New energy station rainstorm disaster short-impending targeted early warning method based on deep learning

Abstract

The application discloses a new energy station rainstorm disaster short-approaching targeted early warning method based on deep learning, aims to solve the problem that the new energy station with surface scattering and many points is difficult to realize differentiated, accurate and advanced early warning under rainstorm weather, divides the station into multiple early warning units, establishes a spatial mapping relationship and a digital decision knowledge base, uses adjacent time radar three-dimensional reflectivity combined with DBSCAN to identify strong convective cloud systems, extracts boundary, centroid, moving direction and state characteristics, screens strong convective clouds with potential threats and binds them with early warning units affected thereby, when the early warning triggering condition is met, fuses cloud system characteristics, initial quantitative rainfall forecast and background characteristics, outputs revised rainfall field of early warning units in the future 0 to 3 hours, combines the spatial mapping relationship, the risk discrimination matrix and the early warning information template, generates and publishes the targeted early warning information, and the method can be used for fine short-approaching early warning of new energy station rainstorm disasters, and is beneficial to improving the pertinence and timeliness of early warning.

Description

Technical Field

[0001] This invention relates to the field of short-term meteorological disaster early warning technology for new energy power stations, and in particular to a short-term targeted early warning method for rainstorm disasters at new energy power stations based on deep learning. Background Technology

[0002] New energy power stations, including wind farms, photovoltaic power stations, and their supporting step-up facilities, are typically characterized by dispersed sites, large coverage areas, and uneven distribution of equipment within individual stations. With the rapid development of the new energy industry, the number of new energy power stations has increased dramatically and they are widely distributed in areas with complex meteorological conditions, such as mountainous areas, coastal areas, river valleys, and high-altitude regions. These areas are also prone to severe convective weather such as short-duration heavy rainfall. Heavy rainfall events are characterized by small spatial and temporal scales, rapid development and evolution, and high destructive potential. They can easily trigger secondary risks such as water accumulation within power stations, slope erosion, road obstruction, and flooding of transformer substations and cable trenches. They can also cause direct physical damage or operational interference to critical facilities such as wind turbine blades, photovoltaic arrays, and step-up equipment, seriously threatening the safe and stable operation of the power system. Unlike conventional energy meteorological services that target administrative regions or large river basins, new energy power station services focus more on whether a specific site, specific zone, or even a specific warning unit will be affected by rain clouds in the next few hours, when rainfall will begin, how long it will last, and what risk level the cumulative rainfall will reach. Therefore, this scenario not only requires short-term nowcasting to have a lead time, but also requires the results to be applicable to the spatial units within the power station. On the other hand, the formation and development of rainstorm processes are influenced by multiple factors, including the generation, movement, merging, and weakening of strong convective cloud systems, as well as the underlying topographic surface conditions, exhibiting strong locality and uncertainty. Existing operational data primarily includes radar 3D reflectivity mosaics, historical observation data, disaster records, initial quantitative rainfall forecast products, and basic geographic information such as topography. How to support accurate and advanced early warning for numerous, scattered new energy power stations within these available data and operational constraints has become a key requirement in practical applications.

[0003] In existing technologies, operational processing for short-term heavy rainfall warnings typically employs methods such as weather radar echo identification, echo extrapolation, numerical weather prediction (NMR) precipitation product correction, automatic weather station precipitation statistics, and threshold discrimination. One approach relies primarily on manual analysis, using indicators such as radar echo intensity, echo top height, and echo movement trends, along with empirical thresholds, to identify severe convective weather processes and assess the risk of future short-term rainfall. Another approach maps numerical weather prediction or quantitative precipitation forecast products to the target area, generating rainfall results for the next few hours through regional averaging, station interpolation, or gridding. These approaches have been applied in urban, watershed, administrative region, and general station-level warnings.

[0004] However, when faced with a massive and dispersed cluster of renewable energy power stations, the application of existing technologies still has the following shortcomings:

[0005] (1) It relies on manual experience and is difficult to scale up, standardize and apply in real time.

[0006] Existing technologies largely rely on forecasters' personal experience and a fully manual operational chain. From manual monitoring of radar echoes and subjective judgment of severe convective cloud systems, to manual matching and experience-based correction of numerical forecast products, and finally to the manual compilation and dissemination of early warning information, the entire process involves multiple sequential steps requiring human intervention. This traditional "human eye + human brain" model, when faced with massive and dispersed renewable energy power plant clusters, suffers from inherent defects such as low efficiency, inconsistent standards, and slow response, making it difficult to meet the demands of large-scale, precise, and real-time modern rainstorm disaster early warning.

[0007] (2) The early warning targets are broad and lack the ability to customize early warning units for differentiated new energy power plants.

[0008] Most existing technologies primarily target administrative regions, unified grids, or single-site locations. Early warning rules and processing procedures typically employ standardized configurations, with limited customization to address the varying degrees of rainstorm-induced disasters affecting different early warning units within a renewable energy power station. Consequently, they struggle to meet the refined operational needs of scenarios characterized by dispersed areas and numerous points of impact. Furthermore, existing solutions often focus on overall prediction of cloud clusters or rainfall fields, lacking a mechanism to establish a clear correspondence between strong convective cloud systems and the affected early warning units. This makes it difficult to directly apply early warning results to specific units, hindering the formation of a targeted early warning closed loop for power station operation and maintenance.

[0009] (3) The lack of a dynamic correction mechanism for precipitation forecast products makes it difficult to adapt to the rapid evolution of local severe convection.

[0010] Existing precipitation forecast products are mostly updated at fixed intervals, and their correction mechanisms rely on post-event statistics or periodic parameter adjustments. They lack the ability to dynamically integrate and calibrate with multi-source data such as real-time radar, satellite, and ground observations. Given the rapid onset and dissipation and uneven spatial distribution of severe convective weather, these forecast products struggle to capture sudden changes in local rainfall intensity, resulting in insufficient timeliness and low accuracy in early warnings. This fails to meet the operational requirements of new energy power plant clusters for "minute-level response and precise localization" of short-duration heavy rainfall.

[0011] Therefore, a short-term targeted early warning method for rainstorm disasters at new energy power stations that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0012] One objective of this invention is to propose a short-term targeted early warning method for rainstorm disasters at new energy power stations based on deep learning. The core technical problem to be solved by this application is: in the application scenario where new energy power stations are scattered and numerous, and the disaster-causing conditions of each early warning unit are different, how to reliably bind strong convective cloud systems with the early warning units affected by them using available radar, historical observations, disaster records, expert rules, initial quantitative rainfall forecasts, and geographic information data, and form a differentiated short-term rainfall discrimination and targeted early warning mechanism for each early warning unit, so as to meet the actual needs of accurate and advanced early warning of heavy precipitation weather at new energy power stations.

[0013] According to an embodiment of the present invention, a deep learning-based short-term targeted early warning method for rainstorm disasters at renewable energy power stations includes:

[0014] S1. Acquire spatial boundary data of new energy power stations, radar monitoring grid parameters and basic geographic information data, divide early warning units, construct benchmark analysis grid, and form spatial mapping relationships;

[0015] S2. Acquire historical observations, disaster records, and expert rules, and configure personalized rules for early warning units to form a digital decision knowledge base that includes cloud system identification rule chain, early warning trigger conditions, risk discrimination matrix, and early warning information templates;

[0016] S3. Receive adjacent radar three-dimensional reflectivity mosaics, align them to the benchmark analysis grid according to the spatial mapping relationship, use DBSCAN to identify strong convective cloud systems based on the cloud system identification rule chain, calculate the boundary, centroid, and direction of movement of the strong convective cloud system, and extract the echo top height, echo bottom height, maximum reflectivity, cloud system area, cloud system volume, and echo gradient to form a cloud system state feature vector.

[0017] S4. Receive the early warning unit, the boundary of the strong convective cloud system, the centroid of the strong convective cloud system, the movement direction of the strong convective cloud system, and the cloud system state feature vector. Calculate the minimum distance from the boundary of each strong convective cloud system to the boundary of the early warning unit. Based on the proximity threat distance threshold, filter out strong convective cloud systems with potential threats. Calculate the angle between the movement direction of the strong convective cloud system and the direction pointing towards the early warning unit. When the angle, maximum reflectivity, and cloud system volume meet the early warning triggering conditions, form an early warning triggering command and a set of cloud system features affecting the early warning unit.

[0018] S5 receives the early warning trigger command, the cloud system feature set affecting the early warning unit, the initial quantitative rainfall forecast product and background features, inputs them into the multi-branch fusion network, fuses the cloud system feature set, relative position features, background features and the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit, and outputs the corrected rainfall field for the early warning unit in the next zero to three hours.

[0019] S6. Receive and correct the rainfall field, spatial mapping relationship, risk discrimination matrix and early warning information template, extract the cumulative rainfall and rainfall start and end time of the early warning unit, determine the early warning level, generate targeted early warning information and send it to the pre-bound receiving end.

[0020] Optionally, S1 includes:

[0021] Acquire spatial boundary data of new energy power stations, and divide the spatial boundary data of new energy power stations into multiple early warning units using early warning unit boundaries;

[0022] Receive radar monitoring grid parameters, determine the grid spacing and grid row and column index of the benchmark analysis grid based on the radar monitoring grid parameters, and extend the benchmark analysis grid to the boundary of the early warning unit;

[0023] Acquire basic geographic information data, assign geographic coordinate labels to the benchmark analysis grid points based on the basic geographic information data, and extract the benchmark analysis grid point sequence from the boundary of the early warning unit.

[0024] Based on the radar monitoring grid parameters and the geographic coordinates of the benchmark analysis grid points, a spatial mapping relationship is formed between the radar monitoring grid and the benchmark analysis grid, and this spatial mapping relationship is then linked to the boundary of the early warning unit and the grid point sequence of the benchmark analysis grid.

[0025] Optionally, S2 includes:

[0026] Acquire historical observations, disaster records, and expert rules, and aggregate historical observations and disaster records according to the boundaries of early warning units;

[0027] Based on historical observations and disaster records, the cloud system identification rule chain parameters corresponding to the early warning unit are determined. The cloud system identification rule chain parameters include at least the maximum reflectivity threshold, echo top height threshold, echo bottom height threshold, cloud system area threshold, cloud system volume threshold, and echo gradient threshold.

[0028] Based on disaster records and expert rules, personalized rule configurations are made for early warning units to form early warning triggering conditions. The early warning triggering conditions include at least the nearby threat distance threshold, the maximum deviation angle threshold, the maximum reflectivity threshold, and the cloud volume threshold.

[0029] Based on historical observations, disaster records, and expert rules, a risk assessment matrix and early warning information template are formed. The cloud system identification rule chain, early warning triggering conditions, risk assessment matrix, and early warning information template are then compiled into a digital decision-making knowledge base.

[0030] Optionally, S3 includes:

[0031] Receive adjacent time-series radar 3D reflectivity mosaics, and map the reflectivity values ​​of adjacent time-series radar 3D reflectivity mosaics to the reference analysis grid according to the spatial mapping relationship to form the reflectivity field of the adjacent time-series reference analysis grid;

[0032] Based on the cloud system identification rule chain, a set of strong convection candidate grid points is selected from the reflectivity field of the benchmark analysis grid in adjacent time intervals. The set of strong convection candidate grid points consists of benchmark analysis grid points that satisfy the constraints of the cloud system identification rule chain.

[0033] DBSCAN is used to identify strong convective cloud systems for candidate grid points. Based on the cloud system identification rule chain, the clustering neighborhood distance parameter and minimum sample number parameter of DBSCAN for identifying strong convective cloud systems are determined, and the clustering results of DBSCAN for identifying strong convective cloud systems are mapped to strong convective cloud system objects.

[0034] The boundary of the strong convective cloud system is calculated based on the connectivity of the outer edge of the grid points of the benchmark analysis grid contained in the strong convective cloud system object, and the centroid of the strong convective cloud system is calculated based on the geographic coordinates of the grid points of the benchmark analysis grid contained in the strong convective cloud system object.

[0035] The pairing relationship of adjacent strong convective cloud systems is determined based on the cloud system identification rule chain, and the movement direction of the strong convective cloud system is calculated based on the displacement vector of the centroid of the adjacent strong convective cloud system.

[0036] For each strong convective cloud system, the echo top height, echo bottom height, and maximum reflectivity are extracted from the reflectivity field of the adjacent time reference analysis grid. The cloud system area is calculated based on the grid coverage of the strong convective cloud system object, and the cloud system volume is calculated based on the height layer coverage of the strong convective cloud system object in the radar three-dimensional reflectivity mosaic.

[0037] The echo gradient is calculated based on the reflectivity difference between adjacent reference grid points inside and outside the boundary of the strong convective cloud system. The echo top height, echo bottom height, maximum reflectivity, cloud area, cloud volume, and echo gradient are combined in a fixed field order to form a cloud system state feature vector.

[0038] Optionally, S4 includes:

[0039] Receive the early warning unit, the boundary of the strong convective cloud system, the centroid of the strong convective cloud system, the direction of movement of the strong convective cloud system, and the cloud system state feature vector, and obtain the boundary of the early warning unit;

[0040] For each strong convective cloud system boundary, extract the boundary point set, calculate the distance between the boundary point set and the early warning unit boundary, and obtain the minimum distance from the strong convective cloud system boundary to the early warning unit boundary;

[0041] The proximity threat distance threshold is determined based on the digital decision-making knowledge base, and strong convective cloud systems with potential threats are screened based on the minimum distance and the proximity threat distance threshold.

[0042] For potentially threatening severe convective cloud systems, the direction pointing towards the early warning unit is determined based on the closest point between the centroid of the severe convective cloud system and the boundary of the early warning unit.

[0043] Calculate the angle between the direction of movement of the strong convective cloud system and the direction pointing towards the early warning unit, and combine the angle with the minimum distance in a fixed field order to form a relative position feature;

[0044] The maximum reflectivity and cloud volume are extracted from the cloud system state feature vector. The warning triggering conditions are determined based on the digital decision knowledge base. The included angle, maximum reflectivity and cloud volume are judged to obtain the warning triggering judgment result.

[0045] Based on the warning trigger determination result, a warning trigger command is generated, and the cloud system state feature vector, relative position feature, strong convective cloud system boundary, strong convective cloud system centroid, and strong convective cloud system movement direction corresponding to the warning trigger determination result are combined to form a cloud system feature set affecting the warning unit.

[0046] Optionally, S5 includes:

[0047] It receives early warning trigger commands, cloud feature sets affecting early warning units, initial quantitative rainfall forecast products and background features, and receives the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit;

[0048] Based on the spatial mapping relationship, the initial quantitative rainfall forecast product is mapped to the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit, forming a grid point rainfall sequence that corresponds one-to-one with the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit.

[0049] The cloud system state feature vector, the movement direction and relative position features of strong convective clouds are extracted from the cloud system feature set affecting the early warning unit. The cloud system state feature vector, the movement direction and relative position features of strong convective clouds are then combined in a fixed field order to form a cloud system input vector set.

[0050] The set of cloud system input vectors is input into the cloud system encoding branch of the multi-branch fusion network. The set of cloud system input vectors is mapped by a fully connected layer to obtain the set of cloud system embeddings. The set of cloud system embeddings is then aggregated to obtain the cloud system embedding vectors.

[0051] Background features are input into the background encoding branch of the multi-branch fusion network. The background features are mapped by a fully connected layer to obtain the background embedding vector. The warning trigger command, cloud embedding vector, and background embedding vector are then input into the fusion layer of the multi-branch fusion network for fusion to obtain the global fusion vector.

[0052] The gridded rainfall sequence is input into the gridded coding branch of the multi-branch fusion network, and the gridded rainfall sequence is mapped by a fully connected layer to obtain the gridded embedding matrix.

[0053] The grid embedding matrix and global fusion vector are input into the grid decoding layer of the multi-branch fusion network to perform grid-by-grid correction on the grid sequence of the benchmark analysis grid within the boundary of the early warning unit, and output the corrected rainfall field for the early warning unit in the next zero to three hours.

[0054] Optionally, S6 includes:

[0055] Receive the corrected rainfall field, spatial mapping relationship, risk discrimination matrix and early warning information template, and determine the correspondence between the corrected rainfall field and the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit based on the spatial mapping relationship;

[0056] The cumulative rainfall of the early warning unit is obtained by hourly summation of the corrected rainfall field, and the start and end times of rainfall are determined based on the continuous periods in the corrected rainfall field where the hourly rainfall is greater than zero.

[0057] The cumulative rainfall and the start and end times of rainfall in the early warning units are determined based on the risk discrimination matrix to establish the early warning level.

[0058] The system fills in the cumulative rainfall, rainfall start and end time and warning level of the warning unit according to the warning information template, generates targeted warning information and sends it to the pre-bound receiving end.

[0059] Optionally, the calculation rule for the minimum distance from the boundary of the strong convective cloud system to the boundary of the early warning unit is as follows: the set of boundary points corresponding to the boundary of the strong convective cloud system is determined as the set of geographic coordinate identifiers of the benchmark analysis grid points constituting the boundary of the strong convective cloud system; the boundary of the early warning unit is discretized into a set of boundary line segments formed by connecting the boundary points in sequence; for each boundary point in the set of boundary points, the point-to-segment distance from the boundary point to each boundary line segment in the set of boundary line segments is calculated and the minimum value is taken as the boundary distance of the boundary point; then, the minimum value of the boundary distances of all boundary points is taken to obtain the minimum distance from the boundary of the strong convective cloud system to the boundary of the early warning unit.

[0060] Optionally, the multi-branch fusion network includes a cloud system encoding branch, a background encoding branch, a fusion layer, a grid point encoding branch, and a grid point decoding layer. The cloud system encoding branch takes a set of cloud system input vectors as input, performs multi-layer fully connected layer mapping on each cloud system input vector in the set to obtain a cloud system embedding set, and performs aggregation on the cloud system embedding set to obtain a cloud system embedding vector. The background encoding branch takes background features as input and performs multi-layer fully connected layer mapping to obtain a background embedding vector. The fusion layer uses the warning trigger command, cloud system embedding vector, and background embedding vector as input to fuse and obtain a global fusion vector. The grid point encoding branch takes the grid point rainfall sequence as input and performs grid point-by-grid multi-layer fully connected layer mapping on the grid point sequence of the benchmark analysis grid within the boundary of the warning unit to obtain a grid point embedding matrix. The grid point decoding layer uses the grid point embedding matrix and the global fusion vector as input to perform grid point-by-grid correction and outputs a corrected rainfall field for the warning unit in the next 0 to 3 hours, which corresponds one-to-one with the grid point sequence of the benchmark analysis grid within the boundary of the warning unit.

[0061] Optionally, the fusion position of the warning trigger command in the multi-branch fusion network is the input end of the fusion layer. After the cloud system encoding branch outputs the cloud system embedding vector and the background encoding branch outputs the background embedding vector, the warning trigger command, the cloud system embedding vector, and the background embedding vector are input into the fusion layer to obtain the global fusion vector. The global fusion vector is then input into the grid decoding layer as the fusion condition input for grid-by-grid correction of the grid embedding matrix. The warning trigger command is not used as the input of the cloud system encoding branch and the grid encoding branch.

[0062] The beneficial effects of this invention are:

[0063] (1) This invention proposes an improved short-term targeted early warning method for rainstorm disasters at new energy power stations. The core improvement lies in refining the early warning object from a general power station or unified grid into early warning units with boundary constraints. Combining historical observations, disaster records, and expert rules, a cloud system identification rule chain, early warning triggering conditions, risk discrimination matrix, and early warning information template are constructed for each early warning unit. Based on this, the method does not only identify strong convective cloud systems as a whole, but further calculates the minimum distance from the boundary of the strong convective cloud system to the boundary of the early warning unit, the angle between the cloud system's movement direction and the direction pointing to the early warning unit, and binds the strong convective cloud system with the early warning units affected by it. Compared with the existing regional early warning algorithms that use a unified threshold and unified partitioning, this scheme transforms whether it will affect a specific early warning unit into a calculable and determinable intermediate quantity, allowing the differences between different units in terms of terrain, disaster tolerance, and historical risk to enter the early warning link. This is beneficial to meeting the business needs of scattered and numerous new energy power stations for differentiated customized early warning.

[0064] (2) This invention proposes a novel rainfall correction method for early warning units. Its technical approach, under conditions where only readily available data such as radar 3D reflectivity mosaic, initial quantitative rainfall forecast products, background geographical features, and early warning triggering results are available, firstly, strong convective cloud systems are identified using DBSCAN, and object-level features such as echo top height, echo bottom height, maximum reflectivity, cloud area, cloud volume, echo gradient, direction of movement, and relative position are extracted. Then, a multi-branch fusion network is used to jointly characterize the object-level cloud system features, unit-level background features, and grid-level rainfall sequences, outputting the corrected rainfall field for the early warning unit for the next 0 to 3 hours. Compared with existing algorithms that rely solely on radar echo extrapolation, numerical rainfall product downsampling, or short-term forecasting based solely on station statistics, this scheme explicitly introduces the cloud system state and cloud-unit relationship affecting the unit into the rainfall correction process. This allows the key intermediate quantity of future rainfall distribution within the early warning unit boundary to be estimated around the threatened cloud system, making it more suitable for serving rainfall extraction, start and end time identification, and subsequent risk assessment within specific units.

[0065] (3) This invention proposes an integrated method that connects identification, triggering, correction, classification, and transmission. It unifies radar monitoring grids, benchmark analysis grids, early warning unit boundaries, and initial quantitative rainfall forecast products into a single processing framework through spatial mapping. The corrected rainfall field is further converted into cumulative rainfall, rainfall start and end times, and early warning levels. Finally, targeted early warning information is generated according to pre-bound receivers. Compared to existing technologies where identification algorithms, forecasting algorithms, and release rules are separated, this solution forms a complete closed loop from identifying strong convective cloud systems to releasing results to specific early warning units. This reduces the conversion steps between regional weather information and executable early warning information within the station, enabling early warning results to directly correspond to specific units and receiving objects within the new energy power station. Therefore, it is more convenient for practical application in real-world operation and maintenance scenarios. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a flowchart of a short-term targeted early warning method for rainstorm disasters at new energy power stations based on deep learning, as proposed in this invention. Detailed Implementation

[0068] In Example 1, reference Figure 1 A short-term targeted early warning method for rainstorm disasters at new energy power stations based on deep learning, comprising:

[0069] S1. Acquire spatial boundary data of new energy power stations, radar monitoring grid parameters and basic geographic information data, divide early warning units, construct benchmark analysis grid, and form spatial mapping relationships;

[0070] S2. Acquire historical observations, disaster records, and expert rules, and configure personalized rules for early warning units to form a digital decision knowledge base that includes cloud system identification rule chain, early warning trigger conditions, risk discrimination matrix, and early warning information templates;

[0071] S3. Receive adjacent radar three-dimensional reflectivity mosaics, align them to the benchmark analysis grid according to the spatial mapping relationship, use DBSCAN to identify strong convective cloud systems based on the cloud system identification rule chain, calculate the boundary, centroid, and direction of movement of the strong convective cloud system, and extract the echo top height, echo bottom height, maximum reflectivity, cloud system area, cloud system volume, and echo gradient to form a cloud system state feature vector.

[0072] S4. Receive the early warning unit, the boundary of the strong convective cloud system, the centroid of the strong convective cloud system, the movement direction of the strong convective cloud system, and the cloud system state feature vector. Calculate the minimum distance from the boundary of each strong convective cloud system to the boundary of the early warning unit. Based on the proximity threat distance threshold, filter out strong convective cloud systems with potential threats. Calculate the angle between the movement direction of the strong convective cloud system and the direction pointing towards the early warning unit. When the angle, maximum reflectivity, and cloud system volume meet the early warning triggering conditions, form an early warning triggering command and a set of cloud system features affecting the early warning unit.

[0073] S5 receives the early warning trigger command, the cloud system feature set affecting the early warning unit, the initial quantitative rainfall forecast product and background features, inputs them into the multi-branch fusion network, fuses the cloud system feature set, relative position features, background features and the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit, and outputs the corrected rainfall field for the early warning unit in the next zero to three hours.

[0074] S6. Receive and correct the rainfall field, spatial mapping relationship, risk discrimination matrix and early warning information template, extract the cumulative rainfall and rainfall start and end time of the early warning unit, determine the early warning level, generate targeted early warning information and send it to the pre-bound receiving end.

[0075] In this embodiment, S1 includes:

[0076] The spatial boundary data of the new energy power station is denoted as ,in, The boundary vertex coordinates are arranged in order of boundary direction. The boundary of the early warning unit used for site zoning is denoted as... ,in, The number of early warning units, For the first The boundary of each early warning unit, for and To perform closed polygon clipping, first determine... Are all boundary vertices located at Internally, when all boundary vertices are located When inside, directly As an early warning unit, when and When boundary intersections exist, the overlapping closed region is reconstructed according to the order of the boundary intersections, and this overlapping closed region is used as an early warning unit. All located in When the grid is external, it does not participate in the calculation, thus obtaining multiple early warning unit boundaries. Each early warning unit boundary maintains a uniform vertex order, which is used to perform grid inclusion determination and boundary localization.

[0077] Let the radar monitoring grid parameters be denoted as ,in, This serves as the starting reference position for the radar monitoring grid. and For radar monitoring grid spacing, and To determine the number of rows and columns in the radar monitoring grid, the coordinates of all boundary vertices of the early warning units are read. The minimum, maximum, minimum, and maximum values ​​of the horizontal and vertical coordinates are then taken to obtain the circumscribed rectangle of the station's early warning range. The coordinates of the lower left corner of this circumscribed rectangle are used as the starting position of the baseline analysis grid. and As the grid spacing for the baseline analysis grid, grid row and column indices are generated sequentially along the horizontal and vertical axes to construct the baseline analysis grid. ,in, For horizontal grid indexing, For vertical grid indexing, each grid point The center coordinates are calculated from the starting position of the reference analysis grid and the grid spacing. The number of horizontal and vertical grid points of the reference analysis grid is determined by dividing the size of the circumscribed rectangle by the corresponding grid spacing and rounding up, so that the boundaries of all early warning units are covered by the reference analysis grid.

[0078] The basic geographic information data is denoted as ,in, For projected coordinates, For each baseline analysis grid point, the corresponding geographic coordinate identifier is used. First read the coordinates of the grid center, then from Select the record with the smallest distance from the center coordinates of the grid point, and assign the geographic coordinate identifier of the record to it. Based on this, a grid point inclusion determination is performed on the boundary of each early warning unit. This includes using ray traversal to determine whether the center coordinates of each grid point are located inside or on the boundary of the corresponding early warning unit. Grid points that meet the determination criteria are extracted in a fixed order of first horizontal indexing and then vertical indexing to form a grid point sequence of the benchmark analysis grid within the boundary of the early warning unit. ,in, For the first The number of grid points within each early warning unit, and a record for each grid point. It consists of six fields, representing the grid row and column index, grid center coordinates, and grid geographic coordinate identifier, respectively;

[0079] The radar monitoring grid is denoted as ,in, For the horizontal indexing of the radar monitoring grid, For the vertical indexing of the radar monitoring grid, a sequence of grid points in the baseline analysis grid within the boundary of each early warning unit is established. Grid records in Read and ,by Using the initial reference position as a baseline, respectively... and Divide by the corresponding grid spacing and take the nearest integer of the result to obtain the horizontal index of the radar monitoring grid. and vertical index ,when and When a point falls within the effective index range of the radar monitoring grid, a spatial mapping relationship is established between the grid record and the radar monitoring grid location, and all corresponding results are written as follows: and will It is used together with the corresponding early warning unit boundary and the reference analysis grid point sequence within the early warning unit boundary, wherein, Provides a fixed order and six-field grid record for grid points within the boundary of the early warning unit. Provides the mapping index of each grid point in the radar monitoring grid. Together, they constitute the direct input for aligning adjacent time-series radar 3D reflectivity mosaics to the baseline analysis grid and mapping the initial quantitative rainfall forecast product to the grid point sequence of the baseline analysis grid within the boundary of the early warning unit. Note: It is assumed that the spatial boundary data of the new energy power station, the boundary of the early warning unit, the radar monitoring grid parameters, and the basic geographic information data use the same projection coordinate system, and that the area where the new energy power station is located is within the coverage area of ​​the radar monitoring grid.

[0080] In this embodiment, S2 includes:

[0081] Let the set of early warning unit boundaries be denoted as , The number of warning units is denoted as the historical meteorological observation sequence. , The number of historical observation entries. and The first The start and end times of each historical observation. For spatial location identification, For observational recording, ten fields sequentially represent maximum reflectivity, echo top height, echo bottom height, cloud area, cloud volume, echo gradient, minimum distance from the boundary of the severe convective cloud system to the boundary of the warning unit, the angle between the direction of movement of the severe convective cloud system and the direction pointing towards the warning unit, cumulative rainfall, and duration of rainfall. The disaster record sequence is denoted as follows: , The number of disaster record entries. and The first The start and end times of each historical observation. For disaster severity levels, threshold-based expert rules are denoted as follows: , For the number of expert rule entries in the threshold category, The five fields represent, in order, the rule name, lower parameter limit, upper parameter limit, sample retention ratio, and rule priority. For each historical observation and each disaster record, a boundary inclusion determination of the early warning unit is performed. When the spatial location identifier falls within... When inside or on the boundary, the corresponding record will be assigned to the first... For each early warning unit, historical observations and disaster records belonging to the same early warning unit are time-bound, with the time-bound condition being: When a historical observation simultaneously meets the time-binding conditions of multiple disaster records, the disaster record with the highest disaster level is selected as the binding result; when the disaster levels are the same, the record with the highest disaster level is selected. The smallest disaster record is used as the binding result;

[0082] For the All historical observations bound to each early warning unit were extracted according to their rule names. , , , , and And for each parameter, only those values ​​below the lower limit of the corresponding parameter are retained. With parameter upper limit Using sample values ​​between these thresholds as candidate thresholds, for the maximum reflectivity threshold, echo top height threshold, cloud area threshold, cloud volume threshold, and echo gradient threshold, the candidate thresholds are first sorted in ascending order of value. Then, each candidate threshold is tried out one by one from the sorted results. When a candidate threshold is adopted, the proportion of historical observations whose sample values ​​are not less than the candidate threshold is first not less than the sample retention proportion. When the current candidate threshold is determined as the corresponding parameter threshold, for the echo bottom height threshold, the candidate thresholds are first sorted from largest to smallest, and then each one is tried from the sorted results. When a candidate threshold is adopted, the proportion of bound historical observations whose sample value is not greater than the candidate threshold is first not less than the sample retention proportion. When the current candidate threshold is determined as the echo bottom height threshold, and multiple threshold-type expert rules exist for the same parameter, the rule priority is used. Threshold determination is performed sequentially from high to low, thus forming the [number]th threshold. The cloud system identification rule chain parameter vector corresponding to each early warning unit The six fields correspond to the maximum reflectivity threshold, echo top height threshold, echo bottom height threshold, cloud area threshold, cloud volume threshold, and echo gradient threshold, respectively. The order of these six fields is the rule execution order when screening strong convection candidate grid points and identifying strong convection cloud systems using DBSCAN.

[0083] For the All early warning units are linked to disaster records; first, determine the highest disaster level. , Indicates the first The highest disaster level that has appeared in the historical samples for each early warning unit, and then select the disaster level that is equal to the one that was bound to it. Historical observations constitute a sample for personalized rule configuration, and the personalized rule configuration sample is extracted. , , and For the proximity threat distance threshold and the maximum deviation angle threshold, candidate thresholds are truncated according to the lower and upper limits of the corresponding expert rules. These thresholds are then sorted by value from smallest to largest, and the candidate thresholds that satisfy the condition that the percentage of samples with values ​​not greater than the candidate threshold is first not less than the sample retention ratio are selected. Similarly, for the maximum reflectivity threshold and the cloud volume threshold, candidate thresholds are truncated according to the lower and upper limits of the corresponding expert rules. These candidate thresholds are then sorted by value from largest to smallest, and the candidate thresholds that satisfy the condition that the percentage of samples with values ​​not less than the candidate threshold is first not less than the sample retention ratio are selected. This process forms the first... The warning triggering condition vector corresponding to each warning unit The four fields correspond to the proximity threat distance threshold, maximum deviation angle threshold, maximum reflectivity threshold, and cloud volume threshold, respectively. Different early warning units use their respective highest disaster level samples to determine the thresholds, thereby completing the personalized rule configuration for the early warning units.

[0084] For the The process of extracting cumulative rainfall from historical observations for each early warning unit. Duration of rainfall It reads the cumulative rainfall grading boundaries, rainfall duration grading boundaries, default disaster level for gaps, and the field order and character format of early warning information given in the expert rules, and uses the cumulative rainfall grading boundaries to determine the risk discrimination matrix. The row index uses the rainfall duration tier boundaries to determine the risk discrimination matrix. The column index maps each bound historical observation to a unique row and column cell, and writes the disaster level associated with that historical observation into the corresponding cell. When multiple bound historical observations correspond to the same cell, the highest disaster level is written. When a cell has no bound historical observations, the default disaster level is written. The warning information template is recorded as follows. The five fields represent, in order, the warning unit identifier fill bit, the warning level fill bit, the cumulative rainfall fill bit, the rainfall start time fill bit, and the rainfall end time fill bit. The cloud system identification rule chain parameter vector corresponding to each early warning unit Early warning trigger condition vector Risk discrimination matrix and early warning information template Write the same knowledge record And they are written into the digital decision-making knowledge base in ascending order of early warning unit number, among which, Provides identification and invocation of strong convective cloud systems. Provided for early warning triggering judgment call, and Provides information for determining early warning levels and generating and calling early warning information.

[0085] In this embodiment, S3 includes:

[0086] The radar three-dimensional reflectivity mosaic sequence is denoted as... ,in, The number of radar times. For the first Each radar time, For the first The radar 3D reflectivity mosaic corresponding to each radar time interval is generated from the same radar monitoring grid. Reflectivity planes of each height layer are indexed by height layer. Formed by stacking, For the horizontal indexing of the radar monitoring grid, For the vertical index of the radar monitoring grid, For height layer index, These are the reflectivity values ​​for the corresponding horizontal and vertical positions, in units of... , For the first The center height of each height layer For the first The thickness of each altitude layer is determined by directly reading it when the radar product provides the altitude layer thickness; otherwise, it is determined by the height difference between the centers of adjacent altitude layers. and , for the Each early warning unit reads the grid point sequence of the benchmark analysis grid. ,in, For the first The number of grid points in the baseline analysis grid within the boundary of each early warning unit, and then the spatial mapping relationship is read. Indexing each grid point ,Will and middle The reflectivity value of the location is written layer by layer. The corresponding height layer forms the reflectivity field of the benchmark analysis grid for adjacent time intervals. and ;

[0087] The first The cloud system identification rule chain parameter vector corresponding to each early warning unit is denoted as follows: The six fields represent, in order, the maximum reflectivity threshold, the echo top height threshold, the echo bottom height threshold, the cloud area threshold, the cloud volume threshold, and the echo gradient threshold. and Each grid point Read the columnar reflectance sequence along the height direction, and take the maximum value of the sequence as the maximum columnar reflectance. Select a reflectance not less than... The height of the center of the highest altitude layer is taken as the echo top height, and the reflectivity is not less than 1. The lowest altitude layer center height is used as the echo bottom height, when there is no reflectivity not less than When the height layer is reached, the grid point is directly removed. When the maximum reflectivity of the columnar layer is not less than [value missing], [the grid point is removed]. The echo peak height is not less than The echo height is not greater than When this occurs, the grid point is written into the strong convection candidate grid point set. or The area of ​​a single cell in the benchmark analysis grid is obtained by multiplying the horizontal grid spacing by the vertical grid spacing. ,Will Divide by Then round up to obtain the minimum number of samples parameter for DBSCAN. The connectivity constraint in the cloud system identification rule chain is set to eight-neighbor connectivity, and the clustering neighborhood distance parameter of DBSCAN is set to... Take the distance from the center of the baseline analysis grid point to the center of the farthest neighboring grid point in the eight-neighbor area as the reference. and The coordinates of the center of each grid point are used as input. DBSCAN is executed on each grid point. When a candidate grid point is in... The number of candidate grid points contained in the neighborhood is not less than When the candidate grid point is identified as a core grid point, all candidate grid points with the same density as the core grid point are then grouped into the same clustering result, resulting in... Temporal clustering results and Temporal clustering results ,in, and These represent the number of clustering results in two adjacent time intervals, and the number of clustering results for each clustering result. or This is the set of grid point indices for the benchmark analysis grid covered by this cluster;

[0088] For each clustering result, the grid points directly adjacent to the four neighbors of the outer grid points of the clustering result are extracted as outer edge grid points. Then, all outer edge grid points are connected end to end in the order of eight-neighbor connectivity to form the boundary of the strong convective cloud system. or For all grid points included in the same clustering result, read the center coordinates from the grid point records. and geographic coordinate identifiers Calculate the arithmetic mean of each to form the centroid of the strong convective cloud system. and ,right Each clustering result at each time step expands outward by one layer of grid points based on eight neighborhoods, forming a matching search index set. In the clustering results, the cluster with the largest number of intersection points with the matching search index set is selected as the paired result. When the number of intersection points is the same, the cluster with the smallest centroid center coordinate distance is selected. When the number of intersection points is zero, it is recorded as unpaired. For paired clustering results, the following is used: Time centroid coordinates subtracted The coordinates of the centroid of time form a displacement vector, and the normalized two-dimensional component is written as the direction of movement of the strong convective cloud system. ,in, For the horizontal component, For the longitudinal component, when the displacement length is zero or unpaired, Written as ;

[0089] right Each clustering result at each time point, from Extract the reflectance values ​​of all grid points and all height layers within the object's coverage area, and select the values ​​that satisfy a reflectance value of not less than [value missing]. The height of the highest layer center is used as the echo top height. Choose a value that satisfies a reflectivity of not less than The lowest height layer center height is used as the echo bottom height. The maximum reflectance value among all reflectance values ​​is taken as the maximum reflectance. Multiply the number of grid points covered by the object by the area of ​​a single grid point. Obtain the cloud system area Within the area covered by the object, the reflectivity must be no less than The three-dimensional voxels are accumulated one by one, and the volume of each three-dimensional voxel is taken as the area of ​​a single cell. With corresponding height layer thickness The sum of the volumes of all three-dimensional voxels is denoted as the cloud system volume. The cloud system volume is obtained using this three-dimensional voxel accumulation method. The quantities involved in the calculation are all grid area, layer thickness, and reflectivity threshold determination results, which can be directly obtained from radar three-dimensional reflectivity mosaic and benchmark analysis grid. A set of adjacent grid point pairs inside and outside the boundary is constructed. Each element is indexed by a grid point located inside the boundary of a strong convective cloud system. An index of a grid point that is adjacent to its four neighbors and located outside the boundary of a strong convective cloud system. The composition is then used to calculate the echo gradient using the following formula. :

[0090] ;

[0091] In the formula, For the first Each early warning unit Time sequence The echo gradient of each clustering result. For the set of adjacent grid points inside and outside the boundary, This represents the number of adjacent grid points inside and outside the boundary. Index of grid points inside the boundary, Indexing the grid points on the outer edge of the boundary. For the first Each early warning unit Time sequence The first benchmark analysis grid point Reflectance values ​​of each height layer For the number of height layers, and The coordinates of the center of the grid point inside the boundary. and Using the coordinates of the center points on the outer side of the boundary, the difference between the maximum reflectivity of the inner and outer cylindrical sections is normalized according to the distance between the centers of adjacent grid points. Then, the arithmetic mean is calculated for all pairs of adjacent grid points inside and outside the boundary to obtain the echo gradient. Subsequently, when... Not less than , Not less than , Not less than At that time, the clustering result is retained as a strong convective cloud system object, and the echo top height, echo bottom height, maximum reflectivity, cloud system area, cloud system volume, and echo gradient are combined in a fixed field order to form a six-dimensional cloud system state feature vector. Each preserved strong convective cloud system object contains a sequence of grid point indices covering the object, a sequence of boundary points of the strong convective cloud system, the centroid of the strong convective cloud system, and a two-dimensional movement direction vector. and the six-dimensional cloud system state feature vector .

[0092] In this embodiment, S4 includes:

[0093] The first The boundary of each early warning unit is denoted as ,in, For the first Number of boundary vertices of each early warning unit For the first The unified projected coordinates of the boundary vertices of each early warning unit are obtained by projecting the geographic coordinates of the boundary vertices of the early warning units onto a plane coordinate system consistent with the baseline analysis grid. These unified projected coordinates will be applied to the first... The severe convective cloud systems of each early warning unit are denoted in object index order as follows: ,in, , The number of strong convective cloud system objects is defined, and each strong convective cloud system object includes a strong convective cloud system boundary. Strong convective cloud system centroid , direction of movement of strong convective cloud systems and cloud system state feature vector ,in, and The unified projected coordinates of the centroid of strong convective cloud systems. and For the geographic coordinates of the centroid of a strong convective cloud system, and These are the lateral and longitudinal components of the direction of movement of the strong convective cloud system, respectively. , , , , , The parameters are, in order: echo top height, echo bottom height, maximum reflectivity, cloud area, cloud volume, and echo gradient. These parameters will be used in the digital decision-making knowledge base. The early warning triggering condition vector corresponding to each early warning unit is denoted as . The four fields represent, in order, the proximity threat distance threshold, the maximum deviation angle threshold, the maximum reflectivity threshold, and the cloud volume threshold. The proximity threat distance threshold... Defined as the maximum boundary distance allowed for strong convective cloud systems to enter the early warning judgment, with units consistent with the uniform projected coordinate length unit, and the maximum deviation angle threshold. Defined as the maximum directional deviation angle allowed to trigger a warning, in degrees;

[0094] Boundaries of the early warning unit Discretize the boundary points sequentially into boundary line segments, and the boundary line segments are formed by... and The sequence is connected to form a boundary that constitutes the strong convective cloud system for each object. The geographic coordinates of the benchmark analysis grid points are identified and converted from the benchmark analysis grid point coordinate records into unified projected coordinates in the same plane coordinate system, forming a sequence of boundary points. ,in, , To determine the number of boundary points, for each boundary point in the boundary point sequence, calculate the point-to-segment distance from that boundary point to all boundary segments of the early warning units. The point-to-segment distance is calculated using the truncation projection coefficient of the boundary point on the corresponding boundary segment. The truncation projection coefficient is less than... Time to take greater than Time to take Located in a closed interval The distance is kept constant within the boundary segment to ensure that it always corresponds to the nearest point on the boundary line. For each boundary point, the minimum value of the distances from all points to the line segment is taken to obtain the boundary distance of that boundary point. Then, the minimum value of the boundary distances of all boundary points is taken to obtain the minimum distance from the boundary of the strong convective cloud system to the boundary of the early warning unit. :

[0095] ;

[0096] In the formula, For the first The first early warning unit and the first The minimum distance from the boundary of a strong convective cloud system to the boundary of the early warning unit between two strong convective cloud systems. Index of boundary points for strong convective cloud systems. For the first Number of boundary points of a strong convective cloud system object Index of the starting point of the boundary line segment of the early warning unit. For the first Number of boundary vertices of each early warning unit and For the first Unified projected coordinates of the boundary points of a strong convective cloud system. and For the first The unified projected coordinates of the boundary vertices of each early warning unit. and In order to be with the first The unified projected coordinates of the next warning unit boundary vertex that is sequentially adjacent to the previous vertex. For the boundary point at the th The truncation projection coefficient on the boundary line segment, Indicates by Modulus calculation to read the distance threshold of nearby threats. Then, press Retain potentially threatening strong convective cloud systems;

[0097] For each preserved strong convective cloud system object, read the centroid of the strong convective cloud system. center coordinates It employs the same search process as that for distances from a point to a line segment to determine the boundary line segment with the smallest distance to the centroid and its corresponding nearest point on all boundary line segments of the early warning units. The two-dimensional vector from the centroid to the nearest point is written as ,in, For the horizontal component, For the longitudinal component, when the centroid does not coincide with the nearest point, and Divide each direction by the length of the two-dimensional vector to obtain the normalized direction pointing to the early warning unit. When the centroid coincides with the nearest point, Written as The direction of movement of strong convective cloud systems and the direction pointing to the early warning unit Perform the angle calculation when or for At that time, the included angle Recorded as In other cases, first calculate the lengths of the two two-dimensional vectors, then sum the product of the horizontal and vertical components and divide by the product of the two lengths to obtain the direction cosine value. Finally, restrict the direction cosine value to a closed interval. Within this range, the inverse cosine of the restricted direction cosine value is taken and converted into an angle value to obtain the angle between the moving direction of the strong convective cloud system and the direction pointing towards the early warning unit. For a strong convective cloud system and an early warning unit, there is only one corresponding direction pointing to the early warning unit. Therefore, the included angle involved in the threshold determination is directly defined as Read the maximum deviation angle threshold Then, press The execution direction consistency determination combines the minimum distance and the included angle in a fixed field order to form a two-dimensional relative position feature. ;

[0098] From cloud system state feature vector Read the third field, maximum reflectivity. and the fifth field cloud system volume Then read the maximum reflectivity threshold. And cloud system volume threshold ,when , and When both are established, the warning trigger determination result will be used. Recorded as In other cases Recorded as In the Within each early warning unit, as long as there is at least one strong convective cloud system that meets the following conditions... Warning trigger command is about to be issued. Recorded as When all strong convective cloud systems satisfy At that time, Recorded as Warning trigger command It is a one-dimensional encoding, and the set of values ​​is All will be satisfied The strong convective cloud system objects are rearranged according to their object index order to obtain the number of cloud systems affecting the early warning unit. For each retained strong convective cloud system, the cloud system state feature vector is... , direction of movement of strong convective cloud systems Relative position features A ten-dimensional object vector is formed by concatenating the fields in a fixed order. The ten fields represent, in order: echo top height, echo bottom height, maximum reflectivity, cloud area, cloud volume, echo gradient, lateral component of the direction of movement, longitudinal component of the direction of movement, minimum distance from the boundary of the strong convective cloud system to the boundary of the warning unit, and the angle between the direction of movement of the strong convective cloud system and the direction pointing towards the warning unit. They are stored side by side in the same object index order. Boundary of strong convective cloud systems Strong convective cloud system centroid and the direction of movement of strong convective cloud systems This forms a set of cloud features that affect the early warning unit. When the multi-branch fusion network is invoked, it directly reads the early warning trigger command. As a one-dimensional trigger input, it directly reads a ten-dimensional object vector. Composition shape is The cloud system input matrix.

[0099] In this embodiment, S5 includes:

[0100] The first The sequence of grid points in the baseline analysis grid within the boundary of each early warning unit is denoted as... ,in, For the first Number of grid points in the baseline analysis grid within the boundary of each early warning unit. For the first The initial quantitative rainfall forecast product is denoted as a base analysis grid of 100 grid points. ,in, For the initial quantitative rainfall forecast product, horizontal grid index, For the initial quantitative rainfall forecast product vertical grid index, This is a rainfall time-leading index, with three time-leading indices corresponding to... Hour, Hour, Hourly rainfall, read the first Spatial mapping relationship of each early warning unit Spatial mapping relationship The initial quantitative rainfall forecast product grid and the baseline analysis grid are obtained by overlaying them in a unified projected coordinate system consistent with the baseline analysis grid, and calculating the overlap area for each intersecting grid pair. For each grid point index... Spatial mapping relationship Directly give the first The set of grid indexes for initial quantitative rainfall forecast products where the areas of intersection occur in the baseline analysis grid cells. , No. Area of ​​each benchmark analysis grid cell , No. Area of ​​each initial quantitative rainfall forecast product grid cell and overlapping area For each rainfall time-leading index The initial quantitative rainfall forecast product is mapped to the first... Based on a baseline analysis grid, the grid point rainfall component is obtained. :

[0101] ;

[0102] In the formula, For the first Within the first early warning unit The benchmark analysis grid points at the _th ... The rainfall amount mapped on the time of rainfall. For early warning unit index, For baseline analysis, grid point index, For rainfall timeliness index, In order to be with the first The set of grid cell indices for initial quantitative rainfall forecast products where the areas of the grid cells intersect in a baseline analysis. For the initial quantitative rainfall forecast product grid cell index, For horizontal grid indexing, For vertical grid indexing, For the first The first benchmark analysis grid cell and the first The overlapping area of ​​grid cells in an initial quantitative rainfall forecast product. For the first Area of ​​each grid cell in the initial quantitative rainfall forecast product For the first Area of ​​each benchmark analysis grid cell. For the initial quantitative rainfall forecast product in the first The first grid cell, the first Rainfall amounts over a given timeframe are ranked according to the order of timeframes. , , Write to the A baseline analysis grid of points was used to generate gridded rainfall sequences. ,in, ;

[0103] Let the background features be denoted as ,in, The average elevation of the early warning unit is obtained by using the arithmetic mean of the digital elevation values ​​corresponding to all grid points of the benchmark analysis grid within the boundary of the early warning unit. The average slope of the early warning unit is obtained by using the arithmetic mean of the terrain slope values ​​corresponding to all grid points in the benchmark analysis grid. The underlying surface type code for the early warning unit is obtained by using the underlying surface category code with the largest area proportion within the boundary of the early warning unit. The early warning trigger command is recorded as... ,in, The number of cloud systems affecting the early warning unit is recorded as For each severe convective cloud system affecting the early warning unit, read the cloud system state feature vector. , direction of movement of strong convective cloud systems and relative position features ,in, ,Will , and A ten-dimensional cloud system input vector is formed by concatenating the data in a fixed field order. Arrange all cloud system input vectors in object index order to form a shape of cloud system input vector set ,when At that time, construct the zero-value cloud system input vector. And write the set of cloud system input vectors as a shape of of ;

[0104] The multi-branch fusion network is configured with a cloud coding branch, a background coding branch, a fusion layer, a grid coding branch, and a grid decoding layer. The actual input to the multi-branch fusion network is the set of cloud input vectors. Background features Warning trigger command and grid-point rainfall sequence The cloud-based coding branch only accepts... For each 10-dimensional cloud system input vector, two fully connected layers are sequentially input, and the output dimensions of each hidden layer are as follows: and After each layer of linear mapping, a linear rectification operation is performed to obtain a shape of... The set of cloud system embeddings is then processed by performing element-wise maximum aggregation along the dimension of the cloud system objects. That is, the maximum value of all cloud system embeddings at the same position is taken for each of the thirty-two feature positions, forming a thirty-two-dimensional cloud system embedding vector. The background encoding branch only receives three-dimensional background features. Two fully connected layers are input sequentially, and the output dimensions of each hidden layer are as follows: and After each layer of linear mapping, a linear rectification operation is performed to form a sixteen-dimensional background embedding vector. The grid-based coding branch only receives grid-based rainfall sequences. For each three-dimensional grid point rainfall vector A grid-by-grid mapping is performed using two fully connected layers with shared parameters, and the output dimensions of each hidden layer are as follows: and After each layer of linear mapping, a linear rectification operation is performed to form a shape of Grid embedding matrix, early warning trigger command Not used as input to the cloud system encoding branch, nor as input to the grid encoding branch, after the cloud system embedding vector and background embedding vector are generated, the one-dimensional warning trigger command, the thirty-two-dimensional cloud system embedding vector, and the sixteen-dimensional background embedding vector are concatenated at the input of the fusion layer in a fixed field order to form a forty-nine-dimensional fusion vector, which is then sequentially input into two fully connected layers, with the output dimensions being as follows: and This forms a 32-dimensional global fusion vector;

[0105] For each benchmark analysis grid point The corresponding 16-dimensional grid point embedding vector and the 32-dimensional global fusion vector are concatenated in field order to form a 48-dimensional grid point fusion vector. Each 48-dimensional grid point fusion vector is then input into the grid point decoding layer. The grid point decoding layer uses a three-layer fully connected layer with shared parameters for grid point-by-grid correction. The output dimensions of the three layers are as follows: , , The first two layers of linear mapping are followed by linear rectification, and the output layer uses linear mapping and reduces the output to less than 100%. The result is truncated as The corrected rainfall vector was obtained. ,in, , , They represent the first Each benchmark analysis grid point is in Hour, Hour, The hourly corrected rainfall data is used to arrange all corrected rainfall vectors in grid index order to form the corrected rainfall field for the next 0 to 3 hours for the early warning unit. The early warning unit will correct the rainfall field shape for the next zero to three hours. The matrix representation has the first dimension corresponding to the grid index in the grid sequence of the benchmark analysis grid within the boundary of the early warning unit, and the second dimension corresponding to the three consecutive rainfall periods. The entire prediction process is completed using a single forward propagation.

[0106] In this embodiment, S6 includes:

[0107] The first The corrected rainfall field corresponding to each early warning unit is denoted as Correcting the rainfall field Depend on The corrected rainfall vectors are composed of elements in index order. For the first The number of grid points in the baseline analysis grid within the boundary of each early warning unit will be the number of grid points in the first early warning unit. The corrected rainfall vector is denoted as ,in, , , , They represent the first Each benchmark analysis grid point is in Hour, Hour, The hourly corrected rainfall is denoted as the spatial mapping relationship. Spatial mapping relationship Store index pairs ,in, For the first The index of the benchmark analysis grid points within the boundary of the early warning unit corresponding to each corrected rainfall vector is denoted as follows: Risk discrimination matrix Each judgment record includes a lower limit for cumulative rainfall, an upper limit for cumulative rainfall, a lower limit for rainfall start time, an upper limit for rainfall start time, a lower limit for rainfall end time, an upper limit for rainfall end time, and a warning level code. The warning information template is recorded as follows: Warning information template It includes fields for warning unit identifier, cumulative rainfall, rainfall start time, rainfall end time, warning level, and sending address;

[0108] Using spatial mapping relationship Correcting the rainfall field Perform grid point correspondence confirmation for each index. Read spatial mapping relationship index pairs in Confirm the first Each corrected rainfall vector is written to the index. After completing the writing of all index pairs for the baseline analysis grid points, hourly rainfall data for each early warning unit is extracted according to three timeframes. Sum by grid index and divide by ,get Hourly Rainfall , will all Sum by grid index and divide by ,get Hourly Rainfall , will all Sum by grid index and divide by ,get Hourly Rainfall ,Will , and By directly adding them together, the cumulative rainfall for the early warning unit can be obtained. The start time of rainfall is recorded as Record the time when the rainfall ends as Both are expressed using an hourly offset relative to the start time of the corrected rainfall field, where the start time corresponds to a time code. After the report is filed Hourly timecode After the report is filed Hourly timecode After the report is filed Hourly timecode Scan in chronological order , , , will be greater than Adjacent time periods are identified as the same continuous rainfall period, and the time code corresponding to the start of the first continuous rainfall period is written as... Write the time code corresponding to the end of the continuous rainfall period as When there are multiple consecutive rainfall periods, the hourly rainfall within each consecutive rainfall period is accumulated, and the consecutive rainfall period with the largest cumulative rainfall is selected as the determining factor. and When the cumulative rainfall is the same during the same period, the continuous rainfall period with the smaller starting time code is selected. , , All At that time, Recorded as and will and All are recorded as ;

[0109] when At that time, read the risk discrimination matrix. All judgment records, first the cumulative rainfall The rainfall was compared with the lower and upper limits of cumulative rainfall in each discrimination record, and then the rainfall start time was determined. The rainfall start time is compared with the corresponding lower and upper limits of the rainfall start time, and then the rainfall end time is determined. A range comparison is performed between the corresponding lower and upper limits of the rainfall termination time. When all three comparisons simultaneously satisfy the same judgment record, the warning level corresponding to that judgment record is determined. When multiple records simultaneously meet the criteria, the record with the highest warning level code is selected as the warning level. When no matching record exists, the warning level will be adjusted. Recorded as not triggering an alert, when At that time, directly raise the warning level Record as no warning triggered;

[0110] Read warning information template The field position will be the first Each early warning unit identifier is written into the early warning unit identifier field, and the cumulative rainfall is recorded. Write the cumulative rainfall field to record the rainfall start time. Write the rainfall start time field and the rainfall end time field. Write the rainfall end time field to set the warning level. Write the warning level field to generate targeted warning information. ,when and At that time, write a no-rainfall flag in the rainfall start time and rainfall end time fields, and read the warning information template. The sending address field in the file is used to obtain the information related to the first... Each early warning unit is pre-bound to a receiving interface address, and targeted early warning information is transmitted according to that interface address. Assembled into a structured message and sent.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0112] This invention, through a closed-loop technical path encompassing early warning unit division, knowledge base customization, cloud system identification and binding, rainfall correction, risk assessment, and targeted dissemination, directly addresses the short-term rainstorm early warning scenario where new energy power stations are scattered, numerous, and have varying disaster-causing conditions among units. Given the uncertainty of the localized generation, changing movement paths, and impact range of strong convective cloud systems, which can easily lead to inaccurate regional early warnings reaching specific power station zones and thus affecting the timing of operation and maintenance responses, this invention, utilizing only readily available data such as radar 3D reflectivity mosaics, historical observations, disaster records, expert rules, initial quantitative rainfall forecast products, and basic geographic information, first establishes a spatial mapping relationship between power station boundaries, early warning units, radar monitoring grids, and benchmark analysis grids. Then, it configures cloud system identification rule chains and early warning triggering conditions for each early warning unit. The system is designed to identify strong convective cloud systems based on risk assessment rules. Then, it identifies these systems using adjacent radar 3D reflectivity mosaics, extracting cloud system boundaries, centroids, movement directions, and state characteristics. Furthermore, it calculates the minimum distance from the cloud system boundary to the warning unit boundary and the angle between the cloud system's movement direction and the direction pointing towards the warning unit. This establishes a stable correspondence between the strong convective cloud system and the affected warning units. Based on this, the system's state characteristics, relative position characteristics, background characteristics, and gridded rainfall sequences are jointly characterized to output the corrected rainfall field for the warning unit within the next 0 to 3 hours. This allows for a reliable estimation of the key intermediate quantity—future rainfall distribution within the warning unit boundary—and subsequently completes the extraction of cumulative rainfall, determination of rainfall start and end times, risk level assessment, and generation of targeted warning information. This achieves an executable warning closed loop for specific warning units.

[0113] Compared to general regional threshold early warning or unified grid rainfall correction schemes, this invention makes targeted improvements to the technical structure to meet the differentiated customization needs of new energy power stations, thereby further enhancing the aforementioned technical effects. Firstly, this invention does not apply uniform rules to the entire power station or administrative region, but rather aggregates historical observations, disaster records, and expert rules into specific early warning units, forming a unit-level cloud system identification rule chain and triggering conditions. This ensures that early warning judgments match the unit's own historical risks and geographical background. Secondly, this invention does not merely identify the cloud cluster itself, but introduces a binding mechanism between strong convective cloud systems and the early warning units affected by them. Through joint determination of minimum distance, directional angle, maximum reflectivity, and cloud system volume, it transforms the determination of whether a particular early warning unit poses a near-term threat into a calculable value. First, this invention provides a filterable and triggerable intermediate criterion, thereby mitigating the generalization problem that easily occurs when issuing warnings solely based on regional echo strength. Second, it proposes a multi-branch fusion network for warning units, which jointly models object-level cloud system input, unit-level background features, and grid-level initial rainfall information in the same model, and performs grid-by-grid correction on the benchmark analysis grid within the unit, avoiding the problem of coarse spatial representation caused by directly transferring external rainfall products to the station. Third, this invention incorporates the identification, triggering, correction, and release links into a unified processing link, enabling the output results to directly correspond to the warning unit, cumulative rainfall, rainfall start and end time, and receiver, thereby improving the feasibility and deliverability of short-term rainstorm warnings for new energy power stations in actual business scenarios.

Claims

1. A short-term targeted early warning method for rainstorm disasters at new energy power stations based on deep learning, characterized in that, include: S1. Acquire spatial boundary data of new energy power stations, radar monitoring grid parameters and basic geographic information data, divide early warning units, construct benchmark analysis grid, and form spatial mapping relationships; S2. Acquire historical observations, disaster records, and expert rules, and configure personalized rules for early warning units to form a digital decision knowledge base that includes cloud system identification rule chain, early warning trigger conditions, risk discrimination matrix, and early warning information templates; S3. Receive adjacent radar three-dimensional reflectivity mosaics, align them to the benchmark analysis grid according to the spatial mapping relationship, use DBSCAN to identify strong convective cloud systems based on the cloud system identification rule chain, calculate the boundary, centroid, and direction of movement of the strong convective cloud system, and extract the echo top height, echo bottom height, maximum reflectivity, cloud system area, cloud system volume, and echo gradient to form a cloud system state feature vector. S4. Receive the early warning unit, the boundary of the strong convective cloud system, the centroid of the strong convective cloud system, the movement direction of the strong convective cloud system, and the cloud system state feature vector. Calculate the minimum distance from the boundary of each strong convective cloud system to the boundary of the early warning unit. Based on the proximity threat distance threshold, filter out strong convective cloud systems with potential threats. Calculate the angle between the movement direction of the strong convective cloud system and the direction pointing towards the early warning unit. When the angle, maximum reflectivity, and cloud system volume meet the early warning triggering conditions, form an early warning triggering command and a set of cloud system features affecting the early warning unit. S5 receives the early warning trigger command, the cloud system feature set affecting the early warning unit, the initial quantitative rainfall forecast product and background features, inputs them into the multi-branch fusion network, fuses the cloud system feature set, relative position features, background features and the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit, and outputs the corrected rainfall field for the early warning unit in the next zero to three hours. S6. Receive the corrected rainfall field, spatial mapping relationship, risk discrimination matrix and early warning information template, extract the cumulative rainfall and rainfall start and end time of the early warning unit, determine the early warning level, generate targeted early warning information and send it to the pre-bound receiving end; S5 includes: It receives early warning trigger commands, cloud feature sets affecting early warning units, initial quantitative rainfall forecast products and background features, and receives the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit; Based on the spatial mapping relationship, the initial quantitative rainfall forecast product is mapped to the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit, forming a grid point rainfall sequence that corresponds one-to-one with the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit. The cloud system state feature vector, the movement direction and relative position features of strong convective clouds are extracted from the cloud system feature set affecting the early warning unit. The cloud system state feature vector, the movement direction and relative position features of strong convective clouds are then combined in a fixed field order to form a cloud system input vector set. The set of cloud system input vectors is input into the cloud system encoding branch of the multi-branch fusion network. The set of cloud system input vectors is mapped by a fully connected layer to obtain the set of cloud system embeddings. The set of cloud system embeddings is then aggregated to obtain the cloud system embedding vectors. Background features are input into the background encoding branch of the multi-branch fusion network. The background features are mapped by a fully connected layer to obtain the background embedding vector. The warning trigger command, cloud embedding vector, and background embedding vector are then input into the fusion layer of the multi-branch fusion network for fusion to obtain the global fusion vector. The gridded rainfall sequence is input into the gridded coding branch of the multi-branch fusion network, and the gridded rainfall sequence is mapped by a fully connected layer to obtain the gridded embedding matrix. The grid embedding matrix and global fusion vector are input into the grid decoding layer of the multi-branch fusion network to perform grid-by-grid correction on the grid sequence of the benchmark analysis grid within the boundary of the early warning unit, and output the corrected rainfall field for the early warning unit in the next zero to three hours.

2. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, S1 includes: Acquire spatial boundary data of new energy power stations, and divide the spatial boundary data of new energy power stations into multiple early warning units using early warning unit boundaries; Receive radar monitoring grid parameters, determine the grid spacing and grid row and column index of the benchmark analysis grid based on the radar monitoring grid parameters, and extend the benchmark analysis grid to the boundary of the early warning unit; Acquire basic geographic information data, assign geographic coordinate labels to the benchmark analysis grid points based on the basic geographic information data, and extract the benchmark analysis grid point sequence from the boundary of the early warning unit. Based on the radar monitoring grid parameters and the geographic coordinates of the benchmark analysis grid points, a spatial mapping relationship is formed between the radar monitoring grid and the benchmark analysis grid, and this spatial mapping relationship is then linked to the boundary of the early warning unit and the grid point sequence of the benchmark analysis grid.

3. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, S2 include: Acquire historical observations, disaster records, and expert rules, and aggregate historical observations and disaster records according to the boundaries of early warning units; Based on historical observations and disaster records, the cloud system identification rule chain parameters corresponding to the early warning unit are determined. The cloud system identification rule chain parameters include at least the maximum reflectivity threshold, echo top height threshold, echo bottom height threshold, cloud system area threshold, cloud system volume threshold, and echo gradient threshold. Based on disaster records and expert rules, personalized rule configurations are made for early warning units to form early warning triggering conditions. The early warning triggering conditions include at least the nearby threat distance threshold, the maximum deviation angle threshold, the maximum reflectivity threshold, and the cloud volume threshold. Based on historical observations, disaster records, and expert rules, a risk assessment matrix and early warning information template are formed. The cloud system identification rule chain, early warning triggering conditions, risk assessment matrix, and early warning information template are then compiled into a digital decision-making knowledge base.

4. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, S3 includes: Receive adjacent time-series radar 3D reflectivity mosaics, and map the reflectivity values ​​of adjacent time-series radar 3D reflectivity mosaics to the reference analysis grid according to the spatial mapping relationship to form the reflectivity field of the adjacent time-series reference analysis grid; Based on the cloud system identification rule chain, a set of strong convection candidate grid points is selected from the reflectivity field of the benchmark analysis grid in adjacent time intervals. The set of strong convection candidate grid points consists of benchmark analysis grid points that satisfy the constraints of the cloud system identification rule chain. DBSCAN is used to identify strong convective cloud systems for candidate grid points. Based on the cloud system identification rule chain, the clustering neighborhood distance parameter and minimum sample number parameter of DBSCAN for identifying strong convective cloud systems are determined, and the clustering results of DBSCAN for identifying strong convective cloud systems are mapped to strong convective cloud system objects. The boundary of the strong convective cloud system is calculated based on the connectivity of the outer edge of the grid points of the benchmark analysis grid contained in the strong convective cloud system object, and the centroid of the strong convective cloud system is calculated based on the geographic coordinates of the grid points of the benchmark analysis grid contained in the strong convective cloud system object. The pairing relationship of adjacent strong convective cloud systems is determined based on the cloud system identification rule chain, and the movement direction of the strong convective cloud system is calculated based on the displacement vector of the centroid of the adjacent strong convective cloud system. For each strong convective cloud system, the echo top height, echo bottom height, and maximum reflectivity are extracted from the reflectivity field of the adjacent time reference analysis grid. The cloud system area is calculated based on the grid coverage of the strong convective cloud system object, and the cloud system volume is calculated based on the height layer coverage of the strong convective cloud system object in the radar three-dimensional reflectivity mosaic. The echo gradient is calculated based on the reflectivity difference between adjacent reference grid points inside and outside the boundary of the strong convective cloud system. The echo top height, echo bottom height, maximum reflectivity, cloud area, cloud volume, and echo gradient are combined in a fixed field order to form a cloud system state feature vector.

5. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, S4 include: Receive the early warning unit, the boundary of the strong convective cloud system, the centroid of the strong convective cloud system, the direction of movement of the strong convective cloud system, and the cloud system state feature vector, and obtain the boundary of the early warning unit; For each strong convective cloud system boundary, extract the boundary point set, calculate the distance between the boundary point set and the early warning unit boundary, and obtain the minimum distance from the strong convective cloud system boundary to the early warning unit boundary; The proximity threat distance threshold is determined based on the digital decision-making knowledge base, and strong convective cloud systems with potential threats are screened based on the minimum distance and the proximity threat distance threshold. For potentially threatening severe convective cloud systems, the direction pointing towards the early warning unit is determined based on the closest point between the centroid of the severe convective cloud system and the boundary of the early warning unit. Calculate the angle between the direction of movement of the strong convective cloud system and the direction pointing towards the early warning unit, and combine the angle with the minimum distance in a fixed field order to form a relative position feature; The maximum reflectivity and cloud volume are extracted from the cloud system state feature vector. The warning triggering conditions are determined based on the digital decision knowledge base. The included angle, maximum reflectivity and cloud volume are judged to obtain the warning triggering judgment result. Based on the warning trigger determination result, a warning trigger command is generated, and the cloud system state feature vector, relative position feature, strong convective cloud system boundary, strong convective cloud system centroid, and strong convective cloud system movement direction corresponding to the warning trigger determination result are combined to form a cloud system feature set affecting the warning unit.

6. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, S6 include: Receive the corrected rainfall field, spatial mapping relationship, risk discrimination matrix and early warning information template, and determine the correspondence between the corrected rainfall field and the grid point sequence of the benchmark analysis grid within the boundary of the early warning unit based on the spatial mapping relationship; The cumulative rainfall of the early warning unit is obtained by hourly summation of the corrected rainfall field, and the start and end times of rainfall are determined based on the continuous periods in the corrected rainfall field where the hourly rainfall is greater than zero. The cumulative rainfall and the start and end times of rainfall in the early warning units are determined based on the risk discrimination matrix to establish the early warning level. The system fills in the cumulative rainfall, rainfall start and end time and warning level of the warning unit according to the warning information template, generates targeted warning information and sends it to the pre-bound receiving end.

7. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 5, characterized in that, The calculation rule for the minimum distance from the boundary of the severe convective cloud system to the boundary of the early warning unit is as follows: the set of boundary points corresponding to the boundary of the severe convective cloud system is determined as the set of geographic coordinate identifiers of the benchmark analysis grid points constituting the boundary of the severe convective cloud system; the boundary of the early warning unit is discretized into a set of boundary line segments formed by connecting the boundary points in sequence; for each boundary point in the set of boundary points, the point-to-segment distance from the boundary point to each boundary line segment in the set of boundary line segments is calculated, and the minimum value is taken as the boundary distance of the boundary point; then, the minimum value of the boundary distances of all boundary points is taken to obtain the minimum distance from the boundary of the severe convective cloud system to the boundary of the early warning unit.

8. The method for short-term targeted early warning of rainstorm disasters at new energy power stations based on deep learning according to claim 1, characterized in that, The multi-branch fusion network includes a cloud system encoding branch, a background encoding branch, a fusion layer, a grid point encoding branch, and a grid point decoding layer. The cloud system encoding branch takes a set of cloud system input vectors as input, and performs multi-layer fully connected layer mapping on each cloud system input vector in the set to obtain a cloud system embedding set. Then, it performs aggregation on the cloud system embedding set to obtain a cloud system embedding vector. The background encoding branch takes background features as input and performs multi-layer fully connected layer mapping to obtain a background embedding vector. The fusion layer takes the warning trigger command, cloud system embedding vector, and background embedding vector as input and performs fusion to obtain a global fusion vector. The grid point encoding branch takes the grid point rainfall sequence as input and performs grid point-by-grid multi-layer fully connected layer mapping on the grid point sequence of the benchmark analysis grid within the boundary of the warning unit to obtain a grid point embedding matrix. The grid point decoding layer takes the grid point embedding matrix and the global fusion vector as input and performs grid point-by-grid correction, outputting a corrected rainfall field for the warning unit in the next 0 to 3 hours that corresponds one-to-one with the grid point sequence of the benchmark analysis grid within the boundary of the warning unit.

9. The method for short-term targeted early warning of rainstorm disasters for new energy power stations based on deep learning according to claim 1, wherein the fusion position of the early warning triggering command in the multi-branch fusion network is the input end of the fusion layer. After the cloud system encoding branch outputs the cloud system embedding vector and the background encoding branch outputs the background embedding vector, the early warning triggering command, the cloud system embedding vector, and the background embedding vector are input into the fusion layer to obtain the global fusion vector. The global fusion vector is then input into the grid decoding layer as the fusion condition input for grid-by-grid correction of the grid embedding matrix. The early warning triggering command is not used as the input of the cloud system encoding branch and the grid encoding branch.

Images