A fan blade icing sub-element reduction scale prediction method and system for complex mountainous areas

Abstract

The application discloses a fan blade icing sub-element reduction scale prediction method and system for complex mountainous areas, and particularly relates to the field of wind farm weather prediction and edge computing, and comprises the following steps: acquiring a weather prediction field, mapping a fan position and a blade sweeping height to a mode grid according to elevation data, and extracting meteorological data according to time; reading an idle period of an edge computing node processor, then reading a storage margin and a link bandwidth, and encoding to obtain fan input data and an edge resource vector; correcting the temperature according to an elevation difference and converting a wet bulb temperature, then correcting the wind speed according to a slope angle, and converting the precipitation to the blade sweeping height after re-distributing the precipitation along an upslope path; through terrain differentiation correction of meteorological elements, the boundary influence of a to-be-refined correction amount on precipitation phase state, icing amount and early warning level is identified, and a refined calculation task is distributed according to an intersection task and an edge resource state.

Description

Technical Field

[0001] This invention relates to the field of wind farm weather forecasting and edge computing technology. More specifically, this invention relates to a method and system for downscaling wind turbine blade icing forecasting based on various factors in complex mountainous terrain. Background Technology

[0002] Existing blade icing forecasts for complex mountain wind farms typically involve spatial interpolation, topographic correction, or downscaling of elements such as temperature, humidity, wind speed, and precipitation from numerical weather predictions. The processor time and memory of edge nodes are allocated based on the task arrival order, computational load, or data error to complete the determination of the precipitation phase at blade height, calculation of icing amount, and generation of early warning results. In wind farms where ridges, canyon exits, windward slopes, and leeward slopes are interspersed, when precipitation occurs, the same edge node needs to process multiple wind turbines, multiple forecast times, and multiple sub-element downscaling tasks before the specified output time. However, the node's processor time and running memory are limited by the installed resources, making it impossible to perform fine-grained calculations on all meteorological elements, terrain units, and blade height layers at the same time. When resources are allocated according to data error or fixed task order, tasks with significant numerical changes that are not enough to change the precipitation phase will continue to occupy computing resources, while tasks with limited numerical changes that are enough to change the distinction between liquid and solid precipitation may be delayed. As a result, node resources are fully occupied while key downscaling tasks are not completed before the data output time. This manifests as repeated switching of the precipitation phase at the blade height in adjacent forecast batches of the same wind turbine, and the icing amount and warning level also show jumps without corresponding meteorological process support. The technical problem this application aims to solve is: under the condition of limited edge computing resources, how to allocate computing resources based on the actual effect of incomplete sub-element downscaling calculations on leaf height precipitation phase, icing results and warning levels. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for downscaling wind turbine blade icing sub-elements in complex mountainous terrain. By implementing topographical differential corrections on meteorological elements, identifying the boundary effects of the correction amounts to be refined on precipitation phase, icing amount, and warning level, and allocating refined calculation tasks based on intersection tasks and edge resource status, the present invention addresses the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for downscaling prediction of wind turbine blade icing elements in complex mountainous terrain, comprising: S1. Obtain the weather forecast field, map the wind turbine location and blade sweep height to the model grid based on the elevation data, and extract meteorological data according to time intervals; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; S2. Correct the temperature based on the elevation difference and convert the wet-bulb temperature. Then, correct the wind speed based on the slope aspect angle. After redistributing the precipitation along the windward uplift path, convert it to the blade sweep height. Combine the wet-bulb temperature sign and the position of the zero-degree isotherm to classify the precipitation phase and obtain the sub-element downscaling results. S3. The Taylor model algorithm is used to refine the correction amount. A second-order polynomial is constructed at the midpoint of the value interval. The polynomial envelope is subtracted from the envelope of the result function to form the remainder interval. The intersection of the polynomial with the phase boundary value or the icing amount boundary value of the warning level is obtained. The first intersection point is associated with the refinement task to form the intersection task set. S4. Based on the intersection task set, a zero-suppression decision graph dynamic programming algorithm is adopted. The zero branch skips the current task, and the one branch deducts the edge resource vector and recalculates the intersection with the refined value. The graph nodes with the same intersection task set and edge resource vector are merged, and the graph branches with negative resource dimension are deleted. The first path is read in ascending order by the number of remaining intersections, and then in ascending order by the resource occupancy to obtain the task allocation result. S5. Execute the detailed tasks according to the task allocation results, update the sub-element downscaling results and edge resource vectors with detailed values, and reconstruct the intersection task set; when the edge resource vector cannot accommodate the remaining tasks, proceed to the next resource cycle until the intersection task set is empty, and output the wind turbine blade icing amount and icing warning level.

[0005] In a preferred embodiment, S1 includes: S11. Locate the model grid point corresponding to the wind turbine location in the digital elevation data, read the adjacent vertical layer along the blade sweep height, and stitch the grid point values ​​of the same meteorological element according to the forecast time to obtain the grid point data block. Digital elevation data refers to the data that records the surface elevation according to the terrain grid under a unified plane coordinate system. Each terrain grid is located by row and column numbers or plane coordinates and corresponds to a surface elevation value. Model grid points refer to the data locations formed by the horizontal discreteness of the weather forecast field. Each model grid point corresponds to a fixed latitude and longitude or plane coordinate, and the grid point values ​​of meteorological elements are stored according to vertical layers and forecast times. The vertical layer refers to the data layer formed by discretizing the weather forecast field along the vertical coordinate of the model. For the wind turbine location, the vertical layer intersecting with the blade sweep height and the vertical layers adjacent to the upper and lower boundaries of the blade sweep height are read. Forecast time refers to the effective time of data determined by the start time and forecast lead time of the weather forecast field. The same forecast time is used to merge meteorological data in the corresponding model grid and vertical layer. Meteorological elements refer to model variables that participate in sub-element downscaling, leaf height and precipitation phase correction, and icing calculation. Each meteorological element has an independent variable code, unit of measurement, and missing data marker. Grid value refers to the numerical record of a meteorological element at a specified model grid point, vertical layer, and forecast time. The numerical record, along with the meteorological element code, spatial location, vertical layer number, and forecast time, is written into the grid data block. S12. Edge computing nodes use the processor idle time period as the time axis, project the storage residence time period and link transmission time period of grid data blocks onto the time axis, extract the common time period of the three, and divide the resource units according to the amount of data that can be completed within the common time period. The storage space remains unoccupied by other tasks during the period from when the edge computing node writes the grid data block to when the grid data block is read from memory after computation. The link transmission period refers to the time interval from the start to the end of the communication link when the idle bandwidth is sufficient to complete the transmission of the grid data block. Data volume refers to the number of bytes of grid data blocks that are transmitted, stored and computed within the shared time period. The capacity is calculated by taking the product of the link bandwidth and the length of the shared time period, the storage capacity, and the number of bytes processed by the processor according to the current task's computation density within the shared time period. The first value of the three capacities in ascending order is taken as the data volume corresponding to the shared time period. S13. Construct an allocation graph using grid data blocks and resource units. Connect allocation edges when resource units can complete grid data block transmission, residence and calculation. Use augmenting paths to match grid data blocks sequentially and deduct the corresponding resource units to obtain wind turbine input data and edge resource vectors. The process of constructing an allocation graph using grid data blocks and resource units is as follows: data nodes are established for each grid data block, and resource nodes are established for each resource unit. The number of bytes, arrival time, and computation duration of each grid data block are read. Then, the common time period, storage capacity, link transmission capacity, and processor processing capacity of each resource unit are read. When a grid data block can complete transmission, residence, and computation within the common time period, an allocation edge is connected between the corresponding data node and resource node to form a bipartite graph. An allocation edge is a graph edge that connects a data node and a resource node. The allocation edge records the number of bytes transmitted, bytes stored, processor runtime, and start and end times of the resource unit occupied by the grid data block. An augmenting path is a path that starts from an unmatched data node, extends along unselected and selected allocation edges alternately, and reaches a resource node with remaining capacity. By exchanging the selection status of each allocation edge in the path, a resource match is established for the starting data node, and the transmission capacity, storage capacity, and processor processing capacity corresponding to the resource node at the end of the path are deducted simultaneously.

[0006] In a preferred embodiment, S2 includes: S21. Read the elevation increment grid by grid along the upwind terrain path from the model grid point to the wind turbine location, and use the elevation increment to recursively correct the temperature. Use the angle between the slope normal and the incoming flow direction to recursively correct the wind speed. Then, redistribute the precipitation according to the water vapor flux in the upwind path. The wet-bulb temperature is obtained iteratively from the corrected temperature, water vapor pressure and air pressure to form a correction task map connected by the terrain path. The reverse-wind terrain path refers to the grid sequence that starts from the terrain grid where the wind turbine is located and backtracks to the model grid point where the wind turbine is located, following the opposite direction of the incoming flow at the current forecast time. During each backtracking, adjacent terrain grids are arranged in ascending order according to the angle between the line connecting the grid centers and the opposite direction of the incoming flow. The first terrain grid is read as the next grid. If the angles are the same, the first terrain grid is read in ascending order according to the grid number. The uphill path refers to rearranging the uphill terrain path according to the direction from the pattern grid point to the wind turbine location, retaining the adjacent grids where the elevation of the next terrain grid is higher than that of the previous terrain grid, and connecting the interconnected grids to form the uphill path; Water vapor flux refers to the mass flow rate of water vapor passing through the interface of adjacent terrain grids with the incoming flow. It is obtained by multiplying the air density, specific humidity, wind speed of the current terrain grid in the normal direction of the interface and the area of ​​the interface. The difference in water vapor flux between two adjacent interfaces is used to determine the precipitation increment of the corresponding terrain grid. The process of recursively correcting temperature and wind speed is as follows: Starting from the temperature and wind speed corresponding to the model grid point, read the elevation increment of the adjacent terrain grid in the upwind terrain path, divide the temperature difference between adjacent vertical layers of the current terrain grid by the inter-layer height difference to obtain the vertical temperature change rate, and then add the product of the vertical temperature change rate and the elevation increment to the corrected temperature of the previous terrain grid; at the same time, project the corrected wind speed vector of the previous terrain grid onto the slope tangent plane of the current terrain grid, use the magnitude of the projection vector as the corrected wind speed of the current terrain grid, and use the current corrected temperature and corrected wind speed as the input of the next terrain grid. The iterative process for obtaining wet-bulb temperature is as follows: First, the dew point temperature is converted from the corrected temperature and water vapor pressure. The dew point temperature is used as the lower bound and the corrected temperature as the upper bound. The midpoint between the upper and lower bounds is read as the candidate wet-bulb temperature. The residual of the wet-bulb equation is calculated based on the saturated water vapor pressure, air pressure, and corrected temperature corresponding to the candidate wet-bulb temperature. When the residual has the same sign as the lower bound residual, the candidate wet-bulb temperature is used to replace the lower bound. When they have different signs, the candidate wet-bulb temperature is used to replace the upper bound. The midpoint and boundary replacement are repeated until the upper and lower bounds are the same after being rounded according to the meteorological data storage precision. The rounded value is then used as the wet-bulb temperature. S22. When merging and correcting task edges with the same input and overlapping terrain paths in the edge computing node graph, processor time slots are allocated according to the data arrival order of the task edges, and storage space is reserved before the corresponding output is read by the subsequent task. When the task edge is executed across processors, the transmission bandwidth is reallocated and the end result of the previous task edge is transmitted to obtain the task allocation table and the correction result within the blade sweep height. The task edge refers to the directed edge in the positive task graph that connects the previous correction task with the next correction task. The task edge indicates that the output of the previous correction task is used as the input of the next correction task, and records the output data volume, generation time, reading time and the processor number where the task is executed. Storage space refers to the address range allocated by the edge computing node from the memory. It is occupied from the moment the task edge outputs and writes data until all subsequent tasks connected to the task edge have completed reading. The number of bytes occupied is determined according to the total number of bytes of data output by the task edge and the corresponding index data. Transmission bandwidth refers to the unit-time transmission capacity allocated from the communication link for transmitting the output of the task side when the two ends of the task side are executed by different processors; the allocated value is obtained by dividing the number of bytes output by the task side by the length of the allocated transmission period, and is released after the transmission is completed; S23. Execute the correction task map according to the task allocation table. When the wet-bulb temperature signs of adjacent vertical layers are different, interpolate to obtain the zero-degree isothermal surface height. Then divide the precipitation according to the overlap range of the blade sweep height and the zero-degree isothermal surface height. Record the overlapping part with positive wet-bulb temperature as liquid precipitation and the overlapping part with negative or zero wet-bulb temperature as solid precipitation to obtain the sub-element downscaling results. The wet-bulb temperature symbol refers to the numerical relationship between the wet-bulb temperature and zero degrees Celsius. A wet-bulb temperature above zero degrees Celsius is recorded as positive, below zero degrees Celsius as negative, and equal to zero degrees Celsius as zero. The process of interpolating to obtain the height of the zero-degree isothermal surface is as follows: read two adjacent vertical layers with different wet-bulb temperature signs, divide the absolute value of the wet-bulb temperature of the lower layer by the sum of the absolute values ​​of the wet-bulb temperatures of the two layers, multiply the resulting ratio by the height difference between the two layers and add it to the height of the lower layer to obtain the height of the zero-degree isothermal surface; when the wet-bulb temperature of any vertical layer is zero, directly take the height of the corresponding vertical layer as the height of the zero-degree isothermal surface.

[0007] In a preferred embodiment, S3 includes: S31. For each correction quantity to be refined, take the midpoint of the value interval as the expansion point, use automatic interval differentiation to obtain the first, second, and third derivative intervals of the result function along the order of operations of precipitation phase correction, icing amount calculation, and warning level determination. Construct a Taylor polynomial with the derivatives within the second term, and generate the remainder term interval from the third derivative interval and the radius of the value interval to obtain a Taylor model that retains the correlation of the correction quantities. Among them, the correction to be refined refers to the single correction in the downscaling results of the sub-elements that has not yet been recalculated according to the target topographic unit, leaf height range or forecast period. Its value range is determined by the lower and upper bounds corresponding to the current correction result and is associated with the source meteorological element, topographic location, leaf height range and forecast period. Automatic differentiation refers to the simultaneous transmission of the current operation value and its first, second, and third derivatives when the result function performs various basic operations, without using finite difference approximation; when encountering addition and subtraction operations, derivatives of the same order are combined respectively; when encountering multiplication, division, and composite operations, the derivatives of each order are updated according to the product rule, quotient rule, and chain rule. Precipitation phase refers to the liquid, solid, or liquid-solid coexistence state of precipitation within the blade sweep height, which is determined by the numerical relationship between wet-bulb temperature and zero degrees Celsius, as well as the positional relationship between the blade sweep height and the zero-degree isotherm height. The second-order term refers to the Taylor expansion term obtained by multiplying the square of the difference between the correction to be refined and the expansion point by half of the second derivative of the result function at the expansion point; the zero-order value, the first-order term, and the second-order term together form a second-order Taylor polynomial. The process of obtaining the first, second, and third derivative intervals of the result function by automatic interval differentiation along the order of operations of precipitation phase correction, icing amount calculation, and warning level determination is as follows: the value interval of the quantity to be refined and corrected is used as the input interval, other corrected quantities are written as constants, and interval addition, subtraction, multiplication, division, and composite function operations are performed item by item according to the order of operations of precipitation phase correction, icing amount calculation, and warning level determination. After each operation, the first, second, and third derivative intervals are updated synchronously according to the chain rule until the result function and its derivative interval within the third order are obtained. The process of constructing a Taylor polynomial using derivatives up to the second order term is as follows: take the midpoint of the interval of the value of the quantity to be refined and corrected as the expansion point, write the function value of the result function at the expansion point as the constant term, multiply the first derivative with the offset of the correction quantity relative to the expansion point to form a linear term, multiply half of the second derivative with the square of the offset to form a quadratic term, and then sum the three terms to obtain the second-order Taylor polynomial. S32. Write the phase boundary value and the icing amount boundary value corresponding to the warning level into the result boundary sequence in the order of operation. Convert the difference between the Taylor polynomial and each result boundary into a Bernstein basis. When the Bernstein coefficients have the same sign, delete the corresponding result boundary. When the Bernstein coefficients have different signs, split the value interval according to the coefficient sign change segment and perform interval Newton mapping. When the mapped interval is contained in the split interval and the derivative interval does not contain zero, record the intersection interval. When the upper and lower boundaries of the split interval are the same after being rounded according to the correction amount storage precision, record the interval to be identified and obtain the candidate intersection table. The phase boundary value refers to the discriminant value corresponding to the phase transition of precipitation; when wet-bulb temperature is used as the discriminant, the phase boundary value is zero degrees Celsius; when the difference between the blade height and the height of the zero-degree isothermal surface is used as the discriminant, the phase boundary value is zero. The ice accumulation threshold refers to the ice accumulation endpoint shared by adjacent ice accumulation warning levels in the warning classification data. Each ice accumulation threshold corresponds to the upper end of the previous warning level and the lower end of the next warning level. The process of converting the difference between the Taylor polynomial and each result boundary into a Bernstein basis is as follows: Let the range of values ​​of the correction to be refined be from the lower bound to the upper bound. Convert the correction to a normalized variable from zero to one. After subtracting the current result boundary from the second-order Taylor polynomial, expand it into a constant term, a first-order term, and a second-order term of the normalized variable. Then convert the three power basis coefficients into second-order Bernstein basis coefficients, so that the difference polynomial can be expressed as a weighted sum of the first-end basis functions, the middle basis functions, and the last-end basis functions. Bernstein coefficients refer to the three weights of the difference polynomial under the quadratic Bernstein basis. When the normalized difference polynomial is written as constant coefficients, linear coefficients and quadratic coefficients, the first Bernstein coefficient is the constant coefficient, the middle Bernstein coefficients are the sum of the constant coefficient and half of the linear coefficient, and the last Bernstein coefficient is the sum of the three power basis coefficients. When Bernstein coefficients have different signs, the process of splitting the value interval according to the coefficient sign-changing segment and performing interval Newton mapping is as follows: First, perform de Castelli split at the midpoint of the value interval to obtain the Bernstein coefficients of the left and right sub-intervals respectively; delete the sub-intervals with three coefficients of the same sign, retain the sub-intervals with coefficients of different signs, and form the Newton mapping interval by subtracting the midpoint value of the difference polynomial from the midpoint of the sub-interval and dividing by the derivative interval of the sub-interval; when the Newton mapping interval has no intersection with the sub-interval, delete the corresponding sub-interval; when there is an intersection, replace the sub-interval with the intersection; when the derivative interval does not contain zero and the intersection is located inside the sub-interval, record the unique intersection interval; continue to perform splitting and mapping on the remaining sub-intervals until the upper and lower bounds of the sub-intervals are the same after being rounded according to the correction amount storage precision.

[0008] In a preferred embodiment, S3 further includes: S33. The edge computing node writes the processor time, storage space and number of bytes transmitted for each interval split and interval Newton mapping in the candidate intersection table into the resource record. Within the limit of the edge resource vector, it constructs a branch search tree, arranges the complete path in ascending order according to the number of intervals to be identified, the total width of the intersection intervals, the processor time occupied, the storage occupied and the number of bytes transmitted. It reads the first path and allocates the corresponding edge resources to perform the operation. Then, according to the result boundary sequence, it reads the first intersection interval or interval to be identified for each quantity to be refined and correlates it with the source refinement task to form an intersection task set.

[0009] In a preferred embodiment, S4 includes: S41. Using the intersection task set as input, recalculate the corresponding Taylor model after skipping the refinement task and after executing the refinement task respectively. Associate the intersection points that no longer intersect with the result boundary after executing the refinement task with the corresponding refinement task, and record the processor time, storage space and transmission bandwidth occupied by the refinement task to obtain the intersection task resource table. The refinement task refers to the task of re-performing the correction calculation after improving the spatial resolution, temporal resolution or vertical stratification of a correction quantity to be refined within the corresponding meteorological elements, topographic location, leaf height range and forecast period. The refinement task takes the source field of the correction quantity to be refined as input, the narrowed value range or recalculated value as output, and records the processor time, storage space and transmission bandwidth occupied during the execution. S42. Based on the intersection task resource table, a zero-suppression decision graph dynamic programming algorithm is used to establish graph nodes. The graph nodes record the remaining intersections, the remaining edge resource vectors, and the corresponding Taylor model. The zero branch retains the current graph node, and the one branch deducts the corresponding processor time, storage space, and transmission bandwidth when the remaining edge resource vectors can undertake the current refinement task. Then, the Taylor model is recalculated with the refinement result, and the intersections that no longer intersect with the result boundary are deleted to obtain candidate graph nodes. The process of establishing graph nodes based on the intersection task resource table using the zero-suppression decision graph dynamic programming algorithm is as follows: refine tasks according to the order of intersections in the result boundary sequence, establish root graph nodes with all intersections, complete edge resource vectors and corresponding Taylor models, and make each layer correspond to a refinement task; generate zero branches that skip the current refinement task and one branch that executes the current refinement task from the current graph node, and write the remaining intersections, remaining edge resource vectors, updated Taylor models and selected refinement tasks after branch processing into the next layer graph node. Graph nodes with the same state field in the same layer share the same successor graph node until all refinement tasks are completed and branches are expanded.

[0010] In a preferred embodiment, S4 further includes: S43. Perform merging and pruning on candidate graph nodes. Merge candidate graph nodes that have the same remaining intersection, remaining edge resource vector, and Taylor model. Delete the latter when the remaining intersection of a candidate graph node is contained in another candidate graph node and the former has no less than the latter in each resource dimension, to obtain the resource allocation graph. S44. Accumulate the remaining number of intersections, the width of the intersection interval, and the resource usage layer by layer along the resource allocation graph. Arrange the paths to the terminal graph nodes in ascending order of remaining number of intersections, ascending order of intersection interval width, and ascending order of resource usage. Read the refined tasks and edge computing node resources corresponding to the first path to obtain the task allocation results.

[0011] In a preferred embodiment, S5 includes: S51. Read the refinement task and corresponding edge resource vector according to the task allocation result. The edge computing node executes the refinement task within the allocated time period, writes the refinement value into the sub-element downscaling result, and deducts the edge resource vector according to the actual occupation to obtain the execution record. The refined value refers to the numerical result obtained after the refinement task re-performs downscaling correction within the corresponding meteorological elements, topographic locations, leaf height ranges, and forecast periods. When the refinement task uses the subdivided topographic grid, vertical layer, or time slice for calculation, the correction results of each subdivided location are read and merged according to the original task's spatial and temporal ranges. The merged value replaces the original correction amount in the sub-element downscaling results. S52. Based on the execution record, update the corresponding Taylor model with the refined value and re-find the intersection point. Delete the tasks that no longer intersect with the result boundary, retain the tasks that still intersect with the result boundary, and then associate the updated intersection point with the source task to obtain the updated intersection point task set.

[0012] In a preferred embodiment, S5 further includes: S53. When the updated intersection task set is not empty and any remaining task has insufficient resource dimension, generate the resource availability interval according to the release time of each resource dimension of the edge resource vector, find the first common time period of each remaining task and the resource availability interval, read the time corresponding to the first task in ascending order of the common time period start as the start time of the next resource cycle, and write the updated intersection task set, Taylor model and sub-element downscaling results into the acceptance record. The resource dimension refers to the three numerical fields in the edge resource vector that represent the processor's executable time, the number of bytes that can be written to the memory, and the number of bytes that can be transmitted through the communication link, respectively. When the demand of the remaining tasks in any numerical field exceeds the corresponding remaining amount, it is recorded as insufficient resource dimension. S54. In the next resource cycle, read the acceptance record, reconstruct the edge resource vector and re-execute the task allocation until the updated intersection task set is empty. Then, calculate the icing amount of the wind turbine blades and determine the icing warning level based on the sub-element downscaling results. The process of calculating the icing amount of wind turbine blades and determining the icing warning level is as follows: Liquid precipitation, solid precipitation, corrected wind speed, and wet-bulb temperature within the blade sweep height are read according to the forecast time. The liquid interception coefficient is read using the liquid interception coefficient table corresponding to the wind turbine blade model, and the solid adhesion coefficient is read using the solid adhesion coefficient table. The product of liquid precipitation and liquid interception coefficient and the product of solid precipitation and solid adhesion coefficient are added together, and then multiplied by the corrected wind speed and the current forecast time length to obtain the icing increment for the current forecast time. The icing increment is accumulated according to the forecast time to obtain the icing amount of the wind turbine blades. Then, the icing amount range containing the icing amount of the wind turbine blades is found in the warning level data, and the corresponding level is used as the icing warning level.

[0013] A component-based downscaling forecasting system for wind turbine blade icing in complex mountainous terrain. The system includes a data mapping module, a component downscaling module, an intersection point identification module, a resource allocation module, and an iterative output module. The data mapping module is used to acquire weather forecast fields, map the wind turbine position and blade sweep height to model grid points based on elevation data, and extract meteorological data on a time basis; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; The sub-item downscaling module is used to correct temperature and convert wet-bulb temperature based on elevation difference, and then correct wind speed based on slope aspect angle. After redistributing precipitation along the windward uplift path, it is converted to the blade sweep height. Combined with the wet-bulb temperature sign and the position of the zero-degree isotherm, the precipitation phase is divided to obtain the sub-element downscaling results. The intersection identification module is used to employ the Taylor model algorithm for the refined correction quantity, construct a second-order polynomial at the midpoint of the value interval, subtract the polynomial envelope from the envelope of the result function to form the remainder term interval, find the intersection point of the polynomial with the phase boundary value or the icing amount boundary value of the warning level, and associate the first intersection point with the refinement task to form the intersection task set. The resource allocation module uses a zero-suppression decision graph dynamic programming algorithm based on the intersection task set. Zero branches skip the current task, while one branch deducts the edge resource vector and recalculates the intersection with the refined value. The module merges graph nodes where the intersection task set and the edge resource vector are the same, deletes graph branches with negative resource dimensions, and reads the first and second paths in ascending order by the number of remaining intersections and then by the amount of resources used to obtain the task allocation result. The iterative output module is used to execute refinement tasks according to the task allocation results, update the sub-element downscaling results and edge resource vectors with refinement values, and reconstruct the intersection task set; when the edge resource vector cannot accommodate the remaining tasks, it will move to the next resource cycle until the intersection task set is empty, and output the wind turbine blade icing amount and icing warning level.

[0014] The technical effects and advantages of this invention are as follows: This solution identifies the actual impact of incomplete downscaling tasks on precipitation phase, icing amount, and warning level, and then allocates marginal resources accordingly, so that critical tasks are completed first, thereby relatively reducing the data output delay caused by ineffective occupation. Temperature, wind speed, and precipitation were corrected along the complex mountain flow path, and precipitation phases were classified by combining wet-bulb temperature and zero-degree isotherm, which relatively improved the stability of liquid precipitation and solid precipitation determination within the blade sweep height. A Taylor model is constructed for the refined correction, and the intersection of the result boundaries is located by using Bernstein's base and interval Newton mapping, thereby distinguishing tasks that will change the phase state or warning level and relatively reducing redundant calculations that have no substantial impact. When processor, memory, and communication link resources are limited, a zero-suppression decision graph is used to jointly compare the remaining intersections, the width of the intersection interval, and the resource usage, so that the task allocation results are more in line with the actual role of each refined task. By continuously updating the Taylor model, the downscaling results of each element, and the intersection task set through a cross-resource cycle succession mechanism, the unfinished tasks converge round by round, thereby relatively suppressing unfounded jumps in icing amount and warning level between adjacent forecast batches. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1-2 The present invention provides a method for downscaling prediction of wind turbine blade icing elements in complex mountainous terrain, comprising: S1. Obtain the weather forecast field, map the wind turbine location and blade sweep height to the model grid based on the elevation data, and extract meteorological data according to time intervals; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; The formation of wind turbine input data requires the simultaneous completion of spatial positioning, height matching, temporal merging, and edge resource arrangement, so that the data read by subsequent sub-factors downscaled have consistent wind turbine numbers, forecast validity times, and data versions. Weather forecast data, digital elevation data, and wind turbine locations are first converted to the same plane coordinate system, and then the vertical layer height is uniformly converted to altitude. Once the current mode's reporting time and data version are locked, other batches of data are not mixed into the same calculation process. The correspondence between the wind turbine location and the weather forecast field is established based on a unified coordinate system; the ground elevation of the digital elevation grid where the wind turbine is located is read, and the ground elevation is added to the hub height to obtain the hub center elevation. Then, the blade length is subtracted from and added to the hub center elevation to obtain the lower and upper boundaries of the blade sweep height. When weather forecasting fields use pressure layers, mixed layers, or model layers for numbering, read the geopotential height or the layer boundary height and convert it to altitude. S11 locates the model grid points containing the wind turbine plane coordinates, reads the vertical layers that intersect with the blade sweep height and are adjacent to the upper and lower boundaries, adds the model start time and forecast lead time to obtain the forecast validity time, arranges the grid point values ​​according to the forecast validity time, vertical layer number and meteorological element code, and writes the unit of measurement, missing measurement mark and data version into the grid point data block. When the wind turbine location falls on the boundary of the pattern grid, the first pattern grid point is read in ascending order of the pattern grid point number; duplicate records are retained for the newest data version and the record written later; when necessary meteorological elements are missing or the blade sweep height exceeds the vertical coverage range, the corresponding grid point data block is stopped from being built. After completing the gridded data blocks, the edge computing nodes delineate the resource scope shared by transmission, residence, and computing according to the time intersection; The processor idle period is written into the time axis according to the start and end times. The transmission duration is obtained by dividing the number of bytes of grid data block by the link idle bandwidth. The link transmission period that can accommodate the transmission duration is located in the link idle interval. The period from the expected write time of the grid data block to the expected completion time of all subsequent tasks constitutes the candidate storage residency period. The number of bytes occupied in the corresponding period shall not exceed the storage capacity. S12 projects the processor idle period, the candidate storage residency period and the link transmission period onto the same time axis, and reads the intersection of the three as the common period. The link transmission capacity is obtained by multiplying the link idle bandwidth by the shared time period length. The storage margin corresponding to the shared time period is used as the storage capacity. The processor processing capacity is obtained by dividing the shared time period length by the processing time of a unit input byte. The three capacities are arranged in ascending order and the first value is read as the data capacity of the resource unit. When the grid data block exceeds the resource unit capacity, it is first split by forecast time, and then by vertical layer number. Meteorological element codes, grid values ​​and missing measurement marks in the same vertical layer are kept in the same sub-block. The resource matching phase establishes a correspondence between grid data blocks and resource units capable of handling data transmission, storage residency, and processor computation; Grid data blocks are written as data nodes, and resource units are written as resource nodes; the number of bytes, arrival time, and processor time requirement of grid data blocks are read, and the common time period and remaining capacity of resource units are read. Only when grid data blocks can complete transmission, writing, calculation, and last read within the common time period are allocation edges connected between the corresponding nodes. S13 processes unmatched data nodes according to the forecast validity time, arrival time and grid data block number, and searches for resource nodes with remaining capacity alternately along the unselected allocation edge and the selected allocation edge from the current data node; After finding an augmenting path, the selected status of the allocated edges within the path is exchanged. Newly selected allocated edges are deducted from the processor capacity, storage capacity, and link capacity simultaneously, while canceled allocated edges are returned to the three capacities simultaneously. Grid data blocks for which no augmentation path is found are written to the pending acceptance table and left to be matched again in the next resource cycle; grid data blocks that have been matched are transmitted, resided and calculated according to the time period recorded by the allocation edge, and then written to the wind turbine input data. The remaining capacity of the resource unit is written to the edge resource vector in the order of the common time period. Taking hourly icing forecast of a ridge wind farm as an example, after the wind turbine coordinates are unified with the weather forecast field, the edge computing nodes read the temperature, humidity, wind speed and precipitation data covering the blade sweep height, and form grid data blocks according to the same forecast validity time. The processor is idle from 2:00 PM to 2:10 PM, the storage space can be maintained until 2:12 PM, and the communication link completes data transmission before 2:30 PM. The shared time period of the three types of resources forms a resource unit. The allocation graph adjusts the existing match through augmenting paths so that the current grid data block can obtain complete processor, memory and communication link resources. The resulting wind turbine input data has a unified spatial location, height range, forecast time, and data version. The edge resource vector synchronously records the remaining capacity after resource allocation, providing a definite data entry point and resource basis for subsequent sub-element downscaling.

[0018] S2. Correct the temperature based on the elevation difference and convert the wet-bulb temperature. Then, correct the wind speed based on the slope aspect angle. After redistributing the precipitation along the windward uplift path, convert it to the blade sweep height. Combine the wet-bulb temperature sign and the position of the zero-degree isotherm to classify the precipitation phase and obtain the sub-element downscaling results. In complex mountainous terrain, model meteorological data need to be corrected for topography along the actual inflow path and to separate liquid precipitation from solid precipitation within the blade sweep height. Before processing, the wind turbine number, forecast validity time, and weather forecast field data version are locked, and then the path generation, meteorological element correction, edge resource allocation, and vertical phase division are completed in sequence. The terrain correction takes the grid where the wind turbine is located as the starting point for backtracking. According to the opposite direction of the incoming flow at the current forecast time, the grids with written paths are deleted from the eight adjacent grids. Then, the next grid is read in ascending order according to the angle between the grid center line and the opposite direction of the incoming flow. When the included angles are the same, they are sorted by grid number. When the horizontal wind speed is zero, the opposite direction of the incoming flow is replaced by the opposite direction from the center of the model grid point to the position of the fan, until the grid where the model grid point is located is reached. After the terrain path is reversed in the opposite direction, S21 reads the elevation increment grid by grid from the pattern grid to the wind turbine position. The temperature vertical change rate is obtained by dividing the temperature difference between the upper and lower vertical layers surrounding the current grid elevation by the interlayer height difference. The product of the change rate and the elevation increment is added to the previous grid to correct the temperature. Stop the current temperature recursion when the vertical layers are the same; construct the slope tangent vector and obtain the slope normal using the difference in elevation between the horizontal and vertical terrain; project the corrected wind speed vector of the previous grid onto the current slope tangent plane; the projected vector and its magnitude are written as the corrected wind speed vector and the corrected wind speed, respectively. The adjacent grids whose elevation rises along the execution direction in the path form the windward slope path. The water vapor flux is calculated by multiplying the air density, specific humidity, the normal projection of the corrected wind speed vector, and the interface area, and the net water vapor inflow is obtained by the difference in flux between adjacent interfaces. The net inflow of positive water vapor is used as the weight for precipitation allocation, so that the total original precipitation within the path remains unchanged. When the total net inflow is not positive, it is allocated according to the proportion of the original precipitation. The dew point temperature is obtained from the corrected temperature and water vapor pressure. Using the dew point temperature and the corrected temperature as the upper and lower bounds, the boundary is repeatedly replaced by the residual of the wet-bulb equation corresponding to the midpoint temperature until the upper and lower bounds are the same after being rounded according to the temperature storage resolution, thus obtaining the wet-bulb temperature. When the water vapor pressure exceeds the saturated water vapor pressure, it is written according to the saturated water vapor pressure. If the residual does not form an opposite sign, an input abnormality is recorded. Temperature recursion, wind speed projection, precipitation redistribution, and wet-bulb temperature calculation are each established as correction task nodes. The task edges are connected according to the input-output relationship to form a correction task graph. After a task enters an edge computing node, repeated computations are compressed and execution resources are configured according to data dependencies. Task edges with consistent meteorological element codes, forecast validity periods, data versions, computation types, and overlapping terrain paths are merged by S22. When paths only overlap at the beginning, task edges before the first forked grid are merged. Processor time slots are allocated according to the arrival time of the output data. When the arrival times are the same, they are sorted according to the forecast validity time, wind turbine number, and task number. After the task output is written to the memory, a read count is established according to the number of subsequent tasks. Each time a read is completed, the count is deducted by one. When the count reaches zero, the storage space is released. When tasks are assigned to different processors, the transmission bandwidth is obtained by dividing the number of bytes output by the task by the transmission time period. The link capacity is released after the end result is transmitted. When the transmission capacity is insufficient, the start time of the next task will be postponed; the task number, processor number, start and end time, input address, output address and predecessor task number are written into the task allocation table, and the corrected temperature, wet-bulb temperature, corrected wind speed vector and redistributed precipitation are written into the correction results within the blade sweep height. Data with input anomaly markers will not be included in phase division. The precipitation phase within the blade sweep height is determined based on the sign change of wet-bulb temperature along the vertical direction; after arranging the wet-bulb temperatures according to the vertical layer elevation, S23 reads adjacent vertical layers one by one; When the signs of the two wet-bulb temperatures are different, divide the absolute value of the lower wet-bulb temperature by the sum of the absolute values ​​of the two wet-bulb temperatures, multiply by the height difference between the layers, and add it to the height of the lower layer to obtain the height of the zero-degree isothermal surface. When the wet-bulb temperature of any layer is zero, directly read the height of the corresponding vertical layer. When there are multiple zero-degree isothermal surfaces, sort them by height and divide the blade sweeping height. When the wet-bulb temperature at the midpoint of each interval is positive, it is recorded as the liquid interval; when it is negative or zero, it is recorded as the solid interval. Precipitation over a period of time is divided according to the length ratio of liquid and solid phase intervals; when recording precipitation by vertical layer, it is first allocated according to layer thickness, and then divided according to the overlap length of phase intervals. When all vertical layers have a positive wet-bulb temperature, all precipitation is written as liquid precipitation; when all are negative or zero, all precipitation is written as solid precipitation. The downscaling results for each element are recorded by wind turbine number, forecast validity period, and data version, including corrected temperature, wet-bulb temperature, corrected wind speed vector, redistributed precipitation, liquid precipitation, solid precipitation, and zero-degree isotherm height. Taking hourly forecast of ridge wind field as an example, when the incoming flow path passes through two separate windward slope uplift intervals, the original precipitation of each interval is redistributed according to the net water vapor inflow; when two wind turbines share the correction task from the mode grid to the first bifurcation grid, the edge computing nodes merge the shared task and execute the bifurcation task respectively. When two zero-degree isothermal surfaces appear within the blade sweep height, the height range is divided into three intervals: liquid, solid, and liquid. Precipitation is recorded as liquid precipitation and solid precipitation respectively according to the overlap length of the intervals, forming a sub-element downscaling result consistent with the terrain path, blade height, and edge resource status.

[0019] S3. The Taylor model algorithm is used to refine the correction amount. A second-order polynomial is constructed at the midpoint of the value interval. The polynomial envelope is subtracted from the envelope of the result function to form the remainder interval. The intersection of the polynomial with the phase boundary value or the icing amount boundary value of the warning level is obtained. The first intersection point is associated with the refinement task to form the intersection task set. Before the downscaling results of each element are included in the icing calculation, it is necessary to identify which corrections will still change the precipitation phase or icing warning level, and to prioritize the allocation of limited marginal resources to the corresponding detailed tasks. After the current model's start time, forecast validity period, and data version are locked, the correction interval model is established sequentially, the intersection points of the result boundaries are narrowed, and the intersection point calculation paths are selected. For corrections that have not yet been recalculated according to target terrain units, blade height ranges, or forecast periods, the original calculation range is first divided into sub-units, and the first and last values ​​of the corrections of the completed sub-units are used to form the lower and upper bounds of the value range, respectively. When a sub-unit has not yet been calculated, read the candidate values ​​obtained by adjacent completed units using the same correction rule. If there are fewer than two valid candidate values, do not build a Taylor model. S31 then uses the midpoint of the value interval as the expansion point, and makes the same refined correction quantity share a common interval variable in the repeated calls in phase correction and icing amount calculation. The wet-bulb temperature discrimination, the difference between the blade height and the height of the zero-degree isothermal surface discrimination, and the calculation of icing amount are all established as result functions. The warning level mapping only provides the icing amount boundary value and does not participate in the derivative transfer. When the result function contains sign judgment, the value interval is split according to the phase boundary value. When it contains coefficient table interpolation, the value interval is split according to the range node in the table. When the division denominator interval contains zero, it continues to split on both sides of the zero point. If it still cannot be separated according to the correction quantity storage precision, write the unidentified mark. Within each smooth interval, the operation value and the first, second and third derivatives are transmitted synchronously. Addition and subtraction operations merge derivatives of the same order. Multiplication, division and composite functions update derivatives according to the product rule, quotient rule and chain rule respectively. Read the function value, first derivative value, and second derivative value at the expansion point to construct the constant term, linear term, and quadratic term. Divide the third derivative interval within the entire value interval by six and multiply it by the cubic interval of the correction offset to obtain the remainder term interval. Write the result function number, operation branch number, value interval, Taylor polynomial, and remainder term interval into the Taylor model. Intersection location starts with sequential encoding of phase boundary value and icing amount boundary value; S32 matches the result boundary with the same dimension for each Taylor model, normalizes the correction quantity to be refined to zero to one according to the lower and upper bounds of the value interval, and expands the Taylor polynomial into constant coefficients, first-order coefficients and second-order coefficients after subtracting the current result boundary. The first Bernstein coefficient is taken as a constant coefficient, the middle Bernstein coefficients are taken as the sum of the constant coefficient and half of the first coefficient, and the last Bernstein coefficient is taken as the sum of the three power base coefficients. When all three Bernstein coefficients are positive or all three are negative, delete the current result boundary; when any coefficient is zero, write the corresponding endpoint or control position as a candidate intersection point; when all three coefficients are zero, write the entire value range as the interval to be identified. When there are coefficients with opposite signs, perform a de Castelli split at the midpoint of the current interval and process the left and right sub-intervals in the order of their lower bounds. For the retained subintervals, the interval of the Newton map is obtained by subtracting the midpoint value of the difference polynomial from the midpoint of the interval and dividing by the interval of the derivative. When the derivative interval includes zero, the midpoint split continues. When the Newton mapping interval and the subinterval have no intersection, the subinterval is deleted. When there is an intersection, the subinterval is replaced by the intersection. When the derivative interval does not contain zero, the intersection is located inside the subinterval, and the Bernstein coefficients still have opposite signs, record the unique intersection interval; when the upper and lower bounds are the same after being rounded according to the correction amount, but do not meet the uniqueness condition, record the interval to be identified. The candidate intersection table is written according to the result function number, result boundary number, intersection interval, unique intersection marker, unidentified marker, source correction amount and refinement task number, and the first intersection is determined according to the result boundary order, the lower bound of the intersection interval and the task number; The allocation of computing resources is based on the number of intervals and the type of operation in the candidate intersection table; the edge computing node reads the processor time, storage bytes and transmission bytes corresponding to one de Castelli split and one interval Newton mapping from the operation record table, and then combines them with the number of sub-intervals to be processed to form the expected resource requirements; S33 establishes a root node based on the candidate intersection state and the remaining edge resource vector. Each time, it selects to execute or postpone an interval operation to form a branch. Executing the branch deducts the processor capacity, storage capacity and link capacity in the corresponding resource unit. When any resource dimension has a negative value, the corresponding branch is deleted. A complete path is formed when each interval in the candidate intersection table is deleted, recorded as a unique intersection interval, or recorded as an interval to be identified, and the resource consumption does not exceed the edge resource vector. The complete path is arranged in order of the number of intervals to be identified, the total width of the intersection intervals, the processor time occupied, the storage usage, and the number of bytes transmitted. The first and second paths are read and the intersection operation is performed. The actual resource usage is written into the intersection operation resource record. When the current resource cycle cannot form a complete path, the status of the processed interval and the unprocessed interval are written into the acceptance record together; after the calculation is completed, the first unique intersection interval or the interval to be identified for each quantity to be refined and corrected is read according to the result boundary sequence, and an association is established with the source refinement task to form an intersection task set; For example, when a wet-bulb temperature range to be refined crosses zero degrees Celsius and the corresponding icing amount crosses an early warning level endpoint, the phase function is first split at the position corresponding to zero degrees Celsius, and then the Taylor model of icing amount is established separately; after the Bernstein coefficient excludes sub-intervals without intersections, the interval Newton mapping shrinks the remaining range into a single intersection interval; Edge computing nodes allocate processor, memory, and communication link resources based on the number of interval splits and mappings. Candidate intervals with insufficient resources are transferred to the next resource cycle. The first intersection is ultimately associated with the corresponding terrain unit, blade height interval, and forecast period for refinement tasks.

[0020] S4. Based on the intersection task set, a zero-suppression decision graph dynamic programming algorithm is adopted. The zero branch skips the current task, and the one branch deducts the edge resource vector and recalculates the intersection with the refined value. The graph nodes with the same intersection task set and edge resource vector are merged, and the graph branches with negative resource dimension are deleted. The first path is read in ascending order by the number of remaining intersections, and then in ascending order by the resource occupancy to obtain the task allocation result. After the intersection task set is formed, it is necessary to determine the refined task combination within the limited processor time, storage space and transmission bandwidth, and ensure that the corresponding result boundary intersections can be eliminated after the task is executed; Therefore, we first estimate the impact of each detailed task on the Taylor model and the intersection state, and then use the zero-suppression decision graph to expand the task selection, compress the repetitive state, and read the execution path with the resource occupancy order first. When the intersection task enters the resource estimation stage, it reads the source meteorological elements, topographic location, leaf height range, forecast period and the range of values ​​to be refined and corrected. It then breaks down the spatial refinement task into sub-grid calculations within the target topographic unit, the vertical refinement task into sub-layer calculations within the leaf height range, and the temporal refinement task into sub-time slice calculations within the forecast period. Since the task has not yet been executed, S41 first constructs a refined result envelope based on the input range of each sub-unit, and then updates the corresponding Taylor model with the result envelope; When the updated Taylor model has no intersection with the result boundary, the corresponding intersection point is recorded as an eliminable intersection point of the current refinement task; if there is still an intersection, the intersection point is retained. The processor requirement is summed by multiplying the number of sub-unit operations by the duration of a single operation; the storage requirement is summed by the number of bytes of input data, intermediate results, and output results; and the transmission requirement is summed by the number of bytes of output across processor task edges. The resulting intersection association, result envelope, and three types of resource requirements are written into the intersection task resource table. When a task is associated with multiple intersections, all associations are retained. When multiple tasks jointly affect an intersection, the associations are recalculated in subsequent branches. After the task order is determined, the tasks are refined by the first intersection point order, result boundary number, number of corrections to be refined, and task number. The root graph node is established by S42 and all remaining intersection points, complete edge resource vector, Taylor model and task layer number are written. Each task layer generates a zero branch and a one branch: the zero branch does not select the current refinement task, inherits the remaining intersection points, remaining edge resource vectors and Taylor model, and increments the task layer number by one; One branch first checks the available time of the processor, storage capacity and link transmission capacity in the same resource unit kernel. When all three types of resources cover the task requirements, the corresponding capacity is deducted. Then, the Taylor model is updated with the refined result envelope and the intersection points that no longer intersect with the result boundary are deleted. When resources are insufficient, no branch is generated; when a branch fails to connect to a terminal, the current graph node is deleted and the predecessor graph node is directly connected to the zero-branch successor node, thereby compressing the unexecutable task state and obtaining candidate graph nodes. As the task layers unfold, candidate graph nodes need to be merged according to state equivalence and resource dominance; S43 first compares the remaining intersection codes, edge resource unit codes, and Taylor model codes within the same task layer. When the result function number, operation branch number, zero-order value, first-order coefficient, second-order coefficient, remainder interval and value interval are all consistent, the candidate graph nodes are merged, and the predecessor path with the highest resource consumption is retained. When two candidate graph nodes are not completely identical, the second graph node is deleted only if the remaining intersection of the first graph node is contained in the second graph node, the processor time period coverage relationship, storage capacity, and link capacity of each resource unit of the first graph node are not less than those of the second graph node, and the Taylor model value envelope formed by the first graph node for the common intersection points contains the corresponding envelope of the second graph node. The intervals to be identified are processed based on the remaining intersections and are not subject to unconditional deletion; the merged graph nodes, predecessor numbers, and branch types together form the resource allocation graph. The path reading phase starts from the root graph node and proceeds along the resource allocation graph to the terminal graph node. The terminal graph node corresponds to the completion of all refined tasks and branch expansion, or the remaining resource units are unable to handle the unprocessed refined tasks. S44 accumulates the remaining number of intersections in each task layer along each complete path, adds the widths of the unique intersection interval and the interval to be identified to form the total width of the intersection interval, and then accumulates the processor time, storage bytes, and transmission bytes respectively. The complete path is first sorted in ascending order by the number of remaining intersections, then sorted in ascending order by the total width of the intersection interval. If the values ​​are the same, the processor time, storage bytes, and transmission bytes are compared in turn. If they are still the same, the first path is read in ascending order by the selected task number sequence. The detailed task, resource unit number, processor time period, storage address range, and transmission time period in the first path are written into the task allocation result for subsequent task execution to read; For example: There are three intersection tasks for the same forecast time of a certain wind turbine. The spatial refinement task can simultaneously eliminate the intersection of wet-bulb temperature phase state and icing amount level. The temporal refinement task only shrinks the icing amount intersection interval. The vertical refinement task needs to occupy another processor and transmit intermediate results. The resource allocation graph expands the branches for skipping and executing the three tasks, deletes paths that exceed the current processor time slot or link capacity, and then merges the remaining intersections, resource states, and graph nodes that are consistent with the Taylor model. After path sorting, the space refinement task that eliminates two intersections and has the highest resource usage is written into the task allocation result, while the remaining tasks are retained for subsequent resource cycles.

[0021] S5. Execute the detailed tasks according to the task allocation results, update the sub-element downscaling results and edge resource vectors with detailed values, and reconstruct the intersection task set; when the edge resource vector cannot accommodate the remaining tasks, proceed to the next resource cycle until the intersection task set is empty, and output the wind turbine blade icing amount and icing warning level. The detailed task execution phase is used to transform the task allocation results into factor-based downscaling results, and to continuously update the intersection task set based on actual resource consumption; During the processing, the wind turbine number, forecast validity time and data version are locked, and detailed calculations, intersection verification, resource cycle acceptance and icing early warning output are completed in sequence. Before a task is started, the edge computing node checks the resource unit number, allocation period, storage address range, and transmission period in the task allocation results, and confirms that the processor, memory, and communication link can still cover the current task requirements; if resources are insufficient, the task will not be started, and the refined task and corresponding Taylor model will be written into the acceptance record. After the resource verification is approved, temperature, wind speed, precipitation and wet-bulb temperature corrections are re-executed in S51 according to the subdivided terrain grid, vertical layer or time slice; Temperature and wet-bulb temperature are weighted and merged according to the product of sub-unit coverage and time length. Wind speed is corrected by first merging vector components and then calculating the modulus. Precipitation is summed according to sub-unit values ​​and the original total precipitation is maintained. When all sub-units are completed, the merged single value is output. If there are still sub-units that are not completed, the value range after shrinkage is formed by the values ​​of the completed sub-units and the boundaries of the incomplete sub-units. The refined value replaces the original correction, and the task number, resource unit number, start and end time, actual processor usage time, number of bytes used for storage, number of bytes transmitted, refined value, and execution status are written into the execution record. At the same time, the edge resource vector is updated according to the actual usage. After the execution record is generated, it is necessary to reassess the impact of the refinement value on the result boundary; S52 reads the refinement value and the source correction amount, and reconstructs the corresponding Taylor model and candidate intersection table; When the updated result function envelope and the result boundary have no intersection, delete the corresponding intersection point. Delete the task when the associated task no longer corresponds to other intersection points. If there is still an intersection, retain the original task. Re-associate and refine the task according to the result function number, result boundary number, and source correction amount for newly generated intersection points. When the current edge resource vector can still handle at least one remaining task, update the intersection task set and return to the resource allocation graph to continue allocation; when all remaining tasks exceed the current remaining amount in at least one of the processor time, storage capacity, or link capacity, they are transferred to resource cycle handling. Resource availability is determined based on the time when the three types of resources can simultaneously meet the task requirements; when processing reaches S53, the processor end time, the last read time of the storage, and the link transmission completion time of the currently executing task are read respectively to form the corresponding resource availability interval; For each remaining task, first filter the intervals where processor capacity, storage capacity, and link capacity all cover the task requirements, and then find the first intersection of the three types of intervals as the common time period; The common time period start points are arranged in ascending order of time. If the start points are the same, they are arranged in order of the first intersection point and the task number. The first time point is written as the start point of the next resource cycle. Update the intersection task set, Taylor model, sub-element downscaling results, incomplete sub-units, input data version and resource cycle number and write them into the acceptance record; The level of refinement is limited by weather forecast field, digital elevation data and time resolution. At least one terrain unit, vertical layer or time slice must be completed in each round. The starting point of the resource cycle is shifted after the resource release time. No more refinement tasks are generated after all sub-units are completed. When the next resource cycle arrives, S54 reads the acceptance record, reconstructs the edge resource vector based on the processor idle time period, storage balance and link idle bandwidth at that time, and then performs task allocation and refinement calculation until the updated intersection task set is empty; The icing calculation uniformly records liquid precipitation and solid precipitation as precipitation intensity per unit time. First, the relative incoming flow velocity is obtained by the corrected wind speed vector, blade angular velocity and blade radius position. Then, the liquid interception coefficient and solid adhesion coefficient are read by the wind turbine blade model, wet bulb temperature storage level, relative incoming flow velocity storage level and blade radius range. When the input is between two gears, it is interpolated according to the gear distance ratio; when it exceeds the range in the table, the boundary gear is read. The product of liquid precipitation intensity and liquid interception coefficient and the product of solid precipitation intensity and solid adhesion coefficient are added together and then multiplied by the forecast duration to obtain the icing increment for the current forecast duration. When liquid and solid coexist, the values ​​are calculated separately and then added together. The total icing amount of the wind turbine blades, expressed in kilograms, is obtained by accumulating the values ​​for the forecast duration. The warning classification data is recorded according to the intervals that are included on the left and not included on the right. The last interval includes the right end. When the ice coverage is equal to the common endpoint of the adjacent intervals, the next level is written. When the forecast time arrives but the intersection task set has not been cleared, the two adjacent warning levels and the incomplete mark are saved, and the definite level is not output. In conjunction with an hourly forecast of the ridge wind field, in a real-world scenario: after the spatial refinement task is completed, the edge computing nodes merge temperature and wet-bulb temperature according to the sub-grid coverage, correct wind speed according to the vector components, and update the sub-element downscaling results with the sum of sub-grid precipitation; when the updated icing amount Taylor model no longer crosses the warning level endpoint, the corresponding intersection and task are deleted. The remaining vertical refinement tasks will be transferred to the next resource cycle due to insufficient current link capacity. They will continue to be executed after the processor, memory and communication link form a shared time period. After the intersection task set is cleared, the composite conversion coefficient is read based on the relative incoming flow velocity, wet bulb temperature and blade model to obtain the cumulative icing amount of the wind turbine blades and the corresponding icing warning level.

[0022] exist Figure 1 It should be noted that: the flowchart, from S1 to S5, represents data mapping and resource coding, element downscaling, result boundary intersection identification, edge resource allocation, and refined execution and result updating in sequence; solid arrows indicate that each processing result is passed to the next stage; the judgment node is used to check whether the intersection task set is empty; if the judgment is not empty, it returns to the resource allocation stage to continue processing; if the judgment is yes, it outputs the wind turbine blade icing amount and icing warning level. exist Figure 2 It should be noted that: the system diagram shows that weather forecast field, digital elevation data and wind turbine parameters are processed sequentially by five functional modules, and output the wind turbine blade icing amount and icing warning level; solid arrows indicate the direction of data and calculation results transmission, the dashed lines pointing from the central edge resource core to each module indicate the support relationship of processor, storage space, transmission bandwidth and resource cycle, and the dashed lines returning from the iterative output module to the resource allocation module indicate that when the intersection task set is not cleared, resources continue to be allocated and refined tasks are executed.

[0023] Furthermore, the present invention also includes a component-based downscaling forecasting system for wind turbine blade icing in complex mountainous terrain. The system includes a data mapping module, a component downscaling module, an intersection point identification module, a resource allocation module, and an iterative output module. The data mapping module is used to acquire weather forecast fields, map the wind turbine position and blade sweep height to model grid points based on elevation data, and extract meteorological data on a time basis; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; The sub-item downscaling module is used to correct temperature and convert wet-bulb temperature based on elevation difference, and then correct wind speed based on slope aspect angle. After redistributing precipitation along the windward uplift path, it is converted to the blade sweep height. Combined with the wet-bulb temperature sign and the position of the zero-degree isotherm, the precipitation phase is divided to obtain the sub-element downscaling results. The intersection identification module is used to employ the Taylor model algorithm for the refined correction quantity, construct a second-order polynomial at the midpoint of the value interval, subtract the polynomial envelope from the envelope of the result function to form the remainder term interval, find the intersection point of the polynomial with the phase boundary value or the icing amount boundary value of the warning level, and associate the first intersection point with the refinement task to form the intersection task set. The resource allocation module uses a zero-suppression decision graph dynamic programming algorithm based on the intersection task set. Zero branches skip the current task, while one branch deducts the edge resource vector and recalculates the intersection with the refined value. The module merges graph nodes where the intersection task set and the edge resource vector are the same, deletes graph branches with negative resource dimensions, and reads the first and second paths in ascending order by the number of remaining intersections and then by the amount of resources used to obtain the task allocation result. The iterative output module is used to execute refinement tasks according to the task allocation results, update the sub-element downscaling results and edge resource vectors with refinement values, and reconstruct the intersection task set; when the edge resource vector cannot accommodate the remaining tasks, it will move to the next resource cycle until the intersection task set is empty, and output the wind turbine blade icing amount and icing warning level.

[0024] Overall Implementation and Explanation: This scheme first maps the weather forecast field to the complex mountainous terrain, wind turbine location, and blade sweep height, and combines the idle status of the processor, memory, and communication link to form wind turbine input data and edge resource vectors; then, it performs terrain correction on temperature, wind speed, and precipitation along the incoming flow direction, calculates wet-bulb temperature and zero-degree isotherm height, and obtains the downscaling results of liquid precipitation, solid precipitation, and other sub-elements within the blade sweep height; For corrections that may still change the precipitation phase or icing warning level, the Taylor model is used to identify the intersections with the phase boundary value or icing amount boundary value. The corresponding refinement tasks are then assigned to the zero-suppression decision map for combination and allocation within limited marginal resources. The refinement results are continuously updated with the sub-element downscaling results and the intersection task set. When resources are insufficient, the process is moved to the next resource cycle until the intersection task set is cleared. Finally, the icing amount of the wind turbine blades is calculated and the icing warning level is determined. Taking hourly icing forecasts for ridge wind farms as an example, weather forecast data is first mapped to the target wind turbine and blade sweep height, and then the temperature, wind speed, and precipitation distribution are corrected along the path of the northwest-flowing mountain slope. When two zero-degree isothermal surfaces appear within the blade sweep height, the height range is divided into three intervals: liquid, solid, and liquid, and the precipitation is then allocated to the corresponding phase. If one of the wet-bulb temperature corrections may still cause a change in the precipitation phase, and another precipitation correction may still cause the icing amount to cross the warning level boundary, the edge computing node will prioritize the refinement tasks that can eliminate more intersections and have higher resource consumption. Tasks with insufficient resources at present will be transferred to subsequent resource cycles for further processing. After all uncertain intersections are eliminated, the icing amount will be accumulated based on liquid precipitation, solid precipitation, relative inflow velocity, and blade parameters, and the corresponding warning level will be output.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for downscaling prediction of wind turbine blade icing elements in complex mountainous terrain, characterized in that, include: S1. Obtain the weather forecast field, map the wind turbine location and blade sweep height to the model grid based on the elevation data, and extract meteorological data according to time intervals; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; S2. Correct the temperature based on the elevation difference and convert the wet-bulb temperature. Then, correct the wind speed based on the slope aspect angle. After redistributing the precipitation along the windward uplift path, convert it to the blade sweep height. Combine the wet-bulb temperature sign and the position of the zero-degree isotherm to classify the precipitation phase and obtain the sub-element downscaling results. S3. The Taylor model algorithm is used to refine the correction amount. A second-order polynomial is constructed at the midpoint of the value interval. The polynomial envelope is subtracted from the envelope of the result function to form the remainder interval. The intersection of the polynomial with the phase boundary value or the icing amount boundary value of the warning level is obtained. The first intersection point is associated with the refinement task to form the intersection task set. S4. Based on the intersection task set, a zero-suppression decision graph dynamic programming algorithm is adopted. The zero branch skips the current task, and the one branch deducts the edge resource vector and recalculates the intersection with the refined value. The graph nodes with the same intersection task set and edge resource vector are merged, and the graph branches with negative resource dimension are deleted. The first path is read in ascending order by the number of remaining intersections, and then in ascending order by the resource occupancy to obtain the task allocation result. S5. Execute the refinement task according to the task allocation result, update the sub-element downscaling result and edge resource vector with the refinement value, and reconstruct the intersection task set; When the edge resource vector cannot accommodate the remaining tasks, it will move to the next resource cycle until the intersection task set is empty, and output the icing amount of the wind turbine blades and the icing warning level.

2. The method for downscaling wind turbine blade icing prediction based on multiple factors in complex mountainous terrain, as described in claim 1, is characterized in that: S1 includes: S11. Locate the model grid point corresponding to the wind turbine location in the digital elevation data, read the adjacent vertical layer along the blade sweep height, and stitch the grid point values ​​of the same meteorological element according to the forecast time to obtain the grid point data block. S12. Edge computing nodes use the processor idle time period as the time axis, project the storage residence time period and link transmission time period of grid data blocks onto the time axis, extract the common time period of the three, and divide the resource units according to the amount of data that can be completed within the common time period. S13. Construct an allocation graph using grid data blocks and resource units. Connect allocation edges when resource units can complete grid data block transmission, residence, and calculation. Use augmenting paths to sequentially match grid data blocks and deduct corresponding resource units to obtain wind turbine input data and edge resource vectors.

3. The method for downscaling wind turbine blade icing prediction based on multiple factors in complex mountainous terrain, as described in claim 2, is characterized in that: S2 includes: S21. Read the elevation increment grid by grid along the upwind terrain path from the model grid point to the wind turbine location, and use the elevation increment to recursively correct the temperature. Use the angle between the slope normal and the incoming flow direction to recursively correct the wind speed. Then, redistribute the precipitation according to the water vapor flux in the upwind path. The wet-bulb temperature is obtained iteratively from the corrected temperature, water vapor pressure and air pressure to form a correction task map connected by the terrain path. S22. When merging and correcting task edges with the same input and overlapping terrain paths in the edge computing node graph, processor time slots are allocated according to the data arrival order of the task edges, and storage space is reserved before the corresponding output is read by the subsequent task. When the task edge is executed across processors, the transmission bandwidth is reallocated and the end result of the previous task edge is transmitted to obtain the task allocation table and the correction result within the blade sweep height. S23. Execute the correction task map according to the task allocation table. When the wet-bulb temperature signs of adjacent vertical layers are different, interpolate to obtain the zero-degree isothermal surface height. Then, divide the precipitation according to the overlap range between the blade sweep height and the zero-degree isothermal surface height. Record the overlapping part with positive wet-bulb temperature as liquid precipitation and the overlapping part with negative or zero wet-bulb temperature as solid precipitation to obtain the sub-element downscaling results.

4. The method for downscaling wind turbine blade icing prediction based on multiple elements in complex mountainous terrain, as described in claim 3, is characterized in that: S3 includes: S31. For each correction quantity to be refined, take the midpoint of the value interval as the expansion point, use automatic interval differentiation to obtain the first, second, and third derivative intervals of the result function along the order of operations of precipitation phase correction, icing amount calculation, and warning level determination. Construct a Taylor polynomial with the derivatives within the second term, and generate the remainder term interval from the third derivative interval and the radius of the value interval to obtain a Taylor model that retains the correlation of the correction quantities. S32. Write the phase boundary value and the icing amount boundary value corresponding to the warning level into the result boundary sequence in the order of operation. Convert the difference between the Taylor polynomial and each result boundary into a Bernstein basis. When the Bernstein coefficients have the same sign, delete the corresponding result boundary. When the Bernstein coefficients have different signs, split the value interval according to the coefficient sign change segment and perform interval Newton mapping. When the mapped interval is contained in the split interval and the derivative interval does not contain zero, record the intersection interval. When the upper and lower boundaries of the split interval are the same after being rounded according to the correction amount storage precision, record the interval to be identified to obtain the candidate intersection table.

5. The method for downscaling wind turbine blade icing prediction based on multiple factors in complex mountainous terrain, as described in claim 4, is characterized in that: S3 also includes: S33. The edge computing node writes the processor time, storage space and number of bytes transmitted for each interval split and interval Newton mapping in the candidate intersection table into the resource record. Within the limit of the edge resource vector, it constructs a branch search tree, arranges the complete path in ascending order according to the number of intervals to be identified, the total width of the intersection intervals, the processor time occupied, the storage occupied and the number of bytes transmitted. It reads the first path and allocates the corresponding edge resources to perform the operation. Then, according to the result boundary sequence, it reads the first intersection interval or interval to be identified for each quantity to be refined and correlates it with the source refinement task to form an intersection task set.

6. The method for downscaling wind turbine blade icing prediction based on multiple elements in complex mountainous terrain, as described in claim 5, is characterized in that: S4 includes: S41. Using the intersection task set as input, recalculate the corresponding Taylor model after skipping the refinement task and after executing the refinement task respectively. Associate the intersection points that no longer intersect with the result boundary after executing the refinement task with the corresponding refinement task, and record the processor time, storage space and transmission bandwidth occupied by the refinement task to obtain the intersection task resource table. S42. Based on the intersection task resource table, a zero-suppression decision graph dynamic programming algorithm is used to establish graph nodes. The graph nodes record the remaining intersections, the remaining edge resource vectors, and the corresponding Taylor model. The zero branch retains the current graph node, and the one branch deducts the corresponding processor time, storage space, and transmission bandwidth when the remaining edge resource vectors can undertake the current refinement task. Then, the Taylor model is recalculated with the refinement result, and the intersections that no longer intersect with the result boundary are deleted to obtain candidate graph nodes.

7. The method for downscaling wind turbine blade icing prediction based on multiple elements in complex mountainous terrain as described in claim 6, characterized in that: S4 also includes: S43. Perform merging and pruning on candidate graph nodes. Merge candidate graph nodes that have the same remaining intersection, remaining edge resource vector, and Taylor model. Delete the latter when the remaining intersection of a candidate graph node is contained in another candidate graph node and the former has no less than the latter in each resource dimension, to obtain the resource allocation graph. S44. Accumulate the remaining number of intersections, the width of the intersection interval, and the resource usage layer by layer along the resource allocation graph. Arrange the paths to the terminal graph nodes in ascending order of remaining number of intersections, ascending order of intersection interval width, and ascending order of resource usage. Read the refined tasks and edge computing node resources corresponding to the first path to obtain the task allocation results.

8. The method for downscaling wind turbine blade icing prediction based on multiple factors in complex mountainous terrain, as described in claim 7, is characterized in that: S5 includes: S51. Read the refinement task and corresponding edge resource vector according to the task allocation result. The edge computing node executes the refinement task within the allocated time period, writes the refinement value into the sub-element downscaling result, and deducts the edge resource vector according to the actual occupation to obtain the execution record. S52. Based on the execution record, update the corresponding Taylor model with the refined value and re-find the intersection point. Delete the tasks that no longer intersect with the result boundary, retain the tasks that still intersect with the result boundary, and then associate the updated intersection point with the source task to obtain the updated intersection point task set.

9. A method for downscaling wind turbine blade icing prediction based on multiple factors in complex mountainous terrain, as described in claim 8, is characterized in that: S5 also includes: S53. When the updated intersection task set is not empty and any remaining task has insufficient resource dimension, generate the resource availability interval according to the release time of each resource dimension of the edge resource vector, find the first common time period of each remaining task and the resource availability interval, read the time corresponding to the first task in ascending order of the common time period start as the start time of the next resource cycle, and write the updated intersection task set, Taylor model and sub-element downscaling results into the acceptance record. S54. In the next resource cycle, read the acceptance record, reconstruct the edge resource vector and re-execute the task allocation until the updated intersection task set is empty. Then, calculate the icing amount of the wind turbine blades and determine the icing warning level based on the sub-element downscaling results.

10. A component-based downscaling forecasting system for wind turbine blade icing in complex mountainous terrain, used to implement the component-based downscaling forecasting method for wind turbine blade icing in complex mountainous terrain as described in any one of claims 1-9, the system comprising a data mapping module, a component downscaling module, an intersection point identification module, a resource allocation module, and an iterative output module, characterized in that: The data mapping module is used to acquire weather forecast fields, map the wind turbine position and blade sweep height to model grid points based on elevation data, and extract meteorological data on a time basis; the edge computing node reads the processor idle time period, then reads the storage space and link bandwidth, and encodes the wind turbine input data and edge resource vector; The sub-item downscaling module is used to correct temperature and convert wet-bulb temperature based on elevation difference, and then correct wind speed based on slope aspect angle. After redistributing precipitation along the windward uplift path, it is converted to the blade sweep height. Combined with the wet-bulb temperature sign and the position of the zero-degree isotherm, the precipitation phase is divided to obtain the sub-element downscaling results. The intersection identification module is used to employ the Taylor model algorithm for the refined correction quantity, construct a second-order polynomial at the midpoint of the value interval, subtract the polynomial envelope from the envelope of the result function to form the remainder term interval, find the intersection point of the polynomial with the phase boundary value or the icing amount boundary value of the warning level, and associate the first intersection point with the refinement task to form the intersection task set. The resource allocation module uses a zero-suppression decision graph dynamic programming algorithm based on the intersection task set. The zero-branch skips the current task, and the one-branch deducts the edge resource vector and recalculates the intersection with the refined value. Merge graph nodes where the task set at the intersection and the edge resource vector are the same, delete graph branches with negative resource dimensions, first sort by the number of remaining intersections in ascending order, then sort by the resource occupancy in ascending order, and read the first and second paths to obtain the task allocation results; The iterative output module is used to execute refinement tasks according to the task allocation results, update the sub-feature downscaling results and edge resource vectors with refinement values, and reconstruct the intersection task set; When the edge resource vector cannot accommodate the remaining tasks, it will move to the next resource cycle until the intersection task set is empty, and output the icing amount of the wind turbine blades and the icing warning level.

Images