A future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration

Abstract

The present invention relates to the technical field of future rainfall prediction, and in particular to a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration. Aiming at the problems that the current existing future rainfall prediction method still has insufficient accuracy, large amount of calculation, and is difficult to meet the timeliness and accuracy, the following scheme is proposed, which includes the following steps: S1: prediction preparation, S2: data analysis, S3: image analysis, S4: data processing, S5: prediction, S6: secondary prediction. The purpose of the present invention is to improve the accuracy of data by performing multi-granularity segmentation on all collected data according to pixels, and at the same time, compare the prediction results through secondary prediction, and finally improve the accuracy of prediction by re-determining the prediction.

Description

Technical Field

[0001] The invention relates to the technical field of future rainfall prediction, and in particular to a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration. Background Art

[0002] Traditional rainfall prediction methods can be roughly divided into two categories: one is based on numerical weather prediction methods, and the other is based on radar echo map extrapolation methods. The numerical weather prediction method solves the relevant equations of fluid mechanics and thermodynamics that describe the weather change process to predict the state of atmospheric movement in the future time period. This method is not accurate enough in real-time forecasting and has a large amount of calculation, making it difficult to meet the requirements of timeliness and accuracy. The radar echo map extrapolation method includes the following: the single centroid method, which performs least squares fitting on the centroid position of a thunderstorm echo at two adjacent moments and then extrapolates the prediction, but when the echo splits or merges, the tracking prediction accuracy is low; the second is the cross-correlation method, which analyzes and calculates the spatial correlation between two consecutive moments to make a short-term extrapolation prediction for the strong convective trajectory, but the prediction accuracy for thunderstorm weather with fast movement and change is relatively low; the third is the method based on optical flow, which is a real-time optical flow algorithm for the radar echo variation method, which uses the changes in pixels in the actual domain in the image sequence and the correlation between adjacent frames, combined with climate and other factors to complete the prediction. However, the optical flow estimation step and the radar echo extrapolation step are separate, and it is very difficult to determine the model parameters. In addition, as the spectral values ​​decrease and disappear quickly, the prediction of real-time rainfall in the region has certain limitations. In the existing short-term heavy rainfall prediction methods, radar echo extrapolation technology is usually used as the main technical means for near-term weather forecasting. Specifically, the echo intensity distribution and the moving speed and direction of the echo body (such as the precipitation area) are determined based on the echo data detected by the meteorological radar, and then the radar echo state after a certain period of time is predicted by linear or nonlinear extrapolation of the echo body. At the same time, the occurrence of climate events itself has a certain degree of randomness. The precipitation process is affected by regional atmospheric circulation, climate zones, tides and other climatic factors. Therefore, under the influence of these random factors, the distribution of short-term precipitation in the region presents a complex trend of change. It is necessary to integrate a large amount of meteorological sample data with multiple features to train the artificial intelligence recognition model. Compared with the traditional severe convective rainfall prediction and forecasting methods, the deep mining and analysis of time and space data are combined to improve the performance of the prediction model.

[0003] However, the existing future rainfall prediction methods still have the problems of insufficient precision, large amount of calculation, and difficulty in meeting timeliness and accuracy. Therefore, we propose a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to solve the problems that the existing future rainfall prediction methods still have insufficient accuracy, large amount of calculation, and difficulty in meeting the timeliness and accuracy requirements, and propose a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting future rainfall based on spatiotemporal bidirectional multi-granularity dynamic integration includes the following steps:

[0007] S1: Forecast preparation: Collect radar echo map sequences and actual rainfall values ​​of the target area at different times;

[0008] S2: Data analysis: Analyze the collected data and generate time data sets according to the length of time;

[0009] S3: Image analysis: perform multi-granularity segmentation on the acquired image and collect data for each sub-point in the target area;

[0010] S4: Data processing: Recombining time data and spatial data to generate functions, and using existing technologies to process the data;

[0011] S5: Make predictions: Make predictions from the generated functions, integrate the data, analyze and model the integrated data, and obtain prediction results;

[0012] S6: Second prediction: Get the data again and make a second prediction, and compare the prediction results;

[0013] Preferably, in S1, radar echo image sequences and actual rainfall values ​​of the target area at different collection times are collected in advance, and all collected radar echo image sequences and actual rainfall values ​​together form an actual rainfall data set of the target area, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0014] Preferably, in S2, the collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time, wherein the time data is segmented into multiple granularities according to the time unit, and the unit can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4), and (t+5). After acquiring the radar data, the data is segmented into multiple granularities according to one of the time units, and is segmented into multiple groups of data of the time unit, and the light vector corresponding to the current time t is The numerical value is processed to generate a function, and the time unit is pushed forward by one unit. The light vector value corresponding to time (t-1) generates a function again until the time is pushed forward to (t-5). The light vector value in the adjacent time unit is predicted by referring to the adjacent time unit and the corresponding light vector data. For example, the value of the light vector in time (t-1) can be used to predict the value of the light vector in time (t+1) with reference to the value of the light vector in time t. Similarly, the light vector value in another time can be predicted based on the values ​​of two adjacent times. When one of the time units is selected, a set of data sets will be generated. Different data sets can be generated according to different time units. All generated data sets are used as the total set of time data in the later stage.

[0015] Preferably, in S3, the collected image data is segmented at multiple granularities according to pixels, data is collected for each sub-point in the target area, and the spatial positions in the cloud map positions can be segmented in order from large to small and according to different spatial areas of the country, province, city, district, street, and community. When different spatial data sets are generated, each spatial position corresponds to a corresponding predicted value, and at the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location. The data is segmented in order from large to small according to the spatial position, the size of the ground radar position and the monitoring area, and the multi-granularity of the space, wherein the multi-granularity segmentation is performed according to the spatial position of the province, city, district, street, and community, and the smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function, and there are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit is switched to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a data set of the spatial level;

[0016] Preferably, in S4, the SVG algorithm is used to realize the recombination of the data set, wherein the data integration is to generate a new data set according to a new combination of time data and space data, and to cross-combine the time-divided data and the space-divided data again to generate a new data set, and to add the data value of the space unit to the data corresponding to different time units, and to use the combined data set to generate a large number of original data sets as the original data sets randomly sampled according to the destination position later;

[0017] Preferably, in said S5, a part of functions is selected from all generated functions according to the time and space of the predicted location, and the data is integrated. If it is impossible to select all the target data sets, a part of the function sets is selected, and the data integration and analysis are performed, wherein it is necessary to select multiple sets of reference data before predicting the target location, and the selected data sets are used as sampled data. The data situation of the target location is predicted based on the data of the data sets, and then the future rainfall situation is predicted based on the rainfall situation corresponding to the collected data values;

[0018] Preferably, in S6, the radar echo image sequence and the actual rainfall value of the target area at different times are reacquired, and after the data is transmitted back, professionals perform data and image analysis, and compare the data and analysis results with the data and analysis results transmitted back for the first time. If the data error is within a reasonable range, a secondary prediction is directly performed, and the prediction result is compared with the first prediction result. If the data error exceeds a reasonable error range, the error data is analyzed by professionals, the error data is corrected through analysis, and the corrected data is used as the final data for secondary prediction, and the prediction results are compared.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. By performing multi-granular segmentation on all collected data according to pixels, the accuracy of the data is improved, and ultimately the accuracy of the prediction is improved.

[0021] 2. Through secondary forecasting and comparing the forecast results, the forecast can be reconfirmed to improve the accuracy of the forecast.

[0022] The purpose of the present invention is to improve the accuracy of the data by performing multi-granularity segmentation on all collected data according to pixels, and at the same time, compare the prediction results through secondary prediction, and finally improve the accuracy of the prediction by re-determining the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration proposed by the present invention. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0025] Embodiment 1

[0026] Reference Figure 1 , a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration, including the following steps:

[0027] S1: Forecast preparation: pre-collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0028] S2: Data analysis: The collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time. When the time data is segmented into multiple granularities, the units can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5). After acquiring the radar data, the data is segmented into multiple groups of data according to one of the time units, and the light vector data corresponding to the current time t is combined into a new data set. The value is processed to generate a function, and the time unit is pushed forward by one unit. The light vector value corresponding to time (t-1) generates a function again until the time is pushed forward to (t-5). The light vector value in the adjacent time unit is predicted by referring to the adjacent time unit and the corresponding light vector data. For example, the value of the light vector in time (t-1) can be used to predict the value of the light vector in time (t+1) by referring to the value of the light vector in time t. Similarly, the light vector value in another time can be predicted based on the values ​​of two adjacent times. When one of the time units is selected, a set of data sets will be generated. Different data sets can be generated according to different time units. All generated data sets are used as the total set of time data in the later stage.

[0029] S3: Image analysis: The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map positions can be segmented in order from large to small, and according to different spatial areas such as country, province, city, district, street, and community. When generating different spatial data sets, each spatial position corresponds to a corresponding predicted value. At the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location. The images are segmented according to the order of spatial positions from large to small, the size of the ground radar position and the monitoring area, and the multi-granularity of the space. The multi-granularity segmentation is performed according to the spatial positions of provinces, cities, districts, streets, and communities. The smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function. There are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit switches to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a spatial level dataset;

[0030] S4: Data processing: SVG algorithm is used to realize the recombination of data sets, in which data integration is to generate a new data set according to a new combination of time data and space data, and to cross-combine the data after time segmentation and the data after space segmentation to generate a new data set, and add the data corresponding to different time units to the data value of the space unit, and use the combined data set to generate a large number of original data sets as the original data sets for random sampling according to the target location in the later stage;

[0031] S5: Make predictions: Select a part of the functions from all the generated functions according to the time and space of the prediction location, and integrate the data. If it is impossible to select all the target data sets, select a part of the function sets, and perform data integration and analysis. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampled data, predict the data situation of the target location according to the data of the data sets, and then predict the future rainfall situation according to the rainfall situation corresponding to the collected data values;

[0032] S6: Secondary prediction: Reacquire the radar echo map sequence and actual rainfall value of the target area at different times. After the data is transmitted back, professionals will conduct data and image analysis, and compare the data and analysis results with the first transmitted data and analysis results. If the data error is within a reasonable range, a secondary prediction will be directly carried out, and the prediction result will be compared with the first prediction result. If the data error exceeds the reasonable error range, professionals will analyze the error data, correct the error data through analysis, use the corrected data as the final data for secondary prediction, and compare the prediction results.

[0033] Embodiment 2

[0034] Reference Figure 1 , a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration, including the following steps:

[0035] S1: Forecast preparation: pre-collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0036] S2: Data analysis: The collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time. When the time data is segmented into multiple granularities, the units can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5). After acquiring the radar data, the data is segmented into multiple groups of data according to one of the time units, and the light vector value corresponding to the current time t is processed to generate a function;

[0037] S3: Image analysis: The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map positions can be segmented in order from large to small, and according to different spatial areas such as country, province, city, district, street, and community. When generating different spatial data sets, each spatial position corresponds to a corresponding predicted value. At the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location. The images are segmented according to the order of spatial positions from large to small, the size of the ground radar position and the monitoring area, and the multi-granularity of the space. The multi-granularity segmentation is performed according to the spatial positions of provinces, cities, districts, streets, and communities. The smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function. There are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit switches to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a spatial level dataset;

[0038] S4: Data processing: SVG algorithm is used to realize the recombination of data sets, in which data integration is to generate a new data set according to a new combination of time data and space data, and to cross-combine the data after time segmentation and the data after space segmentation to generate a new data set, and add the data corresponding to different time units to the data value of the space unit, and use the combined data set to generate a large number of original data sets as the original data sets for random sampling according to the target location in the later stage;

[0039] S5: Make predictions: Select a part of the functions from all the generated functions according to the time and space of the prediction location, and integrate the data. If it is impossible to select all the target data sets, select a part of the function sets, and perform data integration and analysis. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampled data, predict the data situation of the target location according to the data of the data sets, and then predict the future rainfall situation according to the rainfall situation corresponding to the collected data values;

[0040] S6: Secondary prediction: Re-acquire the radar echo map sequence and actual rainfall value of the target area at different times. After the data is transmitted back, professionals will perform data and image analysis, and compare the data and analysis results with the first transmitted data and analysis results. If the data error is within a reasonable range, a secondary prediction will be directly performed, and the prediction results will be compared with the first prediction results. If the data error exceeds the reasonable error range, professionals will analyze the error data, correct the error data through analysis, and use the corrected data as the final

[0041] The data are predicted twice and the prediction results are compared.

[0042] Embodiment 3

[0043] Reference Figure 1 , a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration, including the following steps:

[0044] S1: Forecast preparation: pre-collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0045] S2: Data analysis: The collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time. When the time data is segmented into multiple granularities, the units can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5). After acquiring the radar data, the data is segmented into multiple groups of data according to one of the time units, and the light vector data corresponding to the current time t is combined into a new data set. The value is processed to generate a function, and the time unit is pushed forward by one unit. The light vector value corresponding to time (t-1) generates a function again until the time is pushed forward to (t-5). The light vector value in the adjacent time unit is predicted by referring to the adjacent time unit and the corresponding light vector data. For example, the value of the light vector in time (t-1) can be used to predict the value of the light vector in time (t+1) by referring to the value of the light vector in time t. Similarly, the light vector value in another time can be predicted based on the values ​​of two adjacent times. When one of the time units is selected, a set of data sets will be generated. Different data sets can be generated according to different time units. All generated data sets are used as the total set of time data in the later stage.

[0046] S3: Image analysis: The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map can be segmented in descending order and according to different spatial regions such as country, province, city, district, street, and community. When different spatial data sets are generated, each spatial position corresponds to a corresponding predicted value. At the same time, the data is segmented at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location.

[0047] S4: Data processing: SVG algorithm is used to realize the recombination of data sets, in which data integration is to generate a new data set according to a new combination of time data and space data, and to cross-combine the data after time segmentation and the data after space segmentation to generate a new data set, and add the data corresponding to different time units to the data value of the space unit, and use the combined data set to generate a large number of original data sets as the original data sets for random sampling according to the target location in the later stage;

[0048] S5: Make predictions: Select a part of the functions from all the generated functions according to the time and space of the prediction location, and integrate the data. If it is impossible to select all the target data sets, select a part of the function sets, and perform data integration and analysis. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampled data, predict the data situation of the target location according to the data of the data sets, and then predict the future rainfall situation according to the rainfall situation corresponding to the collected data values;

[0049] S6: Secondary prediction: Reacquire the radar echo map sequence and actual rainfall value of the target area at different times. After the data is transmitted back, professionals will conduct data and image analysis, and compare the data and analysis results with the first transmitted data and analysis results. If the data error is within a reasonable range, a secondary prediction will be directly carried out, and the prediction result will be compared with the first prediction result. If the data error exceeds the reasonable error range, professionals will analyze the error data, correct the error data through analysis, use the corrected data as the final data for secondary prediction, and compare the prediction results.

[0050] Embodiment 4

[0051] Reference Figure 1 , a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration, including the following steps:

[0052] S1: Forecast preparation: pre-collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0053] S2: Data analysis: The collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time. When the time data is segmented into multiple granularities, the units can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5). After acquiring the radar data, the data is segmented into multiple groups of data according to one of the time units, and the light vector data corresponding to the current time t is combined into a new data set. The value is processed to generate a function, and the time unit is pushed forward by one unit. The light vector value corresponding to time (t-1) generates a function again until the time is pushed forward to (t-5). The light vector value in the adjacent time unit is predicted by referring to the adjacent time unit and the corresponding light vector data. For example, the value of the light vector in time (t-1) can be used to predict the value of the light vector in time (t+1) by referring to the value of the light vector in time t. Similarly, the light vector value in another time can be predicted based on the values ​​of two adjacent times. When one of the time units is selected, a set of data sets will be generated. Different data sets can be generated according to different time units. All generated data sets are used as the total set of time data in the later stage.

[0054] S3: Image analysis: The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map positions can be segmented in order from large to small, and according to different spatial areas such as country, province, city, district, street, and community. When generating different spatial data sets, each spatial position corresponds to a corresponding predicted value. At the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location. The images are segmented according to the order of spatial positions from large to small, the size of the ground radar position and the monitoring area, and the multi-granularity of the space. The multi-granularity segmentation is performed according to the spatial positions of provinces, cities, districts, streets, and communities. The smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function. There are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit switches to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a spatial level dataset;

[0055] S5: Make predictions: Select a part of the functions from all the generated functions according to the time and space of the prediction location, and integrate the data. If it is impossible to select all the target data sets, select a part of the function sets, and perform data integration and analysis. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampled data, predict the data situation of the target location according to the data of the data sets, and then predict the future rainfall situation according to the rainfall situation corresponding to the collected data values;

[0056] S6: Secondary prediction: Reacquire the radar echo map sequence and actual rainfall value of the target area at different times. After the data is transmitted back, professionals will conduct data and image analysis, and compare the data and analysis results with the first transmitted data and analysis results. If the data error is within a reasonable range, a secondary prediction will be directly carried out, and the prediction result will be compared with the first prediction result. If the data error exceeds the reasonable error range, professionals will analyze the error data, correct the error data through analysis, use the corrected data as the final data for secondary prediction, and compare the prediction results.

[0057] Embodiment 5

[0058] Reference Figure 1 , a future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration, including the following steps:

[0059] S1: Forecast preparation: pre-collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence;

[0060] S2: Data analysis: The collected data is segmented into multiple granularities according to the time level, and a time data set is generated according to the length of time. When the time data is segmented into multiple granularities, the units can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time, and a new data set is formed according to different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5). After acquiring the radar data, the data is segmented into multiple groups of data according to one of the time units, and the light vector data corresponding to the current time t is combined into a new data set. The value is processed to generate a function, and the time unit is pushed forward by one unit. The light vector value corresponding to time (t-1) generates a function again until the time is pushed forward to (t-5). The light vector value in the adjacent time unit is predicted by referring to the adjacent time unit and the corresponding light vector data. For example, the value of the light vector in time (t-1) can be used to predict the value of the light vector in time (t+1) by referring to the value of the light vector in time t. Similarly, the light vector value in another time can be predicted based on the values ​​of two adjacent times. When one of the time units is selected, a set of data sets will be generated. Different data sets can be generated according to different time units. All generated data sets are used as the total set of time data in the later stage.

[0061] S3: Image analysis: The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map positions can be segmented in order from large to small, and according to different spatial areas such as country, province, city, district, street, and community. When generating different spatial data sets, each spatial position corresponds to a corresponding predicted value. At the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multiple levels of spatial positions according to the spatial geographical location. The images are segmented according to the order of spatial positions from large to small, the size of the ground radar position and the monitoring area, and the multi-granularity of the space. The multi-granularity segmentation is performed according to the spatial positions of provinces, cities, districts, streets, and communities. The smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function. There are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit switches to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a spatial level dataset;

[0062] S4: Data processing: SVG algorithm is used to realize the recombination of data sets, in which data integration is to generate a new data set according to a new combination of time data and space data, and to cross-combine the data after time segmentation and the data after space segmentation to generate a new data set, and add the data corresponding to different time units to the data value of the space unit, and use the combined data set to generate a large number of original data sets as the original data sets for random sampling according to the target location in the later stage;

[0063] S5: Make predictions: Select a part of the functions from all the generated functions according to the time and space of the prediction location, and integrate the data. If it is impossible to select all the target data sets, select some function sets, and perform data integration and analysis. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampling data, predict the data situation of the destination location based on the data in the data sets, and then predict the future rainfall situation based on the rainfall situation corresponding to the collected data values.

[0064] The future rainfall prediction method based on spatiotemporal bidirectional multi-granularity dynamic integration in Embodiments 1, 2, 3, 4 and 5 was tested, and the results were as follows:

[0065]

[0066]

[0067] The future rainfall prediction methods based on spatiotemporal bidirectional multi-granularity dynamic integration obtained in Examples 1, 2, 3, 4 and 5 have significantly improved accuracy compared to existing methods, and Example 1 is the best example.

[0068] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting future rainfall based on spatiotemporal bidirectional multi-granularity dynamic integration, characterized in that: The following steps are involved: S1: Forecast preparation: Collect radar echo image sequences and actual rainfall values ​​of the target area at different times; collect radar echo image sequences and actual rainfall values ​​of the target area at different collection times in advance, and form an actual rainfall data set of the target area from all the collected radar echo image sequences and actual rainfall values, wherein in the actual rainfall data set of the target area, the radar echo image sequences and actual rainfall values ​​at the same collection time are in a one-to-one correspondence; S2: Data analysis: Analyze the collected data and generate a time data set according to the length of time; perform multi-granular segmentation on the collected data according to the time level and generate a time data set according to the length of time, wherein the time data can be segmented into 5 minutes, 10 minutes, 30 minutes, 60 minutes, 120 minutes, 240 minutes and 360 minutes according to the time when performing multi-granular segmentation, and combine different times (t-5), (t-4), (t-3), (t-2), (t-1), t, (t+1), (t+2), (t+3), (t+4) and (t+5) to form a new data set; after acquiring the radar data, perform multi-granular segmentation on the data according to one of the time units and divide it into multiple groups of data of the time unit, process the light vector value corresponding to the current time t to generate a function, and push the time unit forward by one unit, and generate a function again for the light vector value corresponding to the time (t-1) until the time is pushed forward to (t-5), and predict the light vector value in the adjacent time unit by referring to the adjacent time unit and the corresponding light vector data; S3: Image analysis: perform multi-granularity segmentation on the acquired image and collect data for each sub-point in the target area; The collected image data is segmented at multiple granularities according to pixels, and data is collected for each sub-point in the target area. The spatial positions in the cloud map positions can be segmented in order from large to small, and according to different spatial areas of the country, province, city, district, street, and community. When generating different spatial data sets, each spatial position corresponds to a corresponding predicted value. At the same time, the data is cut at multiple granularities according to the spatial position, and the image is divided into multi-level spatial positions according to the spatial geographical location. The data is segmented in order from large to small according to the spatial position, the size of the ground radar position and the monitoring area, and the multi-granularity of the space. The multi-granularity segmentation is performed according to the spatial position of the province, city, district, street, and community. The smallest spatial segmentation unit is adopted, and one of the spatial positions and its corresponding data value is selected to generate a function. There are multiple subset spatial positions under the same spatial level, and each spatial position coordinate corresponds to a value. When the spatial unit is switched to another spatial level, the level is composed of multiple spatial subsets, and the data composition corresponding to the space is changed to a spatial level dataset; S4: Data processing: Recombining time data and spatial data to generate functions, and using existing technologies to process the data; The SVG algorithm is used to realize the recombination of the data set, wherein the data combination is to generate a new data set according to a new combination of time data and space data, and to generate a new data set again according to the cross-combination of the time-divided data and the space-divided data, and to add the data corresponding to different time units to the data of the space unit, and to use the combined data set to generate a large number of original data sets as the original data sets for random sampling according to the target location later; S5: Make predictions: Make predictions from the generated functions, integrate the data, analyze and model the integrated data, and obtain prediction results; S6: Second prediction: Get the data again and make a second prediction, and compare the prediction results.

2. The method for predicting future rainfall based on spatiotemporal bidirectional multi-granularity dynamic integration according to claim 1 is characterized in that: In S5, a part of the functions is selected from all the generated functions according to the time and space of the predicted location, and the data is integrated. If it is impossible to select all the target number sets, a part of the function sets is selected, and the data is integrated and analyzed. Before predicting the target location, it is necessary to select multiple sets of reference data, use the selected data sets as sampled data, predict the data situation of the destination location based on the data in the data sets, and then predict the future rainfall situation based on the rainfall situation corresponding to the collected data values.

3. The method for predicting future rainfall based on spatiotemporal bidirectional multi-granularity dynamic integration according to claim 1 is characterized in that: In S6, the radar echo image sequence and the actual rainfall value of the target area at different times are reacquired, and after the data is transmitted back, professionals perform data and image analysis, and compare the data and analysis results with the data and analysis results transmitted back for the first time. If the data error is within a reasonable range, a secondary prediction is directly performed, and the prediction result is compared with the first prediction result. If the data error exceeds the reasonable error range, the error data is analyzed by professionals, the error data is corrected through analysis, and the corrected data is used as the final data for secondary prediction, and the prediction results are compared.

Images