High mountain valley region hydropower station dam region short-time approaching rainfall forecasting method based on deep learning multi-model fusion

Abstract

The invention relates to the field of meteorological observation, and particularly discloses a high mountain valley region hydropower station dam region short-time approaching rainfall forecasting method based on deep learning multi-model fusion, and the method comprises the steps: obtaining a three-dimensional grid structure of a target region; performing short-time extrapolation prediction on the three-dimensional grid structure through a plurality of time sequence prediction models to obtain a plurality of short-time extrapolation prediction results; based on a deep learning model, performing fusion processing on the plurality of short-time extrapolation prediction results to obtain a fused extrapolation prediction result; and generating a short-time approaching rainfall forecast of the target area based on the fused extrapolation prediction result. According to the method, radar echo extrapolation is performed on the three-dimensional grid structure in combination with a plurality of time sequence prediction models so as to predict the three-dimensional grid structure, the short-time extrapolation prediction results are obtained, the plurality of short-time extrapolation prediction results are fused by using the deep learning model, the short-time approaching rainfall forecast of the target area is generated, the accuracy of the short-time approaching rainfall forecast is improved, and the accuracy of the short-time approaching rainfall forecast is improved. Meteorological disaster prevention and reduction of the dam area of the hydropower station are facilitated, and meteorological guarantee is provided for safe production of the hydropower station.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation, and in particular to a method for short-term nowcasting of precipitation in the dam area of ​​hydropower stations in high mountain valleys based on deep learning multi-model fusion. Background Technology

[0002] Hydropower station dam areas in high mountain and valley terrain typically feature deep canyon topography, low-latitude plateau monsoon climate, and distinct wet and dry seasons. Under the influence of complex terrain, local circulation is significant, and severe weather events such as short-duration heavy rainfall, thunderstorms, and strong winds occur frequently, posing significant meteorological risks to the safe operation of hydropower stations. Traditional extrapolation methods and numerical models based on physical equations have limited ability to characterize local weather in such areas. The generation and dissipation of radar echoes, their deformation, and topographically induced local convection are difficult to accurately describe. Predictions using single models or fusion of multiple models are inaccurate, resulting in insufficient accuracy in 0-2 hour short-term forecasts, which are easily affected by weather, terrain, and model limitations. Therefore, a new method for short-term nowcasting precipitation in high mountain and valley hydropower station dam areas is needed to achieve stable precipitation forecasting in complex regions and environments. Summary of the Invention

[0003] This invention proposes a short-term nowcasting method for hydropower station dam areas in high-altitude valley regions based on deep learning multi-model fusion. It introduces adaptive 3D grid partitioning, a dynamic weighted attention fusion mechanism, and a model accuracy evaluation system to achieve stable precipitation forecasting for complex regions and environments. The method spatially deconstructs the complex terrain of the target area using a 3D grid structure, refining the spatial division of precipitation conditions and radar observation data for each grid cell. Simultaneously, multiple time-series prediction models are combined to extrapolate radar echoes from the 3D grid structure to obtain multiple short-term extrapolation prediction results. These results are then fused using a deep learning model to generate a short-term nowcasting forecast for the target area. This improves the accuracy of short-term nowcasting for hydropower station dam areas in high-altitude valley regions, contributing to meteorological disaster prevention and mitigation in hydropower station dam areas and providing meteorological support for the safe operation of hydropower stations.

[0004] To achieve the above objectives, this invention proposes a short-term nowcasting method for hydropower station dam areas in high-altitude valley regions based on deep learning multi-model fusion. The method includes: A three-dimensional grid structure of the target area is obtained. The three-dimensional grid structure includes grid cells with spatial connection relationships. Each grid cell includes grid cell location data, historical precipitation data of the grid cell, and historical radar observation data of the grid cell. The target area is the dam area of ​​the hydropower station in the high mountain valley area. Multiple time-series prediction models are used to perform short-time extrapolation predictions on the three-dimensional grid structure to obtain multiple short-time extrapolation prediction results. The time-series prediction models are trained based on the three-dimensional grid structure of the target area, historical precipitation data, and historical radar observation data. Each time-series prediction model corresponds to one short-time extrapolation prediction result. Based on a deep learning model, multiple short-term extrapolation prediction results are fused to obtain a fused extrapolation prediction result. Based on the fusion extrapolation prediction results, a short-term nowcast precipitation forecast for the target area is generated.

[0005] This invention presents a technical solution for short-term nowcasting precipitation in the dam area of ​​a hydropower station in a high-altitude valley region, based on a deep learning multi-model fusion method. The method involves acquiring a three-dimensional grid structure of the target area, comprising grid cells with spatial connections. Each grid cell includes grid cell location data, historical precipitation data, and historical radar observation data. The target area is the dam area of ​​a hydropower station in the high-altitude valley region. Multiple time-series prediction models are used to perform short-term extrapolation predictions on the three-dimensional grid structure, resulting in multiple short-term extrapolation prediction results. These time-series prediction models are trained based on the three-dimensional grid structure, historical precipitation data, and historical radar observation data of the target area. Each time-series prediction model corresponds to one short-term extrapolation prediction result. Based on a deep learning model, the multiple short-term extrapolation prediction results are fused to obtain a fused extrapolation prediction result. Finally, a short-term nowcasting precipitation forecast for the target area is generated based on the fused extrapolation prediction result. This invention spatially deconstructs the complex terrain of the target area using a three-dimensional grid structure, finely dividing the precipitation and radar observation data of each grid cell. Simultaneously, it combines multiple time-series prediction models to perform radar echo extrapolation on the three-dimensional grid structure to obtain multiple short-term extrapolation prediction results. A deep learning model is then used to fuse these multiple short-term extrapolation prediction results, and finally, the fused extrapolation prediction results are used to generate short-term nowcast precipitation forecasts for the target area. This improves the accuracy of short-term nowcasting for hydropower station dam areas in high mountain and valley regions, which is helpful for meteorological disaster prevention and mitigation in hydropower station dam areas and provides meteorological support for the safe production of hydropower stations. Attached Figure Description

[0006] Figure 1 A flowchart illustrating a method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the specific process of step S1 in an embodiment of the present invention; Figure 3 This is a schematic diagram of the specific process of step S13 in an embodiment of the present invention; Figure 4This is a schematic diagram of the specific process of step S5 in an embodiment of the present invention; Figure 5 This is a schematic diagram of the specific process of step S51 in an embodiment of the present invention; Figure 6 This is a schematic diagram of the specific process of step S6 in an embodiment of the present invention; Figure 7 This is a schematic diagram of the specific process of step S62 in an embodiment of the present invention; Figure 8 This is a schematic diagram of the specific process of step S624 in an embodiment of the present invention; Figure 9 This is a schematic diagram of the specific process of step S4 in an embodiment of the present invention; Figure 10 This is a schematic diagram of the specific process of step S7 in an embodiment of the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0008] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0009] It should also be noted that when a component is described as "fixed to" or "set on" another component, it can be directly on the other component or there may be an intervening component present. When a component is described as "connected to" another component, it can be directly connected to the other component or there may be an intervening component present.

[0010] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0011] See Figure 1 , Figure 1 This is a flowchart illustrating the short-term precipitation forecasting method for hydropower station dam areas in high mountain and valley regions based on deep learning multi-model fusion, provided in an embodiment of the present invention. The method includes the following steps: S1, obtain the three-dimensional mesh structure of the target area.

[0012] The three-dimensional grid structure includes grid cells with spatial connections. Each grid cell includes grid cell location data, historical precipitation data, and historical radar observation data. The target area is the dam area of ​​a hydropower station in a high mountain valley region.

[0013] S2, the three-dimensional mesh structure is subjected to short-time extrapolation prediction through multiple time-series prediction models to obtain multiple short-time extrapolation prediction results.

[0014] The time-series prediction models are trained based on the three-dimensional grid structure of the target area, historical precipitation data, and historical radar observation data. Each time-series prediction model corresponds to a short-time extrapolation prediction result.

[0015] S3, based on a deep learning model, fuses multiple short-term extrapolation prediction results to obtain a fused extrapolation prediction result.

[0016] S4 generates short-term nowcast precipitation forecasts for the target area based on the fusion extrapolation prediction results.

[0017] In the embodiments of the present invention, it can be understood that the above-mentioned method for short-term now-near-precipitation forecasting of hydropower station dam areas in high mountain valleys based on deep learning multi-model fusion is deployed on a cloud server or a local server in the target area.

[0018] The aforementioned target area can be understood as the specific geographical range requiring refined precipitation forecasting. In this embodiment, it may specifically refer to the dam area of ​​a hydropower station and its surrounding key impact areas in a high mountain and river valley region. This area is typically determined based on meteorological risk analysis, the flood control and dispatching needs of the hydropower station, and the effective detection range of radar.

[0019] The aforementioned three-dimensional grid structure can be seen as a digital and structured data representation of the forecast area and its historical meteorological environment. The three-dimensional grid structure discretizes continuous physical space (longitude, latitude, and altitude) into a series of small, regularly or irregularly arranged units in space.

[0020] Three-dimensional mesh structures can be constructed by processing raw geographic and meteorological data, such as terrain data from digital elevation models (DEMs).

[0021] The aforementioned grid cells are the basic units constituting the "three-dimensional grid structure," serving as the fundamental objects for all calculations and forecasts. Each grid cell represents a small area of ​​three-dimensional space within the target region. Each grid cell is described by three types of core data: grid cell location data, historical precipitation data, and historical radar observation data.

[0022] The grid cell location data describes the geometric location of the grid cell in three-dimensional space. This can be the latitude and longitude coordinates of the grid cell's center point, its altitude, or its row and column index within the entire grid structure. It is obtained by recording the parsing and gridding process of the target area's DEM data. DEM data is a raster dataset containing ground elevation information and can be obtained from Geographic Information System (GIS) data or satellite remote sensing data.

[0023] The historical precipitation data for the aforementioned grid cells represent the observed or accumulated precipitation intensity or amount over a past period at the spatial location represented by that grid cell. This data can be obtained by mapping and interpolating the same ordinate from measured precipitation data from automatic weather stations. Since automatic stations are distributed as discrete points, spatial interpolation methods are needed to transform the station data into a continuous grid field covering the entire target area, which is then matched with a three-dimensional grid structure to assign a historical precipitation data sequence to each grid cell. Examples of such spatial interpolation methods include Kriging interpolation and inverse distance weighting. The aforementioned historical radar observation data for the grid cells can be understood as a sequence of echo intensity over a past period, obtained from meteorological radar observations, covering the spatial location of the grid cell. Radar echo intensity is an indirect reflection of the size and concentration of water droplets in precipitation clouds and is one of the most important inputs for short-term forecasting. Historical radar observation data for the grid cells can be obtained from meteorological radar baseline data. The radar acquires reflectivity factor data in three-dimensional space through scanning. After preprocessing such as quality control and ground clutter suppression, the reflectivity factor data is mapped to a coordinate system and resolution matching the aforementioned three-dimensional grid structure, thereby generating time-series radar echo observation data for each grid cell.

[0024] Specifically, based on the geographical area of ​​the hydropower station dam area and its surroundings, a digital elevation model (DEM), historical automatic weather station precipitation data, and historical radar observation data for the region can be obtained. Then, based on this data, a unified three-dimensional grid coordinate system is constructed. Next, historical precipitation data is spatially interpolated into the grid, and historical radar data is resampled to the grid through coordinate mapping. The spatial location of each grid cell is then determined in conjunction with the DEM data. Finally, a structured dataset is formed, where each grid cell contains its location, historical precipitation sequence, and historical radar echo sequence. This structure provides a standardized input format for subsequent models.

[0025] The three-dimensional mesh structure data is simultaneously input into multiple pre-trained time-series prediction models. These models can be deep learning models capable of handling time-series problems, such as CNN-Attention, Conv-LSTM, and PredRNN++. These models have learned from historical data, specifically precipitation data and radar echo sequences from past moments, to predict radar echo sequences within a short future period (e.g., 0-2 hours). Each model, due to differences in its network structure or training strategy, will produce a slightly different prediction result. For each short-term extrapolation prediction, the result represents a radar echo prediction data field for the same future weather conditions.

[0026] Multiple short-term extrapolated predictions are input into a pre-trained deep learning model. This deep learning model can be a bagging, boosting, or stacking model, or a neural network with an attention mechanism. The deep learning model intelligently analyzes, compares, and weights the information from different preliminary predictions. By learning which model's prediction is more reliable under different weather conditions, the deep learning model combines the advantages of multiple time-series prediction models, suppresses the errors of individual models, and ultimately generates a single prediction that is theoretically more accurate and stable—this is known as fused extrapolated prediction.

[0027] The fusion prediction based on radar echo extrapolation is transformed into the final required precipitation forecast information. Specifically, the fusion extrapolation prediction results are used as input to a transformation model. Through the transformation model (e.g., a quantitative precipitation estimation model based on random forests, neural networks, or ZR relationships), the future echo intensity represented by the extrapolation is transformed or interpreted as the possible precipitation intensity or cumulative amount in each grid cell of the target area within the next 0-2 hours. This embodiment considers the local precipitation-echo statistical relationship, and the final output is an intuitive, gridded short-term nowcast precipitation forecast product, which can be directly used for meteorological risk warning and hydropower station scheduling decisions.

[0028] This embodiment establishes a standard input through data gridding; then, multiple deep learning models are used for parallel preliminary prediction to capture uncertainties; next, another deep learning model is used for intelligent fusion to improve forecast stability; finally, the fused physical quantity predictions are transformed into quantitative precipitation forecasts. By closely integrating the special needs of high-altitude valley hydropower station dam areas and leveraging the advantages of deep learning in processing complex spatiotemporal data relationships, this approach aims to effectively improve the accuracy and practicality of short-term nowcasting precipitation forecasts in this region, providing key technical support for disaster prevention and mitigation.

[0029] Specifically, this embodiment can spatially deconstruct the complex terrain of the target area using a three-dimensional grid structure, finely dividing the precipitation and radar observation data of each grid cell. At the same time, it combines multiple time-series prediction models to perform radar echo extrapolation on the three-dimensional grid structure to obtain multiple short-term extrapolation prediction results. The multiple short-term extrapolation prediction results are then fused using a deep learning model, and the fused extrapolation prediction results are used to generate short-term nowcast precipitation forecasts for the target area. This improves the accuracy of short-term nowcasting for hydropower station dam areas in high mountain and valley regions, helps with meteorological disaster prevention and mitigation in hydropower station dam areas, and provides meteorological support for the safe production of hydropower stations.

[0030] Optionally, step S1 specifically includes: S11 acquires three-dimensional topographic maps, historical precipitation data, and historical radar observation data of the target area.

[0031] S12, based on historical precipitation data and historical radar observation data, determines the actual precipitation range, actual precipitation intensity, forecast precipitation range, and forecast precipitation intensity for each precipitation event within the target area.

[0032] S13, based on the actual precipitation range, actual precipitation intensity, forecast precipitation range, and forecast precipitation intensity, determine the smallest segmentation unit.

[0033] S14 uses the smallest segmentation unit to perform grid segmentation on the three-dimensional topographic map to obtain the initial three-dimensional grid structure.

[0034] The initial three-dimensional mesh structure includes initial mesh volumes with spatial connections, and the initial mesh volumes include the location information of the initial mesh volumes in the three-dimensional terrain map; S15, for each initial grid cell, extract historical precipitation data from the historical precipitation data for a preset duration prior to the current time, and use it as the grid cell's historical precipitation data; and extract historical radar observation data from the historical radar observation data for a preset duration prior to the current time, and use it as the grid cell's historical radar observation data.

[0035] S16. Historical precipitation data and historical radar observation data of the grid are added to the initial grid, and the density is adjusted according to the terrain features of the target area to obtain the three-dimensional grid structure of the target area.

[0036] In this embodiment of the invention, raw, discrete, and multi-source meteorological and geographical data are transformed and organized into a unified, structured three-dimensional data model that is easy for deep learning models to process, thereby obtaining a more refined three-dimensional grid structure. This embodiment does not use a fixed, empirical grid resolution, but rather dynamically determines a minimum segmentation unit based on statistical analysis of historical forecast errors. This unit serves as the basis for constructing the three-dimensional grid structure, making the grid structure more adaptable to the specific spatial scale and forecast error characteristics of the precipitation system in the target area. This improves the accuracy of the three-dimensional grid structure and further enhances the forecast accuracy of short-term nowcast precipitation in the dam areas of hydropower stations in high mountain and valley regions.

[0037] A 3D topographic map of the target area can be understood as a 3D digital model that reflects the surface undulations of the target area (high mountain valley hydropower station dam area and its surroundings). It can be digital elevation model (DEM) or digital terrain model (DTM) data. DTM data can be obtained through satellite remote sensing, aerial photogrammetry, or lidar scanning, and is composed of elevation values ​​of regular grid points.

[0038] Historical precipitation data refers to time-series precipitation data observed and recorded by automatic weather stations in and around a target area over a relatively long period (such as several years). It typically includes the start and end times of precipitation, as well as the amount or intensity of precipitation. Historical precipitation data can be obtained from the operational observation database of a network of automatic weather stations deployed in and around the power station dam area. Automatic weather stations can provide continuous, location-specific, and high temporal resolution precipitation measurements.

[0039] Historical radar observation data can be a sequence of echo intensity data detected by meteorological radar covering the target area during the same period as the aforementioned historical precipitation. Radar data can provide large-scale, high spatiotemporal resolution three-dimensional structure information of precipitation clouds. Historical radar observation data is extracted from the baseline data of weather radar stations covering the target area. The baseline data consists of quality-controlled radar echo data stored in a three-dimensional spherical coordinate or two-dimensional planar position display format.

[0040] It should be noted that the historical precipitation data is aligned in both time and space.

[0041] Each independent precipitation event is identified and separated from historical precipitation data and historical radar observation data. For each precipitation event, the actual precipitation range and actual precipitation intensity are determined. Based on measured data from automatic weather stations, spatial interpolation and analysis can be used to determine the actual spatial range of influence on the ground (such as the area enclosed by isohyets) and the typical or average precipitation intensity within that range.

[0042] The predicted precipitation range and predicted precipitation intensity can be based on the forecast results released at the same time as this precipitation event by a certain historical forecasting method or product (e.g., traditional radar extrapolation forecast or a certain version of numerical model forecast), thereby determining the predicted spatial range and predicted intensity of precipitation corresponding to this precipitation event.

[0043] Considering that the grid size should effectively characterize the spatial distribution of forecast errors, for each precipitation event, its extent error and intensity error can be calculated. The extent error can be the degree of spatial mismatch between the actual and forecast extents, specifically the intersection-union ratio (IUU) of the actual and forecast extents. The intensity error can be the difference or ratio between the actual and forecast intensity.

[0044] Specifically, historical precipitation events can be extracted from all historical precipitation data. From these events, the one with the smallest range error (i.e., the most accurate spatial location forecast) that is greater than zero can be identified. The overlapping and non-overlapping portions of the actual and forecast precipitation ranges within this event are then analyzed. By performing specific modulo operations on the spatial geometric characteristics (such as area and length) of these overlapping and non-overlapping regions, a remainder value is derived and used as the side length of a cube, forming the initial segmentation unit.

[0045] Considering that intensity error also reflects forecasting capability, the average intensity error of all events is calculated, and an adjustment factor is derived accordingly. This adjustment factor is used to adjust the size of the "initial segmentation unit," ultimately yielding the smallest three-dimensional volumetric unit used to segment the entire region, i.e., the minimum segmentation unit.

[0046] In this embodiment, the most accurate precipitation forecast in history is first found, and the deviation scale of this forecast in detail (edge) is analyzed. The grid size is initially determined based on this scale. Then, considering the historically common intensity forecast error level, this initial size is fine-tuned to obtain a theoretically optimal basic grid size that comprehensively considers historical and spatial error characteristics.

[0047] A specific mesh structure is constructed and populated with data using minimum partitioning units. These minimum partitioning units are used as basic cubic elements to perform regular spatial subdivision of the 3D topographic map of the target area. After subdivision, an initial 3D mesh structure is generated, consisting of numerous tightly packed small cubes (initial mesh volumes). Each initial mesh volume automatically obtains its index within the overall structure and the average or center point location information corresponding to its spatial extent.

[0048] For each initial grid cell, based on its spatial location, extract the precipitation observation time series associated with that location from the historical precipitation data, which is traced back a preset duration (e.g., 6 hours) from the current forecast start time, and use it as the grid cell's historical precipitation data.

[0049] Similarly, based on the spatial location of the grid cell, the time series of radar echo intensity covering that location and with the same backtracking duration is extracted from the historical radar observation data and used as the historical radar observation data for that grid cell.

[0050] The extracted historical precipitation data and historical radar observation data of the grid cells are used as attribute values ​​and added to the corresponding initial grid cells. At this point, each grid cell contains its location, historical precipitation sequence, and historical radar echo sequence, thus completing the construction of the three-dimensional grid structure from the initial three-dimensional grid structure to the final target area.

[0051] In one possible embodiment, after dividing the 3D topographic map into a grid using minimum segmentation units, grid cells intersecting with key mountain and valley topographic features are identified within the 3D grid structure and used as topographic grid cells. These key mountain and valley topographic features are areas with strong topographic transitions and significant impacts on airflow and precipitation, such as narrow valley throats, steep windward slopes, and ridgelines. Topographic features are extracted from the topographic grid cells, and the density of the topographic grid cells is adaptively adjusted based on these extracted features.

[0052] Specifically, based on a high-precision DEM, grids that intersect with the mountain and valley terrain are identified as terrain network volumes in the three-dimensional mesh structure. Key terrain features within the coverage area of ​​each terrain grid volume in the three-dimensional mesh structure are calculated. Key terrain features may include: average slope, terrain relief, surface curvature, and surface normal direction. For a terrain grid volume, the density can be adaptively adjusted by a variable density function based on key terrain features such as average slope, terrain relief, surface curvature, and surface normal direction.

[0053] More specifically, based on a high-precision DEM, key topographic feature lines within the target area are pre-identified and vectorized using Geographic Information System (GIS) algorithms. These key topographic feature lines can be ridgelines, valley lines, river centerlines, abrupt slope lines, etc. Among them, ridgelines and valley lines can reflect the topographic framework, river centerlines can reflect the water flow path, and abrupt slope lines can reflect cliffs and plateau edges.

[0054] By traversing each cell in the initial 3D mesh structure, its 3D bounding box is calculated. If the bounding box intersects any of the aforementioned key terrain feature lines, the cell is marked as a terrain grid cell. For each marked terrain grid cell, a set of quantitative key terrain feature vectors T=(S, R, C, θ) is extracted from the high-precision DEM data it covers, where S is the average slope, representing the arithmetic mean of the slopes of all DEM cells within the terrain grid cell; R is the terrain relief, representing the difference between the highest and lowest elevation values ​​within the terrain grid cell; C is the surface curvature, representing the Gaussian curvature or profile curvature at the center point of the terrain grid, used to quantify the concavity and convexity of the terrain, and can be calculated using a second-order differential operator (such as the Zevenbergen-Thorne method based on a 3x3 window); and θ is the surface normal direction, representing the principal curvature direction of the terrain, which is the surface normal angle at the center point of the terrain grid cell, used to indicate the windward slope direction.

[0055] The input to the aforementioned variable density function is the key terrain feature vector T. By normalizing the features of each dimension in the key terrain feature vector T, the potential intensity value of the influence of each terrain grid cell on airflow and precipitation is output. The density of the terrain grid cell is then adjusted according to the intensity value; the higher the intensity value, the higher the density of the terrain grid cell, meaning more grid cells are obtained after secondary segmentation. The aforementioned variable density function is shown below: ; in, The intensity value of the i-th terrain grid cell; Let be the normalized average slope of the i-th terrain grid cell; Let be the average slope of the DEM cell within the i-th terrain grid; This represents the minimum average slope of all grid cells within the entire target area. This represents the maximum average slope of all grid cells within the entire target area; The normalized terrain relief of the i-th terrain grid cell; The topographic relief of the i-th topographic grid cell is the difference between the highest and lowest elevations within the i-th topographic grid cell. This represents the maximum terrain relief within the entire target area; Let be the normalized absolute curvature of the i-th terrain grid volume; The absolute value of the surface curvature at the center point of the i-th terrain grid cell can be used for curvature calculation, which can be done using profile curvature or horizontal curvature. This represents the maximum absolute curvature of the entire target region. Let be the principal curvature direction angle of the i-th terrain grid volume, i.e., the direction with the steepest slope. The prevailing wind direction angle or the current real-time wind direction angle in the target area during a short-term heavy precipitation weather process. The prevailing wind direction angle during a short-term heavy precipitation weather process can be obtained by statistical analysis based on historical weather maps, while the current real-time wind direction angle can be obtained by accessing real-time wind field data. These are the weight coefficients for the corresponding feature terms.

[0056] This can be determined through statistical analysis of radar echo intensity / precipitation intensity and topographic features in historical precipitation events, or through back-optimization based on the grid division effect of training samples. A series of historical precipitation events in different valley and river regions are collected. For each event, the aforementioned topographic features for each local grid cell are calculated. The forecast error for each grid cell is calculated, such as radar echo extrapolation error or quantitative precipitation forecast error. Feature importance analysis tools such as principal component analysis, multiple linear regression, or random forest are used to analyze the explanatory power or contribution of each topographic feature to the forecast error. The relative importance obtained from the analysis is normalized and used as weighting coefficients. The initial value. This can be used in the mesh generation effect evaluation. The initial value is fine-tuned.

[0057] Calculate the intensity values ​​of all terrain grid cells within the target area. Apply K-means clustering or percentile method to naturally divide the intensity values ​​into q clusters. The boundary values ​​between clusters can be used as subdivision thresholds Th1 Th2…Th q Different clusters represent different degrees of influence on airflow and precipitation. Each cluster has a corresponding density division rule, which divides terrain grids that meet the secondary division conditions into smaller grids to improve the density or resolution of the terrain grids. For example, if the intensity value of a terrain grid is within the Th1 cluster, it can be divided without secondary division. If the intensity value is within the Th2 cluster, it can be divided once horizontally and once vertically, thus dividing it into four smaller grids on average. If the intensity value is within the Th3 cluster, it can be divided twice horizontally and twice vertically, thus dividing it into nine smaller grids on average, and so on.

[0058] After density adjustment, the location data of smaller grid cells are updated to the center location of the smaller grid cells. The historical precipitation data and historical radar observation data of the grid cells are inherited from the original grid cells.

[0059] By adaptively adjusting the terrain features within the terrain grid, an adaptive grid partitioning driven by both terrain adaptation and forecast error is formed. This makes the constructed 3D grid structure itself a data container carrying prior knowledge of precipitation physics in high mountain and valley regions. The final constructed 3D grid structure has higher resolution in key terrain areas such as valleys, ridges, and slopes, better matching the spatial heterogeneity of precipitation systems in high mountain and valley regions and improving the accuracy of precipitation forecasting.

[0060] This embodiment analyzes the spatial range and intensity error characteristics between forecasts and actual precipitation events in historical precipitation events, determines a minimum segmentation unit that can finely characterize the spatial details of these errors, and uses this unit to construct the basic grid of the entire forecast area. By adapting it to the specific terrain conditions of the target area, a variable-density three-dimensional grid structure is obtained, which can match the predictability scale of the local precipitation system and complex terrain conditions, thereby improving the accuracy of prediction.

[0061] Optionally, step S13 specifically includes: S131, based on the actual precipitation range and the forecast precipitation range, calculate the range error for each precipitation event, and based on the actual precipitation intensity and the forecast precipitation intensity, calculate the intensity error for each precipitation event.

[0062] S132, perform a remainder operation on the actual precipitation range with the smallest range error and greater than 0, the overlapping range and the non-overlapping range of the forecast precipitation range, and use the remainder as the side length of the smallest segmentation unit to construct the initial segmentation unit of the cube.

[0063] S133, the adjustment factor is determined based on the average of the intensity errors during multiple precipitation events.

[0064] S134, the initial segmentation unit is adjusted by the adjustment factor to obtain the smallest segmentation unit.

[0065] In this embodiment of the invention, for a certain precipitation event, the actual precipitation range (G) and the forecast precipitation range (F) can be represented as a polygonal region on a two-dimensional plane or a binary raster image (1 indicates precipitation, 0 indicates no precipitation).

[0066] The range error can be calculated as (the area of ​​the union of F and G - the area of ​​the intersection of F and G) / the area of ​​G, yielding a relative error. Alternatively, it can be calculated as the average of the shortest distances from points on the boundary of F to the boundary of G. The resulting scalar value characterizing the accuracy of the spatial location prediction for this event is the range error. A smaller range error indicates a more accurate spatial prediction.

[0067] For the same precipitation event, the actual precipitation intensity (I_G) and the forecast precipitation intensity (I_F) are considered. Precipitation intensity represents a scalar value of the intensity of the precipitation event, such as the spatial average intensity or the maximum intensity. The intensity error can be an absolute error |I_F-I_G| or a relative error |I_F - I_G| / I_G.

[0068] Iterate through all historical precipitation events and find the one with the smallest range error value that is greater than 0. The criterion of greater than 0 is to exclude the random case of perfect forecasts (error of 0) and ensure that the selected event is representative and has subtle spatial bias.

[0069] For a selected baseline event, the spatial relationship between its actual precipitation range (G) and forecast precipitation range (F) is analyzed. Spatially, there will be one overlapping range and one or more non-overlapping ranges. The overlapping range is O = G∩F, which is the spatial area where the forecast is correct. The non-overlapping ranges are the parts of G not covered by F, and the parts of F that extend beyond G.

[0070] Calculate the equivalent diameter or length scale along the principal direction for the overlapping range O and each non-overlapping range. Treat the length scales as a set of values ​​and perform a modulo operation. Specifically, these length values ​​can be paired, their absolute differences calculated, and then the average of all differences taken; or, one length can be used as the dividend, the other as the divisor, and the remainder taken as the base length value. Use the base length values ​​obtained in the previous step as the side lengths of the cube to define an initial segmentation unit for the cube. This cube represents the basic volume element derived from the best historical spatial predictions, reflecting its minimum spatial deviation characteristics.

[0071] For all historical precipitation events, the mean intensity error is calculated. This mean reflects the average deviation level of forecast intensity under historical forecasting methods in this region. The corresponding adjustment factor can be found in the forecast mapping table based on the mean. The mapping table records the mapping relationship between the mean and the adjustment factor; that is, each mean corresponds to one adjustment factor.

[0072] In one possible implementation, the adjustment factor is a scaling factor mapped or calculated based on the average intensity error. For example, a baseline intensity error value is set; when the mean intensity error is greater than this baseline, μ > 1 (the mesh needs to be enlarged to reduce sensitivity to spurious strong signals); when the mean intensity error is less than this baseline, μ < 1 (the mesh can be reduced to capture finer structures). The specific functional relationship of μ can be determined empirically or through optimization, for example: μ = 1 + β (mean of intensity error), where β is a preset sensitivity coefficient.

[0073] The final determined grid scale not only considers subtle deviations in spatial positioning but also incorporates the average intensity deviation level of the forecast system, making the grid division more comprehensive and adaptable. A two-stage strategy—selecting the best spatial forecast events to analyze their deviation geometric characteristics and using the average intensity error for global calibration—transforms forecast performance data into a physical spatial scale. This frees the determination of the minimum subdivision unit from subjective experience, giving it data-driven and adaptive characteristics. Consequently, the final constructed 3D grid structure can more scientifically and rationally match the predictability characteristics of historical precipitation in the target area, thereby improving the accuracy of short-term forecasts.

[0074] Optionally, the method for short-term nowcasting of precipitation in the dam area of ​​hydropower stations in high-altitude valleys based on deep learning multi-model fusion also includes step S5, which trains multiple time-series prediction models. Step S5 includes: S51, obtain the first training sample set and multiple time series models to be trained.

[0075] The first training sample set includes a first sample, a second sample, a real precipitation label, and a first radar echo extrapolation label. The first sample is obtained by modifying the historical precipitation data of the grid cell and the historical radar observation data of the grid cell according to the three-dimensional grid structure of the target area. The second sample is obtained by extrapolating the first sample. The real precipitation label is short-time real precipitation data obtained from historical precipitation data, and the first radar echo extrapolation label is short-time radar echo extrapolation data obtained from historical radar observation data.

[0076] S52, train multiple time series models to be trained using the first training sample set to obtain multiple trained time series prediction models.

[0077] In this embodiment of the invention, the time-series model to be trained can be a deep learning network model that has not yet been trained and whose parameters are in a randomly initialized state, or it can be a deep learning network model that has undergone preliminary training. The time-series model to be trained can be a deep learning model capable of handling time-series problems, such as CNN-Attention, Conv-LSTM, or PredRNN++. The time-series model to be trained is designed to process spatiotemporal sequence data, and its network structure typically includes modules such as convolution, recurrent, or attention, enabling it to extract features from historical spatiotemporal meteorological data sequences and predict future sequences.

[0078] Each sample in the first training sample set is a data package containing the input required for model training and the corresponding correct answer (label). The first sample serves as the primary input data for the time-series model to be trained. It is obtained by modifying, formatting, or reconstructing the historical precipitation data and historical radar observation data of the grid cells within the target region's 3D grid structure. These modifications primarily include data standardization, time-series sliding window sampling, and feature stitching preprocessing operations, aiming to create data blocks that conform to the model's input format requirements. It should be noted that while the first sample has the same structure as the 3D grid structure, the historical precipitation data and historical radar observation data of each grid cell differ in both time and numerical value.

[0079] The second sample is obtained from the first sample through some kind of extrapolation. Extrapolation is a spatiotemporal estimation method, such as motion field estimation and translation extrapolation based on optical flow, or preliminary prediction based on linear trends. Its purpose is to provide a rough prior or guiding signal of future states for the time series model to be trained.

[0080] The actual precipitation label is the actual, future surface precipitation data corresponding to the historical precipitation events in the first sample. It is obtained directly from historical precipitation data. It corresponds to a short future time period (e.g., 0-2 hours in the future) after the historical time period represented by the first sample, and is a gridded precipitation field generated through spatial interpolation.

[0081] The aforementioned first radar echo extrapolation label serves as the intermediate prediction target that the time-series model to be trained needs to learn, i.e., real, future radar echo data. This can be obtained from historical radar observation data. The future radar echo observation data corresponding to the same time period as the real precipitation label has been resampled to a format matching the three-dimensional grid structure.

[0082] After constructing or loading the prepared first training sample set, multiple time-series models to be trained are initialized. The goal is to adjust the model parameters through optimization algorithms so that the model's predicted output is as close as possible to the label. The first and second samples are input into the currently trained time-series model. Internally, the model performs feature extraction, fusion, and transformation on the input data through its network layers.

[0083] During training, the time series model to be trained will use two supervisory signals: the first radar echo extrapolation label and the actual precipitation label. The goal is to make its prediction of the future radar echo close to the first radar echo extrapolation label and its prediction of precipitation close to the actual precipitation label.

[0084] The backpropagation algorithm is used to calculate the combined error between the model's prediction and the two labels, and the model's weight parameters are updated based on this error. This process is iterated multiple times on all samples.

[0085] Once the training meets the preset stopping conditions (such as reaching the maximum number of iterations or the loss no longer decreasing significantly), the current time series model to be trained is transformed into a trained time series prediction model. This independent training process is repeated for each model to be trained, eventually resulting in multiple trained time series prediction models. It should be noted that the trained time series prediction model removes the input structure of the second sample but retains the prior parameters of the future state learned from the second sample, which are used to fuse with the features of the 3D mesh structure.

[0086] This embodiment uses the true value of the future state of intermediate radar echoes as an additional supervisory signal to guide the time series prediction model to learn the mapping rules from historical data to future radar echoes and the statistical relationship from relevant features to ground precipitation during the training process. This enables the trained model to have more powerful feature extraction and prediction capabilities, and improves the prediction accuracy for three-dimensional grid structures.

[0087] Optionally, step S52 specifically includes: S521, For each time series model to be trained, the first sample and the second sample are input into the time series model to be trained, and the features of the first sample and the features of the second sample are extracted.

[0088] S522. The features of the first sample and the features of the second sample are fused to obtain the fused features.

[0089] S523 performs regression processing on the fusion features to obtain extrapolation prediction results and precipitation prediction results.

[0090] S524, calculate the first error value between the extrapolated prediction result and the first radar echo extrapolation tag, and calculate the second error value between the precipitation prediction result and the actual precipitation tag.

[0091] S525 uses minimizing the total error of the first and second error values ​​as the optimization objective to adjust the parameters of the time series model to be trained.

[0092] S526, iterate through the parameter adjustment process of the time series model to be trained until the error value is less than the preset error threshold or the number of iterations reaches the number threshold, then stop training and obtain the trained time series prediction model.

[0093] In this embodiment of the invention, the first sample feature and the second sample feature are abstract feature representations extracted by the time-series model to be trained after processing the first sample and the second sample through convolutional layers, respectively. It can be understood that the first sample feature and the second sample feature are no longer the original observation data, but rather high-dimensional tensors that have undergone nonlinear transformation by the neural network and contain the spatiotemporal patterns understood by the model.

[0094] The unified feature representation formed by combining the features of the first sample and the features of the second sample through methods such as concatenation, addition, and attention weighting is thus obtained as the fused feature. The purpose of the fused feature is to integrate historical state information with preliminary extrapolation guidance information.

[0095] The extrapolation prediction result is the predicted output of the time series model to be trained on the future radar echo state. It is a gridded data field, such as the radar reflectivity sequence at multiple future moments.

[0096] The precipitation prediction results are the output of the time series model being trained, representing a gridded quantitative precipitation forecast field. It should be noted that the relevant output structure of the precipitation prediction results can be deleted after training is complete.

[0097] The first and second error values ​​quantify the scalar loss values ​​representing the differences between the extrapolated prediction results and the extrapolated labels of the first radar echo, and the differences between the precipitation prediction results and the actual precipitation labels, respectively. These values ​​can be calculated using a preset loss function. The loss function used for radar echoes can be mean squared error (MSE) or smoothed L1 loss; the loss function used for precipitation can be MSE, mean absolute error (MAE), or a weighted loss considering precipitation imbalance.

[0098] The total error value is the ultimate optimization objective guiding the adjustment of model parameters. It can be a weighted sum or combination of the first and second error values.

[0099] Specifically, a training sample (containing a first sample and a second sample) taken from the first training sample set is input into the time series model currently being trained. The feature extraction network within the time series model processes the first and second samples separately, obtaining first and second sample features representing its deep spatiotemporal pattern. The first and second sample features are then fused into a unified fused feature through a fusion layer within the time series model. This allows the model to learn how to combine historical context with preliminary extrapolation clues.

[0100] Based on fused features, the time series model to be trained simultaneously performs two prediction tasks, each with an output branch. One output branch converts the fused features into extrapolated predictions for future time periods. The other output branch converts the fused features into precipitation predictions for the same future time period. This allows the time series model to simultaneously decode echo data describing atmospheric conditions and precipitation data describing surface effects from the fused features. By approximating the precipitation data with the actual precipitation labels, the extrapolated predictions are further forced to approximate the radar echo extrapolated labels.

[0101] Using the two labels inherent to the samples, the accuracy of the model's two predictions is evaluated. A loss function suitable for regression tasks (such as MSE) can be used to calculate the difference between the extrapolated prediction and the first radar echo extrapolated label. A different loss function (such as MAE) is used to calculate the difference between the precipitation prediction and the actual precipitation label. The first and second error values ​​are then combined into a total error value through a weighted summation. The total error value represents the overall poor performance of the model on the current samples.

[0102] Minimizing the total error is the optimization objective for this iteration and the entire training process. Backpropagation and parameter tuning are performed, and the gradient of the total error with respect to all trainable parameters within the model is calculated using automatic differentiation. Then, optimization algorithms (such as Adam and SGD) are used to make minor adjustments to the model parameters along the gradient descent direction.

[0103] Repeating the above steps constitutes one complete training iteration. During training, all samples in the first training sample set will be traversed and iterated for one epoch.

[0104] Training of the time series model ceases when its performance on the validation set no longer improves, the total error value stabilizes, or the preset maximum number of iterations is reached. At this point, it becomes a trained time series prediction model. The trained time series prediction model will then have the second sample input network and the output branch of the precipitation prediction result removed.

[0105] This embodiment requires the model to simultaneously output and optimize radar echo predictions and precipitation predictions, which helps the time series model learn fusion features that are sensitive to both precipitation and its radar representation and are more physically interpretable, thereby potentially improving the model's generalization ability and the consistency of the final forecast.

[0106] Optionally, the short-term precipitation forecasting method for hydropower station dam areas in high-altitude valleys based on deep learning multi-model fusion also includes step S6 of training the deep learning model. Step S6 specifically includes: S61, obtain the second training sample set and the deep learning model to be trained.

[0107] The second training sample set includes the third sample and the second radar echo extrapolation label. The third sample is obtained by modifying the historical precipitation data of the grid body and the historical radar observation data of the grid body according to the three-dimensional grid structure of the target area. The second radar echo extrapolation label is the short-time radar echo extrapolation data obtained from the historical radar observation data.

[0108] S62, train the deep learning model to be trained based on multiple trained time series prediction models and the second training sample set to obtain a trained deep learning model.

[0109] In this embodiment of the invention, the deep learning model to be trained is a deep learning network architecture for fusing multiple prediction results. Unlike temporal prediction models, the deep learning model to be trained can accept multiple parallel inputs, each input being a prediction result from a temporal prediction model. Its structure includes a feature alignment layer, an attention layer, and a fusion layer. Its parameters are randomly initialized at the start of training.

[0110] The second training sample set can be understood as a dataset specifically used to train the aforementioned deep learning model. Its sample composition differs from that of the first training sample set, and it aims to simulate radar echo extrapolation data from the fusion model.

[0111] The second training sample set includes the third sample and the corresponding second radar echo extrapolation label. The third sample is used to drive the pre-trained time-series prediction model to generate preliminary prediction input data, and the third sample corresponds to the radar echo extrapolation prediction result. The third sample is also obtained by modifying and formatting the historical data in the three-dimensional grid structure of the target area. Its specific form is similar to the first sample, but the time and data corresponding to the historical precipitation data and historical radar observation data of the grid cells are different.

[0112] The second radar echo extrapolation label serves as the supervised target during the training of the deep learning model, representing real, future radar echo data. It's important to note that the direct optimization objective of the deep learning model is the accuracy of the radar echo fusion prediction, not the final precipitation. The second radar echo extrapolation label is obtained from historical radar observation data, corresponding to the real radar echo observation field in a short-term future period following the historical period represented by the third sample. This is physically similar to the first radar echo extrapolation label, but corresponds to a different time and is used for different training stages and models.

[0113] The role of multiple trained time-series prediction models is to independently generate a radar echo prediction sequence based on a third sample. During the training of the fusion model, the parameters of the trained time-series prediction models are fixed and no longer updated.

[0114] Obtain multiple pre-trained time-series prediction models with fixed parameters. These models already possess the ability to predict future radar echoes from historical data. Create a new deep learning model to be trained, with a network structure designed to accept the prediction results from the K time-series prediction models as input.

[0115] For a sample in the second training sample set, its third sample is simultaneously input into each trained time-series prediction model. Each time-series prediction model runs independently and outputs its radar echo prediction result for the same future time period. Assuming there are K time-series prediction models, K radar echo prediction results {P1, P2, ..., P...} are obtained. K}

[0116] The K radar echo prediction results are organized into a tensor suitable for the input of the fusion model by stacking, forming a multi-channel feature map with shape (K, time step, height, width).

[0117] The stacked multi-model prediction tensor is input into the deep learning model to be trained. The deep learning model to be trained learns the correlation, differences and reliability of these predictions through its network layers (including convolutional layers, attention layers, etc.) and outputs a single, fused radar echo prediction result.

[0118] The difference between the fused radar echo prediction and the extrapolated label of the true second radar echo is calculated, and the loss value for this training is obtained using a loss function (such as mean squared error). The gradient of this loss value with respect to all parameters of the deep learning model to be trained is calculated using the backpropagation algorithm. Subsequently, the parameters of the fused model are updated using the gradient descent optimization algorithm. It should be noted that in this backpropagation, the gradients of the parameters of the multiple time-series prediction models acting as feature generators are set not to be updated; that is, the parameters of the multiple time-series prediction models remain unchanged.

[0119] Repeat the processing and calculation for all samples in the second training sample set to complete one training cycle. Repeat this process for multiple cycles until the performance of the fusion model stabilizes (converges) on the validation set. The result at this point is the trained deep learning model.

[0120] This embodiment utilizes a fixed set of diverse time-series prediction models to generate preliminary predictions from multiple perspectives. Then, a dedicated deep learning fusion model is trained to learn how to optimally integrate these preliminary predictions, making its output closer to reality. Decoupling the tasks of generating diverse predictions and intelligent fusion allows the fusion model to focus on learning the complementarity and reliability between different base predictors, effectively reducing training complexity and improving fusion performance, thereby further enhancing prediction accuracy.

[0121] Optionally, step S62 specifically includes: S621, the third sample is input into multiple trained time series prediction models respectively, and the sample prediction results output by each time series prediction model are obtained.

[0122] S622, the sample prediction results output by each time series prediction model are concatenated to form a multi-model fusion feature map.

[0123] S623, input the multi-model fusion feature map into the deep learning model to be trained.

[0124] S624 uses an attention mechanism to dynamically weight and fuse channel-dimensional features from different time-series prediction models in the multi-model fusion feature map, in order to learn and highlight the model features that contribute more to the final forecast in each time-series prediction model.

[0125] S625, based on the weighted fusion of sample features, outputs the final fusion prediction result through the regression layer of the deep learning model to be trained.

[0126] S626, calculate the fusion loss based on the difference between the final fusion prediction result and the extrapolated tag of the second radar echo.

[0127] S627 aims to minimize the fusion loss by adjusting the parameters of the deep learning model to be trained, iterating the parameter adjustment process of the time series model to be trained until the model converges, and obtaining the trained time series prediction model.

[0128] In this embodiment of the invention, the sample prediction result is the prediction result of radar echo extrapolation within a specific future time period output by each trained time-series prediction model after receiving a third sample as input. The specific time period can be any time period within 0-2 hours and can be set by the user. Each time-series prediction model generates one prediction result.

[0129] The multi-model fusion feature map is a multi-dimensional data tensor formed by stacking or splicing multiple sample prediction results along the channel dimension. Its shape is (K, B, T, H, W) or similar, where N represents the number of time-series prediction models, B represents the number of channels of each sample prediction result (such as different prediction times), and T, H, and W represent the time, spatial height, and width dimensions, respectively.

[0130] Attention mechanisms are computational modules embedded in neural networks that enable the network to dynamically focus on more important parts of the input data. In this scheme, the attention mechanism specifically refers to the channel attention mechanism, which assigns different importance weights to channels from different temporal prediction models in the multi-model fusion feature map.

[0131] The feature representation obtained after multi-model fusion feature maps are processed by an attention mechanism. In this process, the predicted features from different models are multiplied by their respective learned weight coefficients, with important features being enhanced and minor features being suppressed.

[0132] A regression layer is one or more neural network layers located at the end of the network in a deep learning model to be trained. These layers are used to map high-dimensional features back to the target prediction space (radar echo field), such as fully connected layers or convolutional layers, and can also be called output layers.

[0133] The fusion loss is used to quantify the difference between the final fused prediction result and the extrapolated label of the actual observed second radar echo. It can be calculated using a preset loss function (such as mean square error (MSE), smoothing L1 loss, etc.).

[0134] Specifically, for the k-th time-series prediction model (k=1, 2, ..., K), the model reads the third sample as input data, calculates and outputs the prediction result of the future radar echo required for the third sample, denoted as the sample prediction result P. k We obtain prediction results for K independent samples {P1, P2, ..., P}. K}

[0135] Let {P1, P2, ..., P K Perform channel concatenation. For example, assuming each Pj is a tensor of shape (B, T, H, W), concatenate them along the first dimension (model dimension) to form a tensor of shape (K). A multi-model fusion feature map (B, T, H, W). The multi-model fusion feature map encapsulates the sample prediction results of all time-series prediction models for the third sample.

[0136] An attention mechanism automatically analyzes information from different models (different channels) in the feature map. The attention mechanism first generates a weight vector representing the importance of each channel. Then, based on the learned pattern, a weight value between 0 and 1 is calculated for each channel. Channels with higher weight values ​​are considered more reliable. The calculated weight vector is multiplied channel-by-channel with the multi-model fusion feature map to obtain the weighted fused sample features. At this point, features from models that contribute significantly are amplified, while those that contribute less are weakened. The weighted fused sample features are then passed to the regression layer of the fusion model. The regression layer performs regression processing on the weighted sample features and maps them back to physical space, outputting a final fusion prediction result with the exact same dimension as the extrapolated label of the second radar echo.

[0137] The pre-defined loss function is called to calculate the difference between the final fused prediction result and the extrapolated label of the second radar echo, thus obtaining the fusion loss. This fusion loss reflects the prediction error of the current fusion model.

[0138] Automatic differentiation is used to calculate the gradient of the fusion loss across all trainable parameters of the deep learning model to be trained. It's important to note that during the training process, the parameters of multiple time-series prediction models are treated as constants and do not participate in gradient calculation or updates. The optimizer updates the parameters of the fusion model using backpropagation based on the calculated gradients. This process iterates through all samples in the second training dataset multiple times until the fusion model's performance on the validation set converges or the preset number of iterations is reached, ultimately yielding the trained deep learning model.

[0139] This embodiment enables deep learning models to do more than simply average or vote on multiple prediction results. Instead, it allows them to dynamically and adaptively evaluate the reliability of different time-series prediction models under different weather scenarios and assign appropriate weights to each time-series prediction, thereby improving the fusion effect of the fused extrapolation prediction results and increasing forecast accuracy.

[0140] Optionally, step S624 specifically includes: S6241 performs global average pooling and global max pooling on the multi-model fusion feature map in the spatial dimension to obtain the first pooling vector and the second pooling vector.

[0141] S6242, the first pooling vector and the second pooling vector are added together after multilayer perceptron processing, and attention weights for each channel are generated through activation function.

[0142] S6243 multiplies the generated attention weights with the multi-model fusion feature map channel by channel to obtain the weighted fusion sample features.

[0143] In this embodiment of the invention, for each channel of the multi-model fusion feature map, the arithmetic mean of the feature values ​​at all spatial locations (in the height H and width W dimensions) within that channel is calculated. This arithmetic mean is used to reflect the global average intensity of the corresponding channel's features. The arithmetic mean of each channel is normalized into a one-dimensional vector, namely the first pooling vector. The length of the first pooling vector is equal to the number of channels B of the input feature map, and each element corresponds to the average intensity of one channel.

[0144] For each channel of the multi-model fusion feature map, find the maximum value of the feature values ​​at all spatial locations within that channel. This maximum value reflects the most salient response of the channel's features. Normalize the maximum values ​​of each channel into a one-dimensional vector, i.e., the second pooling vector. Its length is also B, and each element corresponds to the maximum value of a channel.

[0145] This is used to input the first and second pooling vectors into a shared multilayer perceptron. The multilayer perceptron consists of two fully connected layers with a ReLU activation function in between. Processing the first and second pooling vectors using the same multilayer perceptron ensures consistency in processing. After processing by the multilayer perceptron, the two input vectors are transformed into feature representations in the same semantic space.

[0146] The corresponding elements of the first and second pooling vectors after multilayer perceptron processing are added together to obtain an intermediate vector that integrates average and saliency information. This added vector is then passed through a sigmoid function to map the values ​​to between 0 and 1, ultimately outputting a one-dimensional vector of length B, which represents the attention weights W for each channel. at The closer the weight value is to 1, the more important the corresponding channel's features are in the current fusion decision.

[0147] In one possible embodiment, during the training phase of each time series prediction model, the training performance data of each time series prediction model is recorded. Based on the training performance data of each time series prediction model, the performance vector of that time series prediction model is extracted as a prior feature. The attention weight of the corresponding channel of the time series prediction model in the multi-model fusion graph is adjusted by the performance vector.

[0148] Specifically, during the training of multiple time-series prediction models, a model training performance memory is constructed simultaneously to quantitatively record the training performance data of each model, including its learning characteristics and ability biases. Gradient sensitivity feature vectors, model performance profiles, and learning stability indicators are extracted from the training performance data.

[0149] Among them, the gradient sensitivity feature vector is used to indicate that the temporal model converges faster in a certain direction during training. It can periodically save the gradient g of the loss function with respect to the parameters of the key layer in the k-th model over multiple epochs of model training. k By performing principal component analysis (PCA) on the multiple saved gradients, the first principal component vector p with the largest gradient contribution is extracted. k The first principal component vector p k This represents the most sensitive and active update direction of the model parameters, implying the precipitation evolution pattern that the model is most focused on learning, such as rapid movement, rapid intensification, or topographic retention.

[0150] Model performance profiles are used to describe the functionality of time-series forecasting models. After training, a new sample set can be collected, divided into multiple subsets based on weather type (e.g., convection, stratiform, mixed) and terrain type (e.g., steep slope, valley, plain).

[0151] For the k-th time series prediction model Mk Record its value in each subset D i Mean Absolute Error (MAE) k,i Critical Success Index (CSI) k,i This forms a model representation vector a used to describe the model's functionality. k =[CSI k,1 MAE k,1 CSI k,2 MAE k,2 ,...] .

[0152] The learning stability metric is used to indicate whether a model is stable. It can be calculated by measuring the standard deviation σ of the model's loss on the validation set in the later stages of training. k， σ k The smaller the value, the more stable the model convergence and the more reliable its prediction results.

[0153] For each time series prediction model, the gradient sensitivity feature vector p k、 Model representation of portrait a k and the learning stability index σ k After concatenation, the values ​​are mapped to 0 and 1 using the Sigmoid function, resulting in a one-dimensional representation vector v of length B. k。

[0154] The representation vectors of N time series models are used to construct a representation matrix V = [v1, v2, v3, ..., v2, v3, ..., v4, v5, v6, v7, v8, v9, v1, v1, v1, v2, v3 ... k The attention weight W is calculated using the following formula. at The attention weight vector is obtained by fusing it with the performance matrix V. : ; in, The fusion coefficient can be dynamically set according to the current meteorological scenario. It can calculate the prior confidence of the multi-model fusion feature map and set the fusion coefficient based on the prior confidence. Different prior confidence corresponds to different fusion coefficients. The prior confidence is the average confidence of the short-term extrapolation prediction results of each time series prediction model.

[0155] attention weight vector (Shape [B]) and the input multi-model fusion feature map (shape [K]) Element-wise multiplication is performed on the attention weight vector (B, T, H, W). Specifically, the j-th scalar value of the attention weight vector is multiplied by the feature values ​​of all spatial locations in the j-th channel of the feature map. After element-wise multiplication, the weighted fused sample features are obtained. At this point, the features of each channel in the original feature map are rescaled according to their importance. Important features are enhanced, while minor or irrelevant features are suppressed.

[0156] By incorporating the capability biases and learning characteristics exhibited by the time-series prediction model during training as prior knowledge into real-time prediction, and relying more on performance vectors to adjust attention weights when model prediction uncertainty is high, the deep learning model can learn and highlight model features that contribute more to the final prediction from each time-series prediction model. When predictions are consistent, the attention weights obtained from the current pooling vector are relied upon more, allowing the deep learning model to focus more on the impact of the current scene on the final prediction. Pooling features are used to drive attention to capture the instantaneous patterns of the current scene, while performance vectors provide stable guidance based on historical experience. The weighted fusion of attention weights and performance matrices balances flexibility and stability.

[0157] Attention weights are multiplied channel by channel with the feature map, which enables dynamic recalibration of prediction results from different models. For example, in severe convective weather, the prediction results of robust time series prediction models are automatically weighted up during the fusion process; in orographic precipitation, the prediction results of time series prediction models that are more sensitive to terrain response are weighted up during the fusion process.

[0158] This embodiment comprehensively captures the statistical characteristics of each feature channel through dual-path aggregation using global average pooling and global max pooling. Subsequently, a shared multilayer perceptron learns the nonlinear dependencies between channels and generates weights. Finally, feature recalibration is achieved through channel-by-channel scaling. This enables the deep learning fusion model to dynamically and discriminatively integrate prediction information from different base models based on the specific content of the input data, achieving high-precision feature fusion and further improving prediction accuracy.

[0159] Optionally, step S4 specifically includes: S41 integrates the extrapolated prediction results with precipitation data observed by real-time automatic weather stations and inputs them into a pre-trained precipitation conversion model.

[0160] S42 uses a precipitation conversion model to transform the fused extrapolation prediction results into a gridded quantitative precipitation forecast for the future time period, serving as a short-term nowcast for the target area.

[0161] In this embodiment of the invention, the fusion extrapolation prediction result is an optimized future radar echo intensity prediction field generated by intelligently fusing the outputs of multiple time-series prediction models using a deep learning fusion model. The fusion extrapolation prediction result is a multi-dimensional data grid that characterizes the predicted radar reflectivity of the target area on each grid cell within a future time period (e.g., 0-2 hours).

[0162] Precipitation data observed by real-time automatic weather stations can be surface precipitation observation data collected and transmitted in real time by automatic weather stations deployed in and around the target area at the start of the forecast. It is usually the cumulative precipitation (unit: mm) per minute or hour. Precipitation data observed by real-time automatic weather stations represents the current true state of surface precipitation and has high point-scale accuracy and real-time performance.

[0163] The precipitation conversion model is a machine learning or deep learning model. It establishes a complex mapping relationship between radar echo extrapolated data and precipitation data observed by real-time automatic weather stations to surface precipitation. The precipitation conversion model can be trained using radar echo extrapolated data and precipitation data observed by real-time automatic weather stations as sample data, thereby enabling the model to establish this complex mapping relationship.

[0164] Historically matched radar echo observation data and automatic weather station precipitation observation data are used as input features, while real precipitation data are used as training labels. The precipitation conversion model can be based on random forest, gradient boosting tree, neural network, or physical-statistical joint model (ZR relation calibration model), etc.

[0165] Gridded quantitative precipitation forecasting can be understood as a forecast covering the entire target area with a regular spatial grid, where each grid cell contains the predicted value (in mm) of cumulative precipitation within a specific "future time period" (such as the 1st hour, 2nd hour, etc.). Gridded quantitative precipitation forecasting is spatially discrete and can be directly used for high-precision applications such as flood risk mapping and engineering scheduling.

[0166] The system acquires the fusion extrapolation prediction results (future radar echo prediction sequence) output in real time from the fusion module. Simultaneously, it retrieves the latest real-time precipitation data from automatic weather stations from the data center. Necessary quality control and format standardization are performed on the real-time precipitation data. The station data is correlated with the spatial location of the forecast grid. The fusion extrapolation prediction results serve as the main input features, while the real-time precipitation data serves as an important conditional input or calibration basis, used to capture the actual intensity and spatial distribution of current precipitation to correct potential systematic biases or initial value errors in radar predictions.

[0167] This is achieved by jointly constructing the input tensor for the precipitation transformation model. For example, the input might include: the predicted radar echo field at a future time, the radar observation echo field at the current time, and the analysis field obtained by interpolating precipitation from the current station.

[0168] The transformation model internally processes the input features based on the complex nonlinear relationships it has learned. For example, the transformation model learns that at a specific terrain location (implicit in the grid coordinates), when the radar echo prediction exceeds a certain threshold and there is currently weak precipitation on the ground, the hourly precipitation in the future is likely to increase. After integrating all features, the model calculates a precipitation prediction value for each grid point in the output grid.

[0169] The output of the transformed model is a gridded quantitative precipitation forecast. This forecast covers the target area, with a temporal resolution consistent with the fusion extrapolation prediction results (e.g., every 10 minutes), and a forecast duration covering future time periods (e.g., 0-2 hours). This gridded quantitative precipitation forecast is marked as a short-term nowcast precipitation forecast and distributed to user terminals such as those for hydropower station operation and management.

[0170] This embodiment combines the advantages of deep learning in spatiotemporal prediction with the advantages of statistical models in modeling variable relationships. It utilizes the high spatiotemporal resolution of radar data while ensuring the absolute accuracy of precipitation estimation through ground observations. It can output highly practical quantitative precipitation forecasts that directly support disaster prevention and mitigation decisions, completing the final step from data to decision-making.

[0171] Optionally, the short-term nowcasting method for hydropower station dam areas in high-altitude valleys based on deep learning multi-model fusion further includes an evaluation and verification step S7, used to evaluate the forecast effectiveness after the forecast is generated; the evaluation and verification step S7 includes: S71, obtains measured precipitation data from ground meteorological stations in the target area during the forecast period.

[0172] S72, compare the short-term precipitation forecast with the measured precipitation data, and calculate a preset evaluation index value.

[0173] The evaluation indicators include at least one of the following: structural similarity, correlation coefficient, mean absolute error, root mean square error, hit rate, false alarm rate, critical success index, accuracy, and Heidegger skill score.

[0174] In this embodiment of the invention, after waiting for the forecast period to end, measured precipitation data (G) from all available meteorological stations in the target area during that period is extracted from the observation database. Short-term nowcast precipitation forecasts (R) for the same period and the same area are retrieved from the forecast product library.

[0175] Since forecasts are gridded data while observations are station-based data, spatial matching is required for point-by-point or region-by-region comparisons. The location of each meteorological station is mapped onto the forecast grid, identifying the grid cell containing that station. The forecast value (R) of that grid cell is paired with the observed value (G) of that station to form a "station-grid" data pair. Paired data from all stations are collected to form a dataset for evaluation. For areal rainfall assessment, a comprehensive comparison of the forecast field and the analysis field obtained through station interpolation may be necessary. Indicators can be categorized into three types: continuous variable indicators, categorical variable indicators, and comprehensive skill scores.

[0176] Among them, continuous variable indicators directly assess the error of precipitation values, such as mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (CC).

[0177] Categorical variables classify precipitation events as "occurred" or "not occurred" (e.g., setting a precipitation threshold, such as 0.1 mm / hour) to assess forecast hit and false alarm rates, such as probability of occurrence (POD), false alarm rate (FAR), critical success index (CSI), and accuracy (ACC). These indicators are calculated based on a binary contingency table as shown in Table 1.

[0178] Table 1 ; Integrated skill score: measures the improvement of a forecast relative to a simple reference forecast (such as a climate mean or persistent forecast), such as the Heidegger skill score (HSS).

[0179] For all N valid data pairs, calculate the square of the difference between the predicted and observed values ​​for each point. Sum the squared differences of all points and divide by the total number of pairs N to obtain the average. Finally, take the square root of this average to obtain the RMSE value. The smaller this value, the smaller the overall forecast bias.

[0180] Taking the Critical Success Index (CSI) as an example, firstly, based on a set precipitation threshold, the data from all stations and their corresponding grid points are converted into two categories: "with precipitation" and "without precipitation." Then, the following are counted: the number of stations that predicted and actually received rain (TP), the number of stations that predicted rain but actually received no rain (FP), and the number of stations that predicted no rain but actually received rain (FN). Finally, the CSI is calculated using the formula CSI = TP / (TP + FP + FN). The closer the CSI is to 1, the better the overall forecasting capability for precipitation events.

[0181] To objectively and comprehensively evaluate the consistency, bias, and estimation performance between radar quantitative precipitation inversion values ​​(R) and precipitation observed by ground meteorological stations (G), nine commonly used evaluation indicators were selected.

[0182] Structural similarity (SSIM) is a metric that assesses image quality by comparing the brightness, contrast, and structural information of images. ;

[0183] The correlation coefficient (CC) reflects the degree of correlation between radar precipitation retrieval values ​​and ground rain gauge observations. A value closer to 1 indicates a higher correlation and better estimation results. and These represent the average precipitation values ​​(pixel average values ​​of the image) for the forecast field x and the actual field y, respectively. This represents the covariance between the forecast field x and the actual field y. This represents the variance of the predicted field x. This represents the variance of the actual field y. and It is a pre-defined constant less than 1.

[0184] ; In the formula, N For sample data, These are radar precipitation inversion values. The average hourly precipitation (mm) retrieved from radar precipitation inversion. These are rain gauge observations. The average precipitation (mm) over 1 hour is the data observed by ground automatic weather stations.

[0185] Mean Absolute Error (MAE) represents the average absolute error between radar precipitation retrieval values ​​and rain gauge observations. A lower MAE value indicates a lower error between the radar precipitation retrieval and the rain gauge readings. .

[0186] The root mean square error (RMSE) is used to measure the value retrieved from radar precipitation. Same rain gauge observation value Deviation between: .

[0187] The following Probability of Detection (POD), Failure Rate (FAR), Critical Success Index (CSI), Accuracy (ACC), and Heidegger Skill Score (HSS) are calculated based on a binary contingency table (as shown in Table 1). Probability of Detection (POD) represents the proportion of predicted actual precipitation areas out of all actual precipitation areas:

[0188] .

[0189] False Alarm Rate (FAR) represents the proportion of areas in the forecast precipitation area that actually receive no precipitation out of the total forecast precipitation area.

[0190] .

[0191] The Critical Success Index (CSI) comprehensively considers the three scenarios of hits, misses, and false alarms, and is one of the most commonly used indicators for evaluating the effectiveness of precipitation forecasts. The ideal CSI value is 1; a higher value indicates better accuracy and reliability of the forecast.

[0192] .

[0193] Accuracy (ACC) represents the proportion of samples (including both precipitation and no precipitation) that the model correctly predicted out of the total number of samples.

[0194] .

[0195] The Heidke Skill Score (HSS) represents the forecast accuracy after removing random effects. .

[0196] This embodiment objectively quantifies forecasting capabilities in complex terrain and comprehensively diagnoses the sources of forecast errors. For example, it analyzes MAE and RMSE to understand the overall error level, POD and FAR to understand the ability to capture precipitation events and the occurrence of false alarms, and HSS to understand the model's forecasting skills. It can provide hydropower station operation and management personnel with a quantitative reference on forecast reliability, assisting them in risk assessment and scheduling decisions.

[0197] Those skilled in the art will understand that the short-term precipitation forecasting method for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion proposed in this embodiment of the invention can be applied to short-term precipitation forecasting for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion. The structure of the short-term precipitation forecasting system for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion does not constitute a limitation on the short-term precipitation forecasting system for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion. It may include more or fewer components, or combine certain components, or have different component arrangements.

[0198] A memory, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and computer programs.

[0199] In the short-term precipitation forecasting system for hydropower station dam areas in high mountain and valley regions based on deep learning multi-model fusion, the network interface is mainly used to connect to the backend server and communicate with it; the user interface is mainly used to connect to the client (user end) and communicate with it; and the processor can be used to call the computer program stored in the memory. When the computer program is called and executed by the processor, it implements the steps of the above-mentioned short-term precipitation forecasting method for hydropower station dam areas in high mountain and valley regions based on deep learning multi-model fusion.

[0200] Based on the aforementioned embodiments, this invention also proposes a storage medium storing a computer program. When the computer program is executed by the controller, it implements the aforementioned embodiments' method for short-term and near-term precipitation forecasting of hydropower station dam areas in mountainous and valley regions based on deep learning multi-model fusion.

[0201] The short-term precipitation forecasting system and storage medium for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion of the present invention can implement the steps of the above-mentioned short-term precipitation forecasting method for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion. Therefore, it has at least all the beneficial effects brought about by the technical solutions of the above-mentioned embodiments of the short-term precipitation forecasting method for hydropower station dam areas in mountainous valleys based on deep learning multi-model fusion, which will not be elaborated here.

[0202] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0203] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0204] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0205] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0206] The above description is only a part or preferred embodiment of the present invention. Neither the text nor the drawings should limit the scope of protection of the present invention. All equivalent structural transformations made using the content of the present invention specification and drawings under the overall concept of the present invention, or direct / indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion, characterized in that... The method includes: A three-dimensional grid structure of the target area is obtained. The three-dimensional grid structure includes grid cells with spatial connection relationships. Each grid cell includes grid cell location data, historical precipitation data of the grid cell, and historical radar observation data of the grid cell. The target area is the dam area of ​​the hydropower station in the high mountain valley area. Multiple time-series prediction models are used to perform short-time extrapolation predictions on the three-dimensional grid structure to obtain multiple short-time extrapolation prediction results. The time-series prediction models are trained based on the three-dimensional grid structure of the target area, historical precipitation data, and historical radar observation data. Each time-series prediction model corresponds to one short-time extrapolation prediction result. Based on a deep learning model, multiple short-term extrapolation prediction results are fused to obtain a fused extrapolation prediction result. Based on the fusion extrapolation prediction results, a short-term nowcast precipitation forecast for the target area is generated.

2. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 1, is characterized in that... The steps for obtaining the three-dimensional mesh structure of the target region specifically include: Acquire three-dimensional topographic maps, historical precipitation data, and historical radar observation data of the target area; Based on the historical precipitation data and historical radar observation data, determine the actual precipitation range, actual precipitation intensity, forecast precipitation range, and forecast precipitation intensity for each precipitation event within the target area; Based on the actual precipitation range, the actual precipitation intensity, the forecast precipitation range, and the forecast precipitation intensity, the smallest segmentation unit is determined; The three-dimensional topographic map is divided into grids using the minimum segmentation unit to obtain an initial three-dimensional grid structure. The initial three-dimensional grid structure includes initial grid cells with spatial connectivity, and the initial grid cells include the position information of the initial grid cells in the three-dimensional topographic map. For each initial grid cell, historical precipitation data from the current moment backward for a preset period of time is extracted from the historical precipitation data and used as the grid cell's historical precipitation data; In addition, historical radar observation data with a preset time period preceding the current moment is extracted from the historical radar observation data and used as the grid cell's historical radar observation data. The historical precipitation data and historical radar observation data of the grid are added to the initial grid, and the density is adjusted by the terrain features of the target area to obtain the three-dimensional grid structure of the target area.

3. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 2, is characterized in that... The step of determining the minimum segmentation unit based on the actual precipitation range, the actual precipitation intensity, the forecast precipitation range, and the forecast precipitation intensity specifically includes: Based on the actual precipitation range and the forecast precipitation range, the range error for each precipitation event is calculated, and based on the actual precipitation intensity and the forecast precipitation intensity, the intensity error for each precipitation event is calculated. The actual precipitation range with the smallest range error (greater than 0) is used to perform a remainder operation on the overlapping and non-overlapping ranges of the forecast precipitation range. The remainder is used as the side length of the smallest segmentation unit to construct the initial segmentation unit of the cube. The adjustment factor is determined based on the average of the intensity errors from multiple precipitation events. The initial segmentation unit is adjusted by the adjustment factor to obtain the smallest segmentation unit.

4. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 1, is characterized in that... The method further includes a step of training multiple time-series prediction models, wherein the step of training multiple time-series prediction models includes: A first training sample set and multiple time series models to be trained are obtained. The first training sample set includes a first sample, a second sample, a real precipitation label, and a first radar echo extrapolation label. The first sample is obtained by modifying the historical precipitation data of the grid and the historical radar observation data of the grid according to the three-dimensional grid structure of the target area. The second sample is obtained by extrapolating the first sample. The real precipitation label is short-time real precipitation data obtained from historical precipitation data. The first radar echo extrapolation label is short-time radar echo extrapolation data obtained from historical radar observation data. Multiple time series models to be trained are trained using the first training sample set to obtain multiple trained time series prediction models.

5. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 4, is characterized in that... The step of training multiple time-series models to be trained using the first training sample set to obtain multiple trained time-series prediction models specifically includes: For each of the time series models to be trained, the first sample and the second sample are respectively input into the time series model to be trained, and the features of the first sample and the features of the second sample are extracted. The first sample features and the second sample features are fused to obtain fused features; The fused features are subjected to regression processing to obtain extrapolation prediction results and precipitation prediction results; Calculate a first error value between the extrapolated prediction result and the first radar echo extrapolation tag, and calculate a second error value between the precipitation prediction result and the actual precipitation tag; The parameters of the time series model to be trained are adjusted with the goal of minimizing the total error between the first error value and the second error value. The parameter adjustment process of the time series model to be trained is iterated until the error value is less than the preset error threshold or the number of iterations reaches the number threshold, then the training is stopped and a trained time series prediction model is obtained.

6. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 5, is characterized in that... The method further includes a step of training the deep learning model, which specifically includes: A second training sample set and a deep learning model to be trained are obtained. The second training sample set includes a third sample and a second radar echo extrapolation label. The third sample is obtained by modifying the historical precipitation data of the grid and the historical radar observation data of the grid according to the three-dimensional grid structure of the target area. The second radar echo extrapolation label is short-time radar echo extrapolation data obtained from historical radar observation data. The deep learning model to be trained is trained using multiple pre-trained time series prediction models and the second training sample set to obtain a pre-trained deep learning model.

7. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 6, is characterized in that... The step of training the deep learning model to be trained based on multiple trained time-series prediction models and the second training sample set to obtain a trained deep learning model specifically includes: The third sample is input into multiple trained time series prediction models to obtain the sample prediction results output by each time series prediction model. The sample prediction results output by each time series prediction model are concatenated to form a multi-model fusion feature map. The multi-model fusion feature map is input into the deep learning model to be trained; The attention mechanism is used to dynamically weight and fuse the channel-dimensional features from different time-series prediction models in the multi-model fusion feature map, so as to learn and highlight the model features that contribute more to the final forecast in each time-series prediction model. Based on the weighted fused sample features, the final fused prediction result is output through the regression layer of the deep learning model to be trained; The fusion loss is calculated based on the difference between the final fusion prediction result and the second radar echo extrapolation label; With the goal of minimizing the fusion loss, the parameters of the deep learning model to be trained are adjusted, and the parameter adjustment process of the time series model to be trained is iterated until the model converges, thus obtaining a trained time series prediction model.

8. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion as described in claim 7, is characterized in that... The step of dynamically weighting and fusing the channel-dimensional features from different temporal prediction models in the multi-model fusion feature map using the attention mechanism specifically includes: For the multi-model fusion feature map, global average pooling and global max pooling are performed in the spatial dimension to obtain the first pooling vector and the second pooling vector. The first pooling vector and the second pooling vector are added together after multilayer perceptron processing, and attention weights for each channel are generated by an activation function. The generated attention weights are multiplied channel by channel with the multi-model fusion feature map to obtain the weighted fusion sample features.

9. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion according to any one of claims 1 to 8, characterized in that, The step of generating a short-term nowcast precipitation forecast for the target area based on the fused extrapolation prediction results specifically includes: The fused extrapolation prediction results and the precipitation data observed by real-time automatic weather stations are input together into a pre-trained precipitation conversion model; The precipitation conversion model is used to convert the fusion extrapolation prediction results into gridded quantitative precipitation forecasts for future time periods, which serve as short-term nowcastings for the target area.

10. The method for short-term nowcasting of precipitation in the dam area of ​​a hydropower station in a high-altitude valley region based on deep learning multi-model fusion according to any one of claims 1 to 8, characterized in that, The method further includes an evaluation and verification step for assessing the forecast effectiveness after the forecast is generated; the evaluation and verification step includes: Obtain measured precipitation data from surface meteorological stations in the target area during the forecast period; The short-term precipitation forecast is compared with the measured precipitation data to calculate a preset evaluation index value. The evaluation metrics include at least one of the following: structural similarity, correlation coefficient, mean absolute error, root mean square error, hit rate, false alarm rate, critical success index, accuracy, and Heidegger skill score.

Images