Meteorological prediction method and system based on deep learning

Abstract

The invention relates to the technical field of weather prediction, in particular to a weather prediction method based on deep learning, which comprises the following steps: determining the category of a sub-region according to the number of current-domain rainfall stations and the number of neighborhood rainfall stations corresponding to the sub-region; determining to carry out state matching division or adjustment processing according to the category quantity ratio and the regional distribution uniformity; determining a target optimal area based on the number of drainage points of the adjacent areas and the terrain reference height or based on the fusion times and the terrain fusion difference value; determining a task category according to an inversion factor fluctuation ratio and confirming a distribution sequence according to a priority; and determining or processing the load based on the area span value for task distribution. According to the invention, the rainfall inversion task can be flexibly adjusted, and the rainfall inversion processing efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a meteorological forecasting method based on deep learning. Background Technology

[0002] With the continuous changes in the natural environment, meteorological forecasting technology is of great significance for urban flood control. Among them, the impact of combining rainfall inversion with rain gauge detection results and remote sensing rainfall prediction results on the rainfall prediction accuracy of remote sensing data is crucial. However, in the existing technology, the distribution of rain gauges and urban topography and building features all affect the processing effect of rainfall inversion prediction tasks. Therefore, how to adaptively generate rainfall inversion prediction tasks according to actual scenarios to improve the processing efficiency of rainfall inversion prediction tasks is a technical problem that urgently needs to be solved by those skilled in the art.

[0003] Chinese Patent Publication No. CN109241212A discloses a method for inverting historical rainfall based on a mesoscale numerical atmospheric model and high resolution. The method includes: determining the rainfall data resolution, outputting rainfall data from the mesoscale numerical atmospheric model, dividing control units, calculating correction coefficients, and correcting rainfall data. Through this correction scheme, on the one hand, the spatial resolution information of the rainfall output data from the mesoscale numerical atmospheric model is preserved, reflecting the spatial distribution pattern of rainfall. On the other hand, the control area of ​​each rain gauge is divided using the Thiessen polygon method. Grids within the control area of ​​a particular rain gauge belong to that rain gauge, and the simulated grid rainfall is corrected by the measured rainfall information from that rain gauge. From the initial moment of rainfall, the areal rainfall at each moment is based on the measured data from the rain gauge. By distributing the spatial proportion of rainfall output from the mesoscale numerical atmospheric model to grids of different sizes, it is ensured that the areal rainfall on each control surface is equal to the measured rainfall. In this study, rainfall data output from mesoscale numerical atmospheric models is divided into different control units. Control units are assigned to all grid-based rainfall data within the study area. During this division, if the center point of a grid is located within the control area of ​​a rain gauge, the rainfall data of that grid belongs to the control unit of that rain gauge. This demonstrates that the division of grid-based rainfall data into control units is based solely on the distance between the grid's center point and the rain gauge. However, given the uneven distribution of rain gauges and differences in building features in real-world scenarios, grid division can easily result in poor performance, affecting the processing efficiency of rainfall inversion tasks and ultimately making it difficult to meet user needs for rainfall inversion prediction efficiency. Summary of the Invention

[0004] To address this, the present invention provides a weather forecasting method and system based on deep learning, which overcomes the problems of existing technologies that fail to consider uneven rainfall distribution within the target area, have a single processing method, and cannot flexibly adjust the rainfall inversion task when faced with the condition of missing rainfall information input devices within the target area, resulting in low accuracy of rainfall inversion results.

[0005] To achieve the above objectives, this invention provides a weather forecasting method based on deep learning, comprising: The area of ​​the region is determined based on the uniformity of the station locations. The target region is then evenly divided into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations in each sub-region, the sub-region is determined to be either a Class I or Class II region to be adjusted. Based on the ratio of the number of categories and the uniformity of regional distribution, determine whether to adjust the uniform division method to a state-fitting division or determine whether to adjust the region to be adjusted. The target optimal region corresponding to the second type of region to be adjusted is determined based on the number of drainage points in adjacent areas and the terrain reference height, and the target optimal region corresponding to the first type of region to be adjusted is determined based on the number of fusions and the terrain fusion difference value. The rainfall inversion task for the designated area is determined as a priority task or a baseline task based on the inversion factor volatility ratio of the designated area. Priority tasks and baseline tasks are combined and distributed based on priority. Furthermore, priority tasks are distributed based on the area span value or baseline tasks are distributed based on the remaining task conditions.

[0006] Furthermore, the category of a sub-region is determined based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to the sub-region; If the number of rain gauges in this region and the number of rain gauges in neighboring regions are both 0, then the sub-region is a type of region to be adjusted. If the number of rain gauges in this region is 0 and the number of rain gauges in neighboring regions is greater than 0, then the sub-region is a Class II region to be regulated.

[0007] Furthermore, for regions where the ratio of the number of categories is greater than the preset ratio of the number of categories or the regional distribution uniformity is less than or equal to the preset regional distribution uniformity, the uniform division method is adjusted to a state-fitting division.

[0008] Furthermore, for regions where the ratio of the number of categories is less than or equal to the preset ratio of the number of categories and the regional distribution uniformity is greater than the preset regional distribution uniformity, it is determined that the region to be adjusted will be subject to adjustment.

[0009] Furthermore, the adjustment methods for individual Category II regions to be adjusted include: Based on the number of drainage points in each adjacent area corresponding to the two types of areas to be regulated and the reference elevation of the terrain, the drainage capacity assessment value is determined, and the adjacent area with the largest drainage capacity assessment value is selected as the target optimal area corresponding to the two types of areas to be regulated.

[0010] Furthermore, the adjustment methods for a type of region to be adjusted include: Based on the number of fusions, the target areas with fewer than the preset number of fusions are selected, and the target area with the smallest terrain fusion difference value is selected as the optimal target area.

[0011] Furthermore, for a predefined region where the fluctuation state is such that the ratio of the inversion factor volatility is greater than the preset ratio of the inversion factor volatility, the rainfall inversion task corresponding to the predefined region is designated as a priority task. For a pre-defined region where the fluctuation state is such that the inversion factor volatility ratio is less than or equal to the preset inversion factor volatility ratio, the rainfall inversion task corresponding to the pre-defined region is recorded as the baseline task.

[0012] Furthermore, priority tasks are arranged in descending order of priority to obtain a priority task sequence, and each priority task is distributed sequentially. When distributing a single priority task, the lowest priority baseline task is combined with the priority task and distributed to any processing module.

[0013] Furthermore, based on the remaining conditions of the priority tasks, the priority tasks are distributed to the processing modules with the largest area span values ​​in descending order of priority. Based on the remaining conditions of the baseline task, the baseline task will be randomly assigned to the processing module with the lowest current processing load. The condition for the remaining priority tasks is that there are unassigned priority tasks; the condition for the remaining baseline tasks is that there are unassigned baseline tasks.

[0014] In addition, a forecasting system applying the aforementioned deep learning-based meteorological forecasting method is provided, comprising: The meteorological forecasting unit includes several processing modules used to perform rainfall inversion tasks; The regional division unit, which is connected to the meteorological prediction unit, is used to determine the area of ​​the region based on the uniformity of the station locations, and to divide the target region evenly into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to the sub-region, the sub-region is determined to be a Class I or Class II region to be adjusted. An adjustment analysis unit, which is connected to the region division unit, is used to determine whether to adjust the uniform division method to a state-fitting division or to determine the region to be adjusted based on the category quantity ratio and the regional distribution uniformity. The region search unit is connected to the region division unit and the adjustment analysis unit respectively. It is used to determine the target optimal region corresponding to the two types of regions to be adjusted based on the number of drainage points in adjacent regions and the terrain reference height, and to determine the target optimal region corresponding to the first type of region to be adjusted based on the number of fusions and the terrain fusion difference value. The task distribution unit, which is connected to the region search unit, is used to determine whether the rainfall inversion task of the set region is a priority task or a benchmark task based on the inversion factor fluctuation ratio of the set region, and to combine and distribute the priority task and the benchmark task based on priority, and to determine whether to distribute the priority task based on the region span value or the benchmark task based on the processing load based on the remaining task conditions.

[0015] Compared with the prior art, the beneficial effect of the present invention is that the technical solution of the present invention reflects the rain gauge setting around the location of the sub-region by the number of local rain gauges and the number of neighboring rain gauges corresponding to the sub-region, and determines the category of the sub-region as either a first-class or second-class area to be adjusted, thereby achieving accurate classification of sub-region categories and improving the processing accuracy of subsequent adjustment processing methods for different categories of sub-regions.

[0016] Furthermore, in this invention, the ratio of the number of categories in the sub-regions and the uniformity of regional distribution reflect the difference in quantity and the degree of distribution uniformity between the first-class and second-class regions to be adjusted, thereby determining whether the uniform division method meets the usage requirements. When the uniform division method does not meet the usage requirements, the regional division method is adjusted to state-fit division to improve the accuracy of regional division, making the regional division more consistent with the actual rainfall situation, thereby improving the processing accuracy of the subsequent rainfall inversion task corresponding to each sub-region.

[0017] Furthermore, in this invention, when adjusting the second type of area to be adjusted, the drainage capacity assessment value is determined based on the number of drainage points in each adjacent area corresponding to the second type of area to be adjusted and the reference elevation of the terrain. The drainage capacity assessment value reflects the intensity of the drainage capacity of the adjacent area, thereby determining whether to merge the second type of area to be adjusted into the target optimal area. It is understood that merging the second type of area to be adjusted into the target optimal area will inevitably affect the accuracy of the rainfall inversion after merging. If merged into an adjacent area with poor drainage capacity, it may lead to poor rainfall inversion accuracy in the adjacent area, which may cause safety problems when used for flood control and rescue.

[0018] Furthermore, in the adjustment process for a type of region to be adjusted in this invention, the degree of fusion is reflected by the number of fusions, thereby selecting adjacent regions with a smaller degree of fusion as the initial target region. This reduces the impact of excessive fusion causing a large area and low rainfall inversion accuracy. In addition, the optimal target region is selected based on the terrain fusion difference value to avoid the impact of large terrain differences between the optimal target region and the region to be adjusted after fusion on the rainfall inversion prediction accuracy. This improves the prediction accuracy of the rainfall inversion task of this invention.

[0019] Furthermore, in this invention, the fluctuation ratio of the inversion factor in a predetermined region reflects the degree of change of the influencing factors that interfere with the rainfall inversion task in that predetermined region, thereby correspondingly determining the priority task and the benchmark task, so as to effectively distinguish the priority of the rainfall inversion task and further improve the overall processing efficiency of the rainfall inversion task. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the weather forecasting method based on deep learning according to the present invention; Figure 2 This is a flowchart illustrating how the present invention determines the sub-region category based on the number of local rain gauges and the number of neighboring rain gauges corresponding to the sub-region. Figure 3 This is a flowchart illustrating how the adjustment method is determined based on the ratio of the number of categories and the uniformity of regional distribution in this invention. Figure 4 This is a unit connection diagram of the deep learning-based meteorological forecasting system of the present invention. Detailed Implementation

[0021] To make the objectives and advantages of this invention clearer, the invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0024] Please see Figures 1 to 3 As shown, this invention provides a weather forecasting method based on deep learning, comprising: The area of ​​the region is determined based on the uniformity of the station locations. The target region is then evenly divided into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations in each sub-region, the sub-region is determined to be either a Class I or Class II region to be adjusted. Based on the ratio of the number of categories and the uniformity of regional distribution, determine whether to adjust the uniform division method to a state-fitting division or determine whether to adjust the region to be adjusted. The target optimal region corresponding to the second type of region to be adjusted is determined based on the number of drainage points in adjacent areas and the terrain reference height, and the target optimal region corresponding to the first type of region to be adjusted is determined based on the number of fusions and the terrain fusion difference value. The rainfall inversion task for the designated area is determined as a priority task or a baseline task based on the inversion factor volatility ratio of the designated area. Priority tasks and baseline tasks are combined and distributed based on priority. Furthermore, priority tasks are distributed based on the area span value or baseline tasks are distributed based on the remaining task conditions.

[0025] In this invention, the target area is the scene area that needs to be predicted by rainfall inversion. The scene area can be, but is not limited to, urban areas or rural areas. The rain gauge has facilities to monitor the precipitation at its location, and the distribution of the rain gauge is determined by urban construction.

[0026] Specifically, this invention records parameter records, which record detection values ​​from historical working conditions that demonstrate that the accuracy of meteorological forecasts meets the actual needs of users. These detection values ​​include the number of drainage points, the reference elevation of the terrain, the number of fusions, the fluctuation of rainfall, the fluctuation of temperature, and the fluctuation of wind speed.

[0027] Specifically, the category of a sub-region is determined based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to the sub-region. If the number of rain gauges in this region and the number of rain gauges in neighboring regions are both 0, then the sub-region is a type of region to be adjusted. If the number of rain gauges in this region is 0 and the number of rain gauges in neighboring regions is greater than 0, then the sub-region is a Class II region to be regulated.

[0028] If the number of rain gauges in this region is greater than 0, then the sub-region is a qualified region.

[0029] In this invention, the area of ​​a region is determined based on the uniformity of station locations. Each sub-region is a rectangular area, and the area of ​​the region is the area corresponding to the sub-region. The area of ​​the region is positively correlated with the uniformity of station locations. The uniformity of station locations is denoted as... The formula for calculating the uniformity of site locations is:

[0030] in, , No. The rain gauge station and the distance to the first The distance between the nearest other rain gauges at each rain gauge station .

[0031] The area of ​​the target region is calculated as: Base area × (Preset station location uniformity / Station location uniformity). The values ​​for the base area and the preset station location uniformity are used to understand that a larger base area increases the probability of rain gauges existing in the resulting sub-regions. Station location uniformity reflects the evenness of the distribution of rain gauges within the target region; a higher uniformity further increases the probability of rain gauges existing in the evenly divided sub-regions. Therefore, the higher the user's requirement for the probability of rain gauges existing in the sub-regions, the larger the base area and the higher the value of the preset station location uniformity. One possible value is a base area of ​​600 km² and a preset station location uniformity of [value missing]. .

[0032] For a single sub-region, the number of local rain gauges is equal to the number of rain gauges within that sub-region. The number of rain gauge stations in the corresponding neighborhood is the total number of rain gauge stations in the adjacent areas corresponding to this sub-region. If there is another sub-region that overlaps with at least one corner point of the sub-region, then the other sub-region is the adjacent area of ​​the sub-region. It can be understood that if the sub-region is a rectangle, then the corner point of the rectangle is the intersection point of the two sides.

[0033] Specifically, for regions where the ratio of the number of categories is greater than the preset ratio of the number of categories or the regional distribution uniformity is less than or equal to the preset regional distribution uniformity, the uniform division method will be adjusted to a state-fitting division method.

[0034] The category ratio = number of Class I regions to be adjusted / number of Class II regions to be adjusted. The method for confirming the uniformity of regional distribution is to perform neighbor analysis for each region to be adjusted to obtain the number of invalid neighbors and the number of valid neighbors. The uniformity of regional distribution = (number of invalid neighbors + number of valid neighbors) / number of invalid neighbors. The neighbor analysis for a single region to be adjusted includes: temporarily designating the region to be adjusted as the temporary analysis region to be adjusted; detecting the number of other regions to be adjusted that are of the same category as the temporary analysis region to be adjusted as the sub-invalid neighbor number; and designating the number of regions to be adjusted that are of a different category than the temporary analysis region to be adjusted as the sub-valid neighbor number. The valid neighbor number is the average of the sub-valid neighbor numbers, and the invalid neighbor number is the average of the sub-invalid neighbor numbers. It is easy to understand that the regions to be adjusted include Class I regions to be adjusted and Class II regions to be adjusted.

[0035] The preset category ratio and preset regional distribution uniformity values ​​are understood to reflect the distribution of the first and second types of areas to be adjusted. A larger category ratio indicates a higher proportion of the first type of area to be adjusted, while a lower regional distribution uniformity indicates a worse distribution uniformity between the first and second types of areas to be adjusted. It is easy to see that a smaller category ratio and a larger regional distribution uniformity result in a better uniform division method. Therefore, in specific value settings, the higher the user's requirements for the uniform division method, the smaller the preset category ratio and the larger the preset regional distribution uniformity value. A specific implementation value is provided: a preset category ratio of 1.2 and a preset regional distribution uniformity of 2.

[0036] In the state-fitting partitioning, predefined regions are constructed with each rain gauge as the center. These predefined regions are circular. For the first... Each rain gauge has a corresponding predefined area radius of [radius value missing]. .

[0037] Specifically, for regions where the ratio of the number of categories is less than or equal to a preset ratio of the number of categories and the regional distribution uniformity is greater than a preset regional distribution uniformity, it is determined that the region to be adjusted will be adjusted.

[0038] Specifically, the adjustment methods for a single Category II region to be adjusted include: Based on the number of drainage points in each adjacent area corresponding to the two types of areas to be regulated and the reference elevation of the terrain, the drainage capacity assessment value is determined, and the adjacent area with the largest drainage capacity assessment value is selected as the target optimal area corresponding to the two types of areas to be regulated.

[0039] Once the target optimal region is confirmed, for the second type of region to be adjusted, the second type of region to be adjusted is recorded as its corresponding target optimal region.

[0040] Drainage capacity assessment value = Number of drainage points / Preset number of drainage points + Terrain reference height / Preset terrain reference height; For a single adjacent area, the number of drainage points is the number of facilities with drainage capacity in that adjacent area. Facilities with drainage capacity include, but are not limited to, storm drains, drainage channels, and floodgates. The method for determining the terrain reference height corresponding to a single adjacent area is to uniformly obtain the elevation of several points in that adjacent area and record the average elevation as the terrain reference height. The elevation can be obtained by, but is not limited to, GNSS-RTK measurement and querying public DEM data (such as querying through QGIS or ArcGIS).

[0041] The preset number of drainage points and preset terrain reference height are meant to reflect the drainage capacity of the adjacent area. The higher the user's requirements for drainage capacity, the higher the preset number of drainage points and preset terrain reference height. A method for setting these values ​​is provided: extract the parameters and record the corresponding number of drainage points and terrain reference height, remove outliers from the number of drainage points and terrain reference height, and use the average value of the number of drainage points and terrain reference height after removing outliers as the preset number of drainage points and preset terrain reference height.

[0042] Specifically, adjustments are made sequentially to each type of region to be adjusted in a random order. The adjustment methods for each type of region to be adjusted include: Based on the number of fusions, adjacent areas are filtered to obtain initial target areas with fewer than the preset number of fusions, and the initial target area with the smallest terrain fusion difference value is selected as the optimal target area.

[0043] For a single adjacent region, the number of fusions is the total number of times that adjacent region was recorded as the target optimal region before the current time, and the corresponding terrain fusion difference value is the absolute value of the difference between the terrain reference height of the region to be adjusted and the adjacent region.

[0044] Once the target optimal region is confirmed, for a type of region to be adjusted, that type of region to be adjusted is recorded in its corresponding target optimal region.

[0045] After the adjustment processing of each type I and type II region to be adjusted is completed, the target optimal region and qualified region obtained are recorded as the set region respectively.

[0046] The preset number of fusions is understood to be such that the more fusions a target area is selected, the larger its area will be. However, if rainfall is inverted for that target area, the correction effect of the rainfall prediction result based on the rain gauge will be worse. Therefore, the higher the user's requirement for the accuracy of rainfall inversion prediction, the lower the preset number of fusions will be. A method is provided to extract the parameters that meet the user's needs and record the corresponding number of fusions. Outliers in the number of fusions are removed, and the average value of the number of fusions after removing outliers is taken as the preset number of fusions.

[0047] Specifically, for a pre-defined region where the fluctuation ratio of the inversion factor is greater than the preset fluctuation ratio of the inversion factor, the rainfall inversion task corresponding to the pre-defined region is designated as a priority task; For a pre-defined region where the fluctuation state is such that the inversion factor volatility ratio is less than or equal to the preset inversion factor volatility ratio, the rainfall inversion task corresponding to the pre-defined region is recorded as the baseline task.

[0048] For a single defined area, the method for determining the corresponding inversion factor volatility ratio is as follows: the ratio with the largest value among the first, second, and third ratios is recorded as the inversion factor volatility ratio. This invention applies several inversion influencing factors, including rainfall volatility, temperature volatility, and wind speed volatility. Rainfall volatility is the absolute value of the difference between the maximum and minimum unit rainfall detected by rain gauges within the defined area in the most recent monitoring period. Wind speed volatility is the maximum wind speed detected at the corresponding installation location of the rain gauge within the defined area in the most recent monitoring period. The absolute value of the difference between the maximum wind speed and the minimum wind speed is used. The temperature fluctuation is the absolute value of the difference between the maximum temperature and the minimum temperature at the installation location of the rain gauge in the most recent monitoring period. After each fixed time interval, the rainfall within that fixed time interval and the wind speed and temperature at the end of the fixed time interval are measured. The rainfall within the fixed time interval is the unit rainfall. The rain gauge is equipped with a temperature sensor to detect the ambient temperature at the installation location of the rain gauge and a wind speed sensor to detect the ambient wind speed at the installation location of the rain gauge.

[0049] In this invention, the first ratio = rainfall fluctuation / preset rainfall fluctuation, the second ratio = temperature fluctuation / preset temperature fluctuation, and the third ratio = wind speed fluctuation / preset wind speed fluctuation. The values ​​of the preset rainfall fluctuation, preset temperature fluctuation, and preset wind speed fluctuation are understood to be such that higher values ​​have a greater impact on the rainfall inversion task, resulting in lower accuracy of the inversion results. Therefore, the higher the accuracy requirements of the rainfall inversion results, the higher the values ​​of the preset rainfall fluctuation, preset temperature fluctuation, and preset wind speed fluctuation. A method for determining these values ​​is provided: extracting the corresponding rainfall fluctuation, temperature fluctuation, and wind speed fluctuation records that meet user needs; removing outliers from the rainfall fluctuation, temperature fluctuation, and wind speed fluctuation; and using the average values ​​of the rainfall fluctuation, temperature fluctuation, and wind speed fluctuation after removing outliers as the preset rainfall fluctuation, preset temperature fluctuation, and preset wind speed fluctuation, respectively.

[0050] The monitoring cycle is a cyclical period, with each monitoring cycle having the same duration. Within a single monitoring cycle, there are several fixed time intervals. It is easy to understand that the higher the accuracy of the user's acquisition of rainfall fluctuations, temperature fluctuations, and wind speed fluctuations, the shorter the corresponding fixed time intervals and monitoring cycle durations will be. It is worth noting that the fixed time interval duration is shorter than the monitoring cycle duration, providing one possible value: a single monitoring cycle of 1 hour and a fixed time interval duration of 5 minutes.

[0051] Specifically, priority tasks are arranged in descending order of priority to obtain a priority task sequence. Each priority task is then distributed sequentially. When distributing a single priority task, the lowest priority baseline task is combined with the priority task and distributed to any processing module.

[0052] The priority confirmation method for the priority tasks is determined by the user. The user can determine the priority of the set area based on the urgency of the flood control needs corresponding to the set area. The priority is the priority corresponding to the priority task. The priority can be determined based on, but is not limited to, the number of key protection facilities in the set area, whether there are people living there, and the probability of water accumulation leading to flooding. This is content that is already known to those skilled in the art and will not be elaborated here. Specifically, when distributing priority tasks, they are sent to the processing module with the lowest current processing load.

[0053] Specifically, for the remaining conditions of priority tasks, priority tasks are distributed to the processing module with the largest area span value in descending order of priority. Based on the remaining conditions of the baseline task, the baseline task will be randomly assigned to the processing module with the lowest current processing load. The condition for the remaining priority tasks is that there are unassigned priority tasks; the condition for the remaining baseline tasks is that there are unassigned baseline tasks.

[0054] For a single remaining priority task and any processing module, the average distance between the corresponding set area location of the priority task and the corresponding set area of ​​each rainfall inversion task in the processing module is recorded as the area span value. The distance between any two set areas is the distance between the center points of the areas of the two set areas. The area span value reflects the concentration of set areas in the processing module. It can be understood that the smaller the area span value, the closer the distance between the set area location of the remaining priority task and the corresponding set area location of the group and task in the processing module is. If the processing module crashes, it is easy to cause a delay in the processing of rainfall inversion tasks in a large area of ​​set areas, which may lead to system crash. The processing load is the current CPU utilization rate of the processing module, which is common knowledge and will not be elaborated here.

[0055] Please see Figure 4 The diagram shown is a unit connection diagram of the deep learning-based meteorological forecasting system of the present invention. A deep learning-based meteorological forecasting system is provided, comprising: The meteorological forecasting unit includes several processing modules used to perform rainfall inversion tasks; The regional division unit, which is connected to the meteorological prediction unit, is used to determine the area of ​​the region based on the uniformity of the station locations, and to divide the target region evenly into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to the sub-region, the sub-region is determined to be a Class I or Class II region to be adjusted. An adjustment analysis unit, which is connected to the region division unit, is used to determine whether to adjust the uniform division method to a state-fitting division or to determine the region to be adjusted based on the category quantity ratio and the regional distribution uniformity. The region search unit is connected to the region division unit and the adjustment analysis unit respectively. It is used to determine the target optimal region corresponding to the two types of regions to be adjusted based on the number of drainage points in adjacent regions and the terrain reference height, and to determine the target optimal region corresponding to the first type of region to be adjusted based on the number of fusions and the terrain fusion difference value. The task distribution unit, which is connected to the region search unit, is used to determine whether the rainfall inversion task of the set region is a priority task or a benchmark task based on the inversion factor fluctuation ratio of the set region, and to combine and distribute the priority task and the benchmark task based on priority, and to determine whether to distribute the priority task based on the region span value or the benchmark task based on the processing load based on the remaining task conditions.

[0056] The processing module is a computing device with data processing capabilities. It is equipped with a rainfall prediction model based on remote sensing data. Specifically, after receiving a rainfall inversion task, the processing module predicts the rainfall for a complete monitoring cycle after the current time for the designated area corresponding to each rainfall inversion task. After the monitoring cycle ends, it obtains the average rainfall detected by each rain gauge station in the designated area during the monitoring cycle. The rainfall during the monitoring cycle is the rainfall corresponding to the installation location of the rain gauge station during the monitoring cycle. Users can determine whether the model accuracy meets the usage requirements based on the deviation between the rainfall detected by the rain gauge station and the rainfall predicted by the rainfall prediction model. This is content that is already known to those skilled in the art and will not be elaborated here.

[0057] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A weather forecasting method based on deep learning, characterized in that, include: The area of ​​the region is determined based on the uniformity of the station locations. The target region is then evenly divided into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations in each sub-region, the sub-region is determined to be either a Class I or Class II region to be adjusted. Based on the ratio of the number of categories and the uniformity of regional distribution, determine whether to adjust the uniform division method to a state-fitting division or determine whether to adjust the area to be adjusted. The target optimal region corresponding to the second type of region to be adjusted is determined based on the number of drainage points in adjacent areas and the terrain reference height, and the target optimal region corresponding to the first type of region to be adjusted is determined based on the number of fusions and the terrain fusion difference value. The rainfall inversion task for the designated area is determined as a priority task or a baseline task based on the inversion factor volatility ratio of the designated area. Priority tasks and baseline tasks are combined and distributed based on priority. Furthermore, priority tasks are distributed based on the area span value or baseline tasks are distributed based on the remaining task conditions.

2. The weather forecasting method based on deep learning according to claim 1, characterized in that, The category of a sub-region is determined based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to that sub-region. If the number of rain gauges in this region and the number of rain gauges in neighboring regions are both 0, then the sub-region is a type of region to be adjusted. If the number of rain gauges in this region is 0 and the number of rain gauges in neighboring regions is greater than 0, then the sub-region is a Class II region to be regulated.

3. The weather forecasting method based on deep learning according to claim 2, characterized in that, For regions where the ratio of the number of categories is greater than the preset ratio of the number of categories or the regional distribution uniformity is less than or equal to the preset regional distribution uniformity, the uniform division method will be adjusted to a state-fitting division method.

4. The weather forecasting method based on deep learning according to claim 2, characterized in that, For regions where the ratio of the number of categories is less than or equal to the preset ratio of the number of categories and the regional distribution uniformity is greater than the preset regional distribution uniformity, the regions to be adjusted are determined to be adjusted.

5. The weather forecasting method based on deep learning according to claim 4, characterized in that, The adjustment methods for a single Category II region to be adjusted include: Based on the number of drainage points in each adjacent area corresponding to the two types of areas to be regulated and the reference elevation of the terrain, the drainage capacity assessment value is determined, and the adjacent area with the largest drainage capacity assessment value is selected as the target optimal area corresponding to the two types of areas to be regulated.

6. The weather forecasting method based on deep learning according to claim 4, characterized in that, The adjustment methods for a certain type of region to be adjusted include: Based on the number of fusions, adjacent areas are filtered to obtain initial target areas with fewer than the preset number of fusions, and the initial target area with the smallest terrain fusion difference value is selected as the optimal target area.

7. The deep learning-based weather forecasting method according to any one of claims 3, 5, or 6, characterized in that, For a pre-defined region where the fluctuation state is such that the ratio of the inversion factor volatility is greater than the preset ratio of the inversion factor volatility, the rainfall inversion task corresponding to the pre-defined region is designated as a priority task. For a pre-defined region where the fluctuation state is such that the inversion factor volatility ratio is less than or equal to the preset inversion factor volatility ratio, the rainfall inversion task corresponding to that pre-defined region is recorded as the baseline task.

8. The weather forecasting method based on deep learning according to claim 7, characterized in that, Priority tasks are arranged in descending order of priority to obtain a priority task sequence. Each priority task is then distributed sequentially. When distributing a single priority task, the lowest priority baseline task is combined with the priority task and distributed to any processing module.

9. The weather forecasting method based on deep learning according to claim 8, characterized in that, For the remaining conditions of priority tasks, the priority tasks are distributed to the processing modules with the largest area span values ​​in descending order of priority. Based on the remaining conditions of the baseline task, the baseline task will be randomly assigned to the processing module with the lowest current processing load. The condition for the remaining priority tasks is that there are unassigned priority tasks; the condition for the remaining baseline tasks is that there are unassigned baseline tasks.

10. A forecasting system applying the deep learning-based meteorological forecasting method according to any one of claims 1 to 9, characterized in that, include: The meteorological forecasting unit includes several processing modules used to perform rainfall inversion tasks; The regional division unit, which is connected to the meteorological prediction unit, is used to determine the area of ​​the region based on the uniformity of the station locations, and to divide the target region evenly into several sub-regions. Based on the number of local rain gauge stations and the number of neighboring rain gauge stations corresponding to the sub-region, the sub-region is determined to be a Class I or Class II region to be adjusted. An adjustment analysis unit, which is connected to the region division unit, is used to determine whether to adjust the uniform division method to a state-fitting division or to determine the region to be adjusted based on the category quantity ratio and the regional distribution uniformity. The region search unit is connected to the region division unit and the adjustment analysis unit respectively. It is used to determine the target optimal region corresponding to the two types of regions to be adjusted based on the number of drainage points in adjacent regions and the terrain reference height, and to determine the target optimal region corresponding to the first type of region to be adjusted based on the number of fusions and the terrain fusion difference value. The task distribution unit, which is connected to the region search unit, is used to determine whether the rainfall inversion task of the set region is a priority task or a benchmark task based on the inversion factor fluctuation ratio of the set region, and to combine and distribute the priority task and the benchmark task based on priority, and to determine whether to distribute the priority task based on the region span value or to analyze the benchmark task based on the processing load based on the remaining task conditions.

Images