing, including feature factor extraction and missing value handling, to construct a deep learning model. The model is trained and the multi-source fusion precipitation inversion results are output based on real-time radar data. Furthermore, the method identifies errors in the precipitation area and intensity of intelligent grid precipitation forecasts, uses "phase correction" technology to correct the positional errors of the rainbands in the intelligent grid forecasts, and evaluates the accuracy of the deep learning model's prediction performance based on the short-term nowcast precipitation amount. This achieves intelligent prediction of short-term nowcast precipitation.

Description

Technical Field

[0001] This invention provides a deep learning model for extrapolating and predicting short-term near-term precipitation based on multi-source fusion data. It involves meteorological monitoring, multi-source data processing, computer network algorithm applications, and belongs to the field of model algorithm development. Background Technology

[0002] With the intensification of global warming, the water vapor content in the atmosphere has changed significantly. Numerical simulations and satellite observations both 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 increased significantly, posing a serious threat to human life, property, and ecological stability. Precipitation forecasting plays a crucial role in meteorology, especially short-term, imminent precipitation forecasting. In disaster prevention, accurate precipitation forecasts provide governments and relevant departments with critical information to promptly activate emergency plans, organize evacuations, and reduce casualties and property losses caused by secondary disasters such as landslides, mudslides, and urban flooding triggered by precipitation. In construction, precipitation can lead to water accumulation at construction sites, affecting construction quality. Accurate precipitation forecasts can help construction companies adjust their work plans in advance, rationally arrange work processes, and avoid project delays due to weather conditions. Precipitation forecasts can also provide early warning information, enabling construction companies to take safety measures in advance and ensure the personal safety of construction workers.

[0003] In precipitation forecasting research, traditional numerical modeling studies utilize high-resolution numerical models such as CMA-MESO, CMA-GD, and CMA-SH to predict precipitation. These models can simulate atmospheric physical processes and predict the spatiotemporal distribution of precipitation. However, due to model errors, including defects in physical parameterization schemes, computational errors, and flaws in dynamic equations, the results of numerical predictions have significant uncertainties. Statistical forecasting, on the other hand, is based on historical precipitation data and relevant meteorological elements to establish statistical models that predict future precipitation. With in-depth research into the physical mechanisms influencing climate anomalies and the interactions between various oceanic and atmospheric systems, selecting physically significant factors has become an important means of statistical forecasting. Based on the analysis of the statistical and physical relationships between various factors and predicted quantities, various mathematical and statistical methods are used to establish forecasting models, which remains the main means of climate prediction to this day.

[0004] In recent years, with the development of technology, artificial intelligence forecasting methods based on deep learning, relying on models 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 possess powerful image recognition and nonlinear simulation capabilities, demonstrating good forecasting ability for persistent and significant weather anomalies. Deep neural networks can not only model complex nonlinear systems but also provide models with higher levels of abstraction, thereby improving the model's feature extraction capabilities. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to improve the accuracy of prediction. Therefore, this invention provides a short-term now-near 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 MIM network. Based on the model prediction results and station observation data, station matching is performed to evaluate the prediction accuracy of the model and optimize the prediction effect of the model.

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

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

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

[0009] Step 2: Use the neighborhood method and bilinear interpolation to preprocess the collected data. Based on the annual climate values ​​of the stations, supplement the missing values ​​of the stations. Then use the bilinear interpolation method to grid the observation data and interpolate the station data onto the grid field.

[0010] Step 3: Based on the mapping method of the modeling, construct the precipitation observation dataset from input to output using the data after the preprocessing steps;

[0011] Step 4: Based on the data update status, periodically acquire multi-source data through the data acquisition interface and process it to obtain product data with the same spatiotemporal 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, use the MIM network to extract and map spatiotemporal features, construct a deep convolutional network to extract fusion features, and output quantitative precipitation fusion inversion results;

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

[0014] Step 7: Correct the precipitation products provided by radar and satellite using the probability density function matching method to generate the optimal multi-source fused precipitation inversion product.

[0015] In step one, the observation data includes precipitation, temperature, humidity, air pressure, wind direction and wind speed data parsed from the original report, with a data interval of 1 minute.

[0016] In step two, the bilinear interpolation method involves performing linear interpolation once in each of the two directions, given the function f in Q. 11 =(x 1, y1) Q 12 = (x1, y2), Q 21 = (x2, y1) and Q 22 = the values ​​of the four points (x2, y2),

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

[0018]

[0019] The coordinates of linear interpolation R1 are (x, y1), and the coordinates of linear interpolation R2 are (x, y2), where x1 < x < x2;

[0020] Then, linear interpolation is performed in the Y direction to obtain the interpolated grid point feature value f(P):

[0021]

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

[0023]

[0024] Where 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 7: When correcting the precipitation location in the smart grid forecast, a combination of Fourier method and objective diagnostic evaluation method is used. By performing fast Fourier transform on the short-term now-near-precipitation forecast field and the smart grid forecast field at the same moment, the spectral spatial deviation of precipitation observation and forecast at the same moment is calculated. Based on the spectral spatial deviation, the precipitation distribution in the numerical forecast is corrected.

[0026] Step seven involves correcting the precipitation locations predicted by the smart grid. The main process is as follows:

[0027] 7.1. The overall displacement of precipitation level relative to short-term nowcast is obtained by using the fast Fourier transform method, and the precipitation position is corrected accordingly.

[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 moment is R. a (x,y); Using the Discrete Fourier Transform technique, R f (x,y) and R a (x,y) transformed to the frequency domain:

[0029]

[0030] Where u = 0, 1, 2, ..., N-1, and assume R... a (u,v) is determined only by R a (u,v) is obtained through a simple translation, ignoring the intensity variation of the precipitation field. Then, according to the properties of the Fourier transform, we can obtain:

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

[0032] In the formula R a (u,v) and R f (u,v) represent R respectively a (x,y) and R f The Fourier transform of (x,y), where x0 and y0 are the shift parameters to be determined, and their cross-power spectrum is:

[0033]

[0034] In the above formula, R f (u,v) * For R f The complex conjugate of (u,v) is obtained by inversely transforming exp(-j(ux0+vy0)) to obtain the two-dimensional pulse 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 the target area through objective diagnostic assessment methods, and then describe the characteristics of the target area, including the intensity and shape of the short-term precipitation field and the smart grid forecast field;

[0036] The objective diagnostic assessment method includes two steps: target identification and target feature description and matching. First, a convolution operation is performed on the intelligent grid-predicted precipitation field, using a convolution function.

[0037]

[0038] In the formula, f is the original precipitation field of the smart grid. It is a filtering function, where variables (x,y) and (u,v) are grid coordinates. It is a circular filter determined by the radius R; then, the precipitation threshold T is used to binarize the precipitation area. After binarization, sporadic precipitation and weak precipitation in the precipitation field will be filtered out. The boundary of the "target" is delineated by the coefficient field of binarization. These delineated areas are the "identified target".

[0039] 7.3. After identifying the short-term precipitation forecast field and the smart grid precipitation forecast field, the attributes of the targets in the two fields are calculated, including the precipitation center, area and intensity. Then, the targets in the two fields are matched. After correcting the phase and intensity of the smart grid forecast field, the dynamic weighting method is used to fuse it with the short-term precipitation forecast.

[0040] The features used in the matching process include the centroid distance deviation between the two targets, the shortest distance between the target boundaries, the tilt angle deviation, the ratio of the overlapping areas of the targets, and the ratio of the target areas. Different weights are assigned to these five components to obtain the value function I. j ,

[0041]

[0042] C i It represents the confidence distribution of each feature, reflecting the confidence level of each feature. F i,j It is the membership function of each feature, w i These are the weights corresponding to each feature, and the total value function I is obtained after each target match. j I j The larger the value of I, the greater the similarity between the matched targets; when I j A score greater than 0.7 indicates a successful pairing, calculated by determining the distance x between the paired targets. i The total displacement x0 of the precipitation field is obtained.

[0043]

[0044] Among them, w i The weight of each target region is equal to the ratio of the target area to the sum of the areas of M targets. The intelligent grid precipitation field is corrected by the displacement to obtain the final intelligent grid precipitation field.

[0045] The main technique used for correcting precipitation intensity is the Weibull distribution function method. The Weibull distribution function is as follows:

[0046]

[0047] The probability density function is:

[0048]

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

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

[0051] In the formula, QPF modified For the adjusted precipitation distribution field, CDF| QPE and CDF| model These are the Weibull probability distribution functions for the short-term precipitation forecast field and the smart grid forecast field, respectively;

[0052] 7.4 After correcting the phase and intensity of the smart grid forecast field, a dynamic weighting method is used to fuse it with the short-term precipitation forecast. During the fusion process, the weights of the short-term precipitation forecast and the smart grid forecast need to be adjusted over time. Within a shorter forecast period, the short-term precipitation forecast has a larger weight; as the forecast period lengthens, the weight of the smart grid forecast increases. The specific formula is as follows:

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

[0054] In the formula, 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 Let w(t) represent the rainfall forecast by the smart grid at time t, and w(t) represent the weighting coefficients of the smart grid. The weights are calculated using the empirical equation of the hyperbolic tangent function:

[0055]

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

[0057] This invention proposes a short-term now-near precipitation prediction method based on multi-source fusion data and MIM (Memory in Memory) network. By fusing radar, CLDAS and station observation data, a spatiotemporally matched multi-source input feature is constructed. The memory module of the MIM network is used to enhance the modeling capability of nonlinear precipitation processes, and the prediction model is dynamically evaluated and optimized, thereby improving the accuracy of short-term now-near precipitation forecast. Attached Figure Description

[0058] Figure 1 This is a flowchart of the present invention;

[0059] Figure 2 This is a bilinear interpolation plot;

[0060] Figure 3 To output the training dataset;

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

[0062] Figure 5 The diagram shows the non-stationary module (MIM-N) and the stationary module (MIM-S).

[0063] Figure 6 A diagram of a MIM network with three MIMs and one ST-LSTM;

[0064] Figure 7 Flowchart for nowcast forecast correction. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

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

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

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

[0069] Step Two: Preprocess the collected data using the neighborhood method and bilinear interpolation. Based on the station's annual climate values, supplement the missing values ​​at the station (time series before and after the missing values ​​and data from nearby stations are relatively complete, removing outliers less than or equal to 0 or exceeding historical extremes). Then, use bilinear interpolation to grid the observed data, interpolating the station data onto the gridded field to form cumulative precipitation every 6 minutes. The core of bilinear interpolation is performing linear interpolation in two directions. Suppose we want to obtain the value of the unknown function f at point P = (x, y), and we know the value of function f at point Q... 11 =(x 1, y1) Q 12 = (x1, y2), Q 21 = (x2, y1) and Q 22 The values ​​of the four points (x2, y2) are shown in the attached figure. Figure 2 As shown.

[0070] First, calculate the linear interpolation value 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, linear interpolation is performed in the Y direction to obtain the interpolated grid point feature value f(P):

[0074]

[0075] Step 3: Based on the modeling mapping method, the preprocessed data is used to establish a precipitation observation dataset with an input-to-output "20 frames → 20 frames" model. The output training dataset is shown in the attached figure. Figure 3 As shown.

[0076] Step 4: Based on data updates, periodically acquire multi-source data (radar, satellite, automatic weather station, etc.) from the same time period within two hours via the data acquisition interface. Use spatial range cropping and spatial resolution scaling to obtain product data with the same spatiotemporal resolution. Perform data normalization on the acquired multi-source data, normalizing it to between 0 and 1 using a threshold setting. Supplementing missing data makes the originally high-precision observation data more complete in the time series, and the samples in the dataset are closer to the actual observation values. Simultaneously, data normalization effectively reduces the variability range of precipitation in extreme precipitation events, allowing deep learning models to better capture precipitation variation characteristics during training. When performing deep learning model inference, if the input data is too large, the model will not run. Data compression of radar data can be achieved through downsampling to adapt to the precipitation prediction model.

[0077] The formula for normalizing data is:

[0078]

[0079] Where 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.

[0080] Step 5: Based on radar and satellite data, a MIM network is used for spatiotemporal feature extraction and mapping. A deep convolutional network is constructed to extract and fuse features for precipitation extrapolation and prediction, and quantitative precipitation fusion inversion results are output. Natural spatiotemporal processes can be highly unstable in many ways, such as the accumulation of radar echoes, deformation, or dissipation at higher levels in precipitation forecasting. According to Cramer's law, any non-stationary process can be decomposed into a deterministic time-varying polynomial plus a zero-mean random term. The MIM module utilizes the difference signal between adjacent recurring states and two cascaded, self-updating memory modules to simulate the non-stationary and near-stationary characteristics in spatiotemporal dynamics. By stacking multiple MIM blocks, the model has the opportunity to capture higher-order non-stationarity, gradually stabilizing the spatiotemporal process and making future sequences more predictable. The structure diagram of the MIM network is attached. Figure 4 To be continued Figure 6 As shown.

[0081] Figure 4 In the diagram, the left image shows an ST-LSTM block, and the right image shows a proposed Memory-In-Memory (MIM) block. MIM aims to introduce two cyclic 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 the MIM block, non-stationary modules (MIM-N) and stationary modules (MIM-S) are interconnected in a cascaded structure, and non-stationarity is modeled by difference.

[0083] Step 6: Using the test set data, perform inversion based on the constructed model, and evaluate the model's performance based on ground-observed precipitation data. Select the optimal model and verify it using traditional verification indicators, which mainly include TS score, false alarm rate, missed alarm rate, and hit rate.

[0084] Step 7: When correcting the precipitation location in the smart grid forecast, a combination of Fourier method and objective diagnostic evaluation method is used. By performing fast Fourier transform on the short-term now-near-precipitation forecast field and the smart grid forecast field at the same moment, the spectral spatial deviation of precipitation observation and forecast at the same moment is calculated. Based on the spectral spatial deviation, the precipitation distribution in the numerical forecast is corrected.

[0085] Step seven involves correcting the precipitation locations predicted by the smart grid. The main process is as follows:

[0086] 7.1. The overall displacement of precipitation level relative to short-term nowcast is obtained by using the fast Fourier transform method, and the precipitation position is corrected accordingly.

[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 moment is R. a (x,y). Using the Discrete Fourier Transform technique, R... f (x,y) and R a (x,y) is transformed into the frequency domain.

[0088]

[0089] Where u = 0, 1, 2, ..., N-1, and assume R... a (u,v) is determined only by R a (u,v) is obtained through a simple translation, ignoring the intensity variation of the precipitation field. Then, according to the properties of the Fourier transform, we can obtain:

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

[0091] In the formula R a (u,v) and R f (u,v) represent R respectively a (x,y) and R fThe Fourier transform of (x,y) is given, where x0 and y0 are the shift parameters to be determined. Their cross-power spectrum is:

[0092]

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

[0094] 7.2. Identify the target area through objective diagnostic assessment methods, and then describe the characteristics of the target area, including the intensity and shape of the short-term precipitation field and the smart grid forecast field;

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

[0096]

[0097] In the formula, f is the original precipitation field of the smart grid. This is a filtering function. The variables (x, y) and (u, v) are grid point coordinates. Filtering function It is a circular filter determined by the radius R. The purpose of convolution is to smooth the original precipitation field, making it more continuous and filtering out some small-scale systems that are not predictable. Then, a precipitation threshold T is used to binarize the precipitation area. After binarization, sporadic and weak precipitation in the precipitation field will be filtered out. The boundary of the "target" can be delineated through the binarized coefficient field. These delineated areas are the "identified targets".

[0098] 7.3. After identifying the short-term precipitation forecast field and the smart grid precipitation forecast field, the attributes of the targets in the two fields are calculated, including the precipitation center (the precipitation location of the target), area, intensity, etc. Then, the targets in the two fields are matched. After correcting the phase and intensity of the smart grid forecast field, the dynamic weighting method is used to fuse it with the short-term precipitation forecast.

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

[0100]

[0101] C i This represents the confidence distribution of each feature, reflecting the reliability of each feature. F i,j It is the membership function for each feature. i These are the weights corresponding to each feature. After each target match, the total value function I is obtained. j I j The larger the value of I, the greater the similarity between the matched targets. j A score greater than 0.7 indicates a successful pairing. The distance x between the paired targets is calculated. i The total displacement x0 of the precipitation field can be obtained.

[0102]

[0103] Where w i The weight of each target region is equal to the ratio of the target area to the sum of the areas of M target regions. The final smart grid precipitation field is obtained by correcting the displacement.

[0104] The main technique used for correcting precipitation intensity is the Weibull distribution function method. The Weibull distribution function is as follows:

[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. Then, the smart grid precipitation field can be adjusted based on the radar-estimated field, specifically:

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

[0110] In the formula, QPF modified For the adjusted precipitation distribution field, CDF| QPE and CDF| model These are the Weibull probability distribution functions for the short-term precipitation forecast field and the smart grid forecast field, respectively.

[0111] 7.4 After correcting the phase and intensity of the smart grid forecast field, a dynamic weighting method is used to fuse it with the short-term precipitation forecast. During the fusion process, the weights of the short-term precipitation forecast and the smart grid forecast need to be adjusted over time. Within a shorter forecast period, the short-term precipitation forecast has a larger weight; as the forecast period lengthens, the weight of the smart grid forecast increases. The specific formula is as follows:

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

[0113] In the formula, 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 Let w(t) represent the rainfall forecast by the smart grid at time t, and w(t) represent the weighting coefficients of the smart grid. The weights are calculated using the empirical equation of the hyperbolic tangent function:

[0114]

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

Claims

1. A short-term nowcasting precipitation prediction method based on multi-source fusion data and MIM network, characterized in that, Includes the following steps: Step 1: Collect observation data from nearby ground stations in the target area; Step 2: Use the neighborhood method and bilinear interpolation to preprocess the collected data. Based on the annual climate values ​​of the stations, supplement the missing values ​​of the stations. Then use the bilinear interpolation method to grid the observation data and interpolate the station data onto the grid field. Step 3: Based on the mapping method of the modeling, construct the precipitation observation dataset from input to output using the data after the preprocessing steps; Step 4: Based on the data update status, periodically acquire multi-source data through the data acquisition interface and process it to obtain product data with the same spatiotemporal 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, use the MIM network to extract and map spatiotemporal features, construct a deep convolutional network to extract fusion features, and output quantitative precipitation fusion inversion results; Step 6: Take test set data, perform inversion based on the model, and evaluate the model based on ground-observed precipitation data to select the optimal model; Step 7: Correct the precipitation products provided by radar and satellite using the probability density function matching method to generate the optimal multi-source fused precipitation inversion product; In step seven, when correcting the precipitation location in the smart grid forecast, a combination of Fourier method and objective diagnostic evaluation method is used. By performing fast Fourier transform on the short-term now-near precipitation forecast field and the smart grid forecast field at the same moment, the spectral space deviation of precipitation observation and forecast at the same moment is calculated. Based on the spectral space deviation, the precipitation distribution in the numerical forecast is corrected. Step seven involves correcting the precipitation locations predicted by the smart grid. The main process is as follows: 7.

1. The overall displacement of precipitation level relative to short-term nowcast is obtained by using the fast Fourier transform method, and the precipitation position is corrected accordingly. 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 moment is R. a (x,y); Using the Discrete Fourier Transform technique, R f (x,y) and R a (x,y) transformed to the frequency domain: ； ； Where u = 0, 1, 2, ..., N-1, and assume R... a (u,v) is determined only by R f (u,v) is obtained through a simple translation, ignoring the intensity variation of the precipitation field. Then, according to the properties of the Fourier transform, we can obtain: ； In the formula R a (u,v) and R f (u,v) represent R respectively a (x,y) and R f The Fourier transform of (x,y) is given, where x0 is the translation parameter of the precipitation field in the x-direction and y0 is the translation parameter of the precipitation field in the y-direction. Both are unknowns to be solved, and their cross-power spectrum phase factor is: ； In the above formula, R f (u,v) * For R f The complex conjugate of (u,v) is obtained by inversely transforming exp(−j(ux0+vy0)) to obtain the two-dimensional impulse function l(x−x0,y−y0). The translation parameters can be determined by finding the peak position of its phase correlation coefficient. 7.

2. Identify the target area using objective diagnostic assessment methods, and then describe the characteristics of the target area, including the intensity and shape of the short-term precipitation field and the smart grid forecast field; The objective diagnostic assessment method includes two steps: target identification and target feature description and matching. First, a convolution operation is performed on the intelligent grid-predicted precipitation field, using a convolution function. ； In the formula, f is the original precipitation field of the smart grid, φ is the filtering function, and the variables (x,y) and (u,v) are the grid coordinates. The filtering function φ is a circular filter determined by the radius of influence R. Then, the precipitation threshold T is used to binarize the precipitation area. After binarization, sporadic precipitation and weak precipitation in the precipitation field will be filtered out. The boundary of the "target" is delineated by the coefficient field of the binarization. These delineated areas are the "identified targets". 7.

3. After identifying the short-term precipitation forecast field and the smart grid precipitation forecast field, the attributes of the targets in the two fields are calculated, including the precipitation center, area and intensity. Then, the targets in the two fields are matched. After correcting the phase and intensity of the smart grid forecast field, the dynamic weighting method is used to fuse it with the short-term precipitation forecast. The features used in the matching process include the centroid distance deviation between the two targets, the shortest distance between the target boundaries, the tilt angle deviation, the ratio of the overlapping areas of the targets, and the ratio of the target areas. Different weights are assigned to these five components to obtain the value function I. j , ； C i It represents the confidence distribution of each feature, reflecting the confidence level of each feature. F i,j It is the membership function of each feature, w i These are the weights corresponding to each feature, and the total value function I is obtained after each target match. j I j The larger the value of I, the greater the similarity between the matched targets; when I j A score greater than 0.7 indicates a successful pairing. This is determined by calculating the x-direction displacement deviation d of the i-th successfully paired target. i The total displacement of the precipitation field in the x-direction is obtained by weighting by area. ； Among them, w i The weight of each target region is equal to the ratio of the target area to the sum of the areas of M targets. The intelligent grid precipitation field is corrected by the displacement to obtain the final intelligent grid precipitation field. The technique used to correct precipitation intensity is the Weibull distribution function method. The Weibull distribution function is: ； The probability density function is: ； In the formula, α is the shape parameter, α > 0; β is the scale parameter, β > 0; and is the location parameter. When , the solution of the above expression becomes a two-parameter Weibull distribution. Assuming that the short-term forecast precipitation field and the smart grid precipitation field follow a Weibull distribution, their distribution functions are obtained according to the above. Then, the smart grid precipitation field can be adjusted according to the radar estimated field. The specific relationship is: ； In the formula, QPF modified For the adjusted precipitation distribution field, CDF| QPE and CDF| model These are the Weibull probability distribution functions for the short-term precipitation forecast field and the smart grid forecast field, respectively; 7.4 After correcting the phase and intensity of the smart grid forecast field, a dynamic weighting method is used to fuse it with the short-term precipitation forecast. During the fusion process, the weights of the short-term precipitation forecast and the smart grid forecast need to be adjusted over time. Within a shorter forecast period, the short-term precipitation forecast has a larger weight; as the forecast period lengthens, the weight of the smart grid forecast increases. The specific formula is as follows: ； In the formula, 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 Let w(t) represent the rainfall forecast by the smart grid at time t, and w(t) represent the weighting coefficients of the smart grid. The weights are calculated using the empirical equation of the hyperbolic tangent function: ； Where w1 and w6 represent the endpoint weights of the smart grid at times t=1 and t=6, respectively, and are set to 1 to make the weight curve change smoothly.

2. The short-term nowcasting precipitation prediction method based on multi-source fusion data and MIM network according to claim 1, characterized in that: In step one, the observation data includes precipitation, temperature, humidity, air pressure, wind direction and wind speed data parsed from the original report, with a data interval of 1 minute.

3. The short-term nowcasting precipitation prediction method based on multi-source fusion data and MIM network according to claim 1, characterized in that: In step two, the bilinear interpolation method involves performing linear interpolation once in each of the two directions, given the function f in Q. 11 = (x 1, y1) Q 12 = (x1, y2), Q 21 = (x2, y1) and Q 22 = the values ​​of the four points (x2, y2), First, calculate the linear interpolation value 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 point feature values. f (P): 。 4. The short-term nowcasting precipitation prediction method based on multi-source fusion data and MIM network according to claim 1, characterized in that: In step four, the formula for normalizing the data is: ； in, X ´ indicates the value of multi-source data. Min This represents the minimum value of the data. Max This indicates the maximum value of the data.

Images