Ground-based laser radar near-ground wind field modeling method and system based on data fusion

Abstract

The application provides a ground-based laser radar near-ground wind field modeling method and system based on data fusion, and relates to the technical field of wind field modeling. The method comprises the following steps: firstly, using a ground-based laser radar to perform multiple scans on a target terrain area to obtain an observation data set; then, obtaining ground surface temperature data matched in space and time with the three-dimensional point cloud data, and fusing the three-dimensional point cloud data to generate a coupled data set; generating initial wind field feature vectors at different time instants based on the coupled data set; then, processing the vectors through a long short-term memory network to generate target wind field feature vectors; then, inputting the target vectors into a wind field inversion model constructed based on a high-order Taylor expansion polynomial to calculate three-dimensional wind field data and turbulence characteristic data; finally, constructing a near-ground wind field model of the target terrain area through an adaptive interpolation algorithm according to the above data. The application improves the fineness and accuracy of the near-ground wind field modeling of complex terrain.

Description

Technical Field

[0001] This application relates to the field of wind field modeling technology, and in particular to a ground-based lidar near-surface wind field modeling method and system based on data fusion. Background Technology

[0002] Ground-based lidar plays an important role in near-surface wind field observation, enabling the acquisition of wind measurement data with high spatiotemporal resolution. Its modeling methods have positive application prospects for wind energy resource assessment and atmospheric environment monitoring.

[0003] In existing technologies, to improve the accuracy of wind field modeling, related methods typically introduce multi-source data fusion. For example, some schemes combine lidar wind measurement data with meteorological station observations or numerical weather prediction data to achieve data complementarity, while other schemes focus on using time series analysis methods to process wind field data to capture the temporal variation characteristics of wind speed. These methods aim to construct more accurate and reliable wind field models.

[0004] However, existing methods still have room for improvement in their accuracy in depicting key physical processes of wind fields when modeling the fine structure of wind fields in complex terrain. Furthermore, they do not fully utilize the deep spatiotemporal correlations of multi-source heterogeneous data. Therefore, the level of detail in near-surface wind field modeling under complex terrain in existing technologies needs further improvement. Summary of the Invention

[0005] This application provides a ground-based lidar near-surface wind field modeling method and system based on data fusion, in order to solve the problems of low precision and low accuracy in the modeling of near-surface wind fields in complex terrain in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for modeling near-surface wind fields using ground-based lidar based on data fusion, comprising:

[0007] The target terrain area is scanned using at least one ground-based lidar with multiple scanning methods to obtain an observation dataset of the target terrain area;

[0008] Obtain land surface temperature data that is spatiotemporally matched with each 3D point cloud data in the observation dataset, and fuse the land surface temperature data with the corresponding 3D point cloud data to generate a coupled dataset;

[0009] Based on the coupled dataset, initial wind field feature vectors at different times are generated;

[0010] Based on the initial wind field feature vectors at different times, and combined with a long short-term memory network, a target wind field feature vector is generated.

[0011] The target wind field feature vector is input into the wind field inversion model, and the target wind field feature vector is processed by the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data. The wind field inversion model is constructed based on a high-order Taylor expansion polynomial.

[0012] Based on the three-dimensional wind field data and the turbulence feature data, an adaptive interpolation algorithm is used to construct a near-surface wind field model of the target terrain area.

[0013] Optionally, generating the target wind field feature vector based on the initial wind field feature vectors at different times, combined with a long short-term memory network, includes:

[0014] The initial wind field feature vectors at different times are sorted in chronological order to obtain the input feature set;

[0015] The input feature set is input into the memory unit of the long short-term memory network, and the memory unit calculates the unit state data at the current moment based on the input feature set and the historical feature information stored in the memory unit.

[0016] Based on the unit state data, generate the feature output for the current moment;

[0017] The feature outputs from multiple consecutive time points are merged and transformed to generate the target wind field feature vector.

[0018] Optionally, the memory unit calculates the current unit state data based on the input feature set and the historical feature information stored in the memory unit, including:

[0019] The memory unit receives the current input feature vector from the input feature set and inputs the input feature vector and the unit state data from the previous time to the first control module of the memory unit;

[0020] The first control module calculates the forgetting weight value based on the input feature vector and the unit state data of the previous time step, and performs a dot product operation between the forgetting weight value and the unit state data of the previous time step to obtain the first intermediate state data.

[0021] The input feature vector at the current moment and the unit state data at the previous moment are input to the second control module and the third calculation module in the memory unit. The second control module calculates the updated weight value, and the third calculation module calculates the candidate state data.

[0022] The state update module of the memory unit performs a dot product operation on the updated weight value and the candidate state data to obtain the second intermediate state data, and then adds the first intermediate state data and the second intermediate state data to obtain the unit state data at the current time.

[0023] Optionally, the step of inputting the target wind field feature vector into the wind field inversion model, and processing the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data includes:

[0024] The target wind field feature vector is input into the input module of the wind field inversion model, and the target wind field feature vector is analyzed to separate the first input parameter and the second input parameter.

[0025] The first input parameters are processed through the residual network module of the wind field inversion model to obtain the first deep feature parameters.

[0026] The second input parameters are weighted and fused using the attention mechanism module of the wind field inversion model to obtain the second deep feature parameters.

[0027] The first physical relationship function of the first calculation module in the wind field inversion model is used to calculate the first deep feature parameters and generate the first calculation result for each spatial point.

[0028] The second deep feature parameters are calculated using the second physical relationship function of the second calculation module in the wind field inversion model to generate the second calculation result for each spatial point.

[0029] The first calculation results of all spatial points are summarized to form three-dimensional wind field data, and the second calculation results of all spatial points are summarized to form turbulence characteristic data.

[0030] Optionally, the step of constructing a near-surface wind field model of the target terrain area based on the three-dimensional wind field data and the turbulence feature data using an adaptive interpolation algorithm includes:

[0031] Based on the digital elevation model of the target terrain region, determine the spatial grid of the area to be modeled;

[0032] The spatial coordinates of each spatial point in the three-dimensional wind field data are used as the first interpolation node, and the spatial coordinates of each spatial point in the turbulence feature data are used as the second interpolation node.

[0033] Based on the spatial density distribution of the first interpolation node and the second interpolation node, the interpolation calculation method for each spatial grid is adaptively selected;

[0034] Based on the interpolation calculation method, the three-dimensional wind speed vector and turbulence intensity information corresponding to each spatial grid are calculated respectively;

[0035] Based on the three-dimensional wind speed vector and turbulence intensity information of each spatial grid, a wind field model of the near-surface of the target terrain region is constructed.

[0036] Optionally, the step of fusing the surface temperature data with the corresponding three-dimensional point cloud data to generate a coupled dataset includes:

[0037] Each 3D point cloud data in the observation dataset is labeled as a spatial point, and the spatial coordinates of each spatial point are determined;

[0038] Based on the spatial coordinates of each spatial point, the corresponding temperature value is extracted from the surface temperature data, and based on the temperature value, a temperature attribute is assigned to each spatial point to form a spatial point set.

[0039] The spatial coordinates and temperature attributes of the spatial points in the three-dimensional point cloud data are jointly encoded to generate a coupled dataset.

[0040] Optionally, generating initial wind field feature vectors at different times based on the coupled dataset includes:

[0041] Extract the spatiotemporal subset corresponding to each time step in the coupled dataset, and perform the following steps for each spatiotemporal subset:

[0042] The spatiotemporal subset is input into a convolutional neural network, and the first feature extraction layer of the convolutional neural network generates the first spatial distribution feature of the spatial point set based on the spatial coordinates and three-dimensional point cloud data of the spatiotemporal subset.

[0043] The second feature extraction layer of the convolutional neural network is used to associate the first spatial distribution feature with the temperature attribute of the spatiotemporal subset, and the associated data is processed to generate the second spatial distribution feature.

[0044] The output layer of the convolutional neural network transforms the second spatial distribution features to output an initial wind field feature vector.

[0045] Secondly, this application provides a ground-based lidar near-surface wind field modeling system based on data fusion, comprising:

[0046] The scanning module is used to scan the target terrain area using at least one ground-based lidar through multiple scanning methods to obtain an observation dataset of the target terrain area;

[0047] The acquisition module is used to acquire surface temperature data that is spatiotemporally matched with each three-dimensional point cloud data in the observation dataset, and to fuse the surface temperature data with the corresponding three-dimensional point cloud data to generate a coupled dataset.

[0048] The first generation module is used to generate initial wind field feature vectors at different times based on the coupled dataset;

[0049] The second generation module is used to generate a target wind field feature vector based on the initial wind field feature vectors at different times, combined with a long short-term memory network.

[0050] The input module is used to input the target wind field feature vector into the wind field inversion model, and process the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data. The wind field inversion model is constructed based on a high-order Taylor expansion polynomial.

[0051] The construction module is used to construct a near-surface wind field model of the target terrain area based on the three-dimensional wind field data and the turbulence feature data, using an adaptive interpolation algorithm.

[0052] Thirdly, this application provides an electronic device, comprising:

[0053] Memory, used to store computer programs;

[0054] A processor is configured to execute the computer program to implement the steps of the ground-based lidar near-surface wind field modeling method based on data fusion as described in the first aspect above.

[0055] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the ground-based lidar near-surface wind field modeling method based on data fusion as described in the first aspect above.

[0056] This application provides a ground-based lidar near-surface wind field modeling method based on data fusion. This method offers the following advantages: First, by scanning the target terrain area using multiple methods, more comprehensive wind field observation data can be obtained. Second, fusing surface temperature data with the observation data allows the dataset to simultaneously contain wind and temperature information, enriching the physical dimensions of the data. Third, generating feature vectors based on the fused coupled dataset provides discriminative initial features for subsequent analysis. Fourth, combining long short-term memory networks to process feature vectors at different times helps capture the dynamic patterns of wind field evolution over time. Fifth, using an inversion model based on higher-order Taylor expansion to process the target feature vectors enables more accurate calculation of the detailed features of the three-dimensional wind field and turbulence. Finally, through an adaptive interpolation algorithm and based on the calculated wind field and turbulence data, a wind field model that better reflects the actual terrain distribution is constructed.

[0057] Furthermore, firstly, by organizing the temporal features in an orderly manner, a structured input is provided to the network; then, the memory unit combines historical and current information for calculation, so that the generated state data can effectively represent the dynamic continuity of the wind field; based on this state data, the output is generated, which retains the key information at each moment; finally, the outputs of multiple moments are merged and transformed to generate a target feature vector that can comprehensively reflect the temporal evolution law of the wind field.

[0058] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart illustrating a ground-based lidar near-surface wind field modeling method based on data fusion, provided for embodiments of this application;

[0061] Figure 2 A schematic diagram illustrating a specific implementation of a ground-based lidar near-surface wind field modeling method based on data fusion, provided in this application embodiment;

[0062] Figure 3 This is a schematic diagram of a ground-based lidar near-surface wind field modeling system based on data fusion, provided as an embodiment of this application. Detailed Implementation

[0063] For modeling near-surface wind fields in complex terrain, existing technologies typically combine lidar observation data with data from other meteorological sources and use traditional time-series analysis methods to construct models. However, when modeling areas with undulating terrain, there is still room for improvement in characterizing the internal physical processes of the wind field and utilizing the deep spatiotemporal correlations of multivariate data, resulting in a limited degree of reflection of the fine structure of the wind field in the final model.

[0064] To address the aforementioned issues, this application provides a ground-based lidar near-surface wind field modeling method based on data fusion. This method first fuses multi-scan mode observation data acquired by lidar with spatiotemporally matched surface temperature data to construct a dataset containing wind-temperature coupling information. Next, a temporal feature extraction network processes this dataset to capture the dynamic changes in the wind field over time. Then, an inversion model based on specific physical principles is used to calculate the three-dimensional wind field and turbulence data from deep features. Finally, adaptive spatial interpolation is used to construct the final wind field model. This method enhances the model's ability to characterize the three-dimensional structure and turbulence features of wind fields under complex terrain by collaboratively calculating multi-source observation data, temporal dynamic features, and physically constrained models, thereby improving the precision of near-surface wind field modeling.

[0065] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] The core of this application is to provide a method for modeling near-surface wind fields using ground-based lidar based on data fusion. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0067] Step 101: Use at least one ground-based lidar to scan the target terrain area using multiple scanning methods to obtain an observation dataset of the target terrain area.

[0068] In step 101, a ground-based lidar is an optical detection device installed on the ground. This device emits lasers into the atmosphere and receives signals reflected back by particles such as aerosols. It measures wind speed by analyzing changes in the signals. Multiple scanning modes refer to the lidar device using more than one spatial motion mode of the laser beam for measurement. For example, the laser beam can be swung up and down in the vertical direction to detect wind at different altitudes, or it can be rotated and scanned in the horizontal plane to detect wind in different horizontal directions.

[0069] The target terrain region is a specific geographical area that needs to be modeled for wind field, such as a mountainous or hilly area; the observation dataset refers to the collection of raw measurement data reflecting the wind field conditions of the target terrain region, which is collected by ground-based lidar in one or more scans.

[0070] In this embodiment of the application, a ground-based lidar set up in the target terrain area is controlled to work according to a variety of preset scanning modes. Then, the lidar emits lasers into the air and receives the echo signals reflected by the atmosphere. By processing and solving these echo signals, an observation dataset describing the wind field conditions in the area is finally obtained.

[0071] Step 102: Obtain surface temperature data that is spatiotemporally matched with each 3D point cloud data in the observation dataset, and fuse the surface temperature data with the corresponding 3D point cloud data to generate a coupled dataset.

[0072] Among them, surface temperature data refers to information that reflects the distribution of Earth's surface temperature, obtained through satellite remote sensing. Coupled dataset is a structured data set generated after joint coding processing, in which each data record contains both spatial location and temperature information.

[0073] In this embodiment, step 102 includes the following process:

[0074] Step 1021: Mark each three-dimensional point cloud data in the observation dataset as a spatial point and determine the spatial coordinates of each spatial point.

[0075] In step 1021, a spatial point is a logical identifier assigned to each original three-dimensional point cloud data in the observation dataset. A spatial point uniquely represents a specific wind measurement location with a three-dimensional position in space. Spatial coordinates are information used to accurately describe the location of this spatial point in geographic space, and usually include values ​​of three dimensions: longitude, latitude, and altitude.

[0076] In this embodiment of the application, the observation dataset is first traversed to identify each three-dimensional point cloud data; then each three-dimensional point cloud data is defined as a spatial point, and the geographical location information contained in the point is read, which constitutes the spatial coordinates of the spatial point.

[0077] In practical applications, if a 3D point cloud data contains location information of 123 degrees east longitude, 456 degrees north latitude, and 1895 meters above sea level, then the point is marked as a spatial point with spatial coordinates of [123, 456, 1895]. It should be understood that the above values ​​of east longitude, north latitude, and altitude, as well as other related values ​​given later, are all virtual and hypothetical geographical information, not real geographical information.

[0078] Step 1022: Extract the corresponding temperature value from the surface temperature data according to the spatial coordinates of each spatial point, and assign a temperature attribute to each spatial point based on the temperature value to form a spatial point set.

[0079] In step 1022, the temperature attribute refers to an additional data field added to each spatial point, the value of which is the surface temperature value corresponding to the spatial coordinates of that point; the spatial point set refers to the data set consisting of all spatial points that have been assigned the temperature attribute.

[0080] In this embodiment of the application, based on the spatial coordinates of each spatial point obtained in step 1021, a query is performed in the mapping relationship between geographic coordinates and temperature values ​​formed by the surface temperature data to find the temperature value recorded by the coordinate point closest to the spatial coordinate; then the queried temperature value is added as a new attribute data and associated with the corresponding spatial point; finally, all spatial points associated with the temperature attribute are gathered together to form the spatial point set.

[0081] In practical applications, if the coordinates of a spatial point obtained in step 1021 are [123, 456, 1895], then the point with coordinates closest to [123, 456] is found in the surface temperature data. The temperature value recorded at this point is 18.5 degrees Celsius. Therefore, an attribute is added to this spatial point, and the temperature attribute value is recorded as 18.5. After performing this operation on all spatial points, a spatial point set is formed.

[0082] Step 1023: Jointly encode the spatial coordinates and temperature attributes of the three-dimensional point cloud data and the spatial point set to generate a coupled dataset.

[0083] In this embodiment, the original three-dimensional point cloud data corresponding to each spatial point, its spatial coordinate value, and the temperature attribute value attached to it are written together into a new data record structure, wherein the data record structure is a complete record, and each data item is arranged in a predetermined field order; then this operation is performed on all points in the spatial point set, and finally all records are summarized to generate the final coupled dataset.

[0084] In practical applications, for the aforementioned spatial points [123, 456, 1895] and their corresponding 3D point cloud data and temperature attribute 18.5, they are encoded into a data record, which may contain fields such as "coordinates", "point cloud parameters", and "temperature". After performing this encoding on all spatial points, these records are combined into a table, thus generating the coupled dataset.

[0085] This application, through step 102, can effectively correlate wind field observation data from different sources with surface temperature data to form a structured coupled dataset, providing a unified and standardized data foundation for extracting wind field features with both spatial and temperature dimensions in subsequent steps.

[0086] Step 103: Based on the coupled dataset, generate initial wind field feature vectors at different times.

[0087] The initial wind field feature vector is a one-dimensional array consisting of multiple numerical elements. This array is used to quantify the initial state of the wind field in a local area of ​​the Earth's surface at a specific moment.

[0088] In this embodiment, step 103 includes the following process:

[0089] Step 1031: Extract the spatiotemporal subset corresponding to each time step in the coupled dataset, and perform the following steps for each spatiotemporal subset:

[0090] In step 1031, the spatiotemporal subset refers to a subset of data selected from the coupled dataset in which all data records have the same collection timestamp.

[0091] In this embodiment of the application, firstly, based on the time tag recorded by each data record in the coupled dataset, all data records belonging to the same acquisition time are grouped into a set to obtain a spatiotemporal subset; then, for each different acquisition time, this grouping operation is repeated to obtain multiple spatiotemporal subsets corresponding to different times.

[0092] In practical applications, if the coupled dataset contains data at three time points T1, T2, and T3, then all records at time T1 are taken as the spatiotemporal subset S1, all records at time T2 are taken as the spatiotemporal subset S2, and all records at time T3 are taken as the spatiotemporal subset S3.

[0093] Step 1032: Input the spatiotemporal subset into a convolutional neural network, and generate the first spatial distribution feature of the spatial point set based on the spatial coordinates and three-dimensional point cloud data of the spatiotemporal subset through the first feature extraction layer of the convolutional neural network.

[0094] In step 1032, the first feature extraction layer is a data processing layer in the convolutional neural network, which performs preliminary analysis and integration of spatial location information in the input data; the first spatial distribution feature is an intermediate calculation result, which reflects the aggregation, dispersion or arrangement pattern of spatial points in the spatiotemporal subset in three-dimensional space.

[0095] In this embodiment of the application, the spatiotemporal subset S1 is input into a convolutional neural network. The first feature extraction layer of the network reads and calculates the spatial coordinates and three-dimensional point cloud data of each spatial point in the spatiotemporal subset S1. The calculation process scans the input data through a preset convolutional kernel to extract the primary rules about the spatial positional relationships contained in the data, and then generates the first spatial distribution feature F1 to describe such spatial distribution rules.

[0096] In practical applications, the first feature extraction layer of this convolutional neural network may contain three 3×3 convolutional kernels. These kernels perform sliding calculations on the input data, and the extracted first spatial distribution feature F1 can be a feature set containing 128 feature maps.

[0097] Step 1033: Using the second feature extraction layer of the convolutional neural network, the first spatial distribution feature is associated with the temperature attribute of the spatiotemporal subset, and the associated data is processed to generate the second spatial distribution feature.

[0098] In step 1033, the second feature extraction layer is another data processing layer in the convolutional neural network located after the first feature extraction layer. The second spatial distribution feature is a more complex spatial feature representation generated by further integrating temperature attribute information on the basis of the first spatial distribution feature.

[0099] In this embodiment of the application, the second feature extraction layer of the convolutional neural network receives the first spatial distribution feature F1 output by the previous layer, and reads the temperature attribute value corresponding to each spatial point in the spatiotemporal subset S1; the layer performs a weighted combination operation on the elements in the first spatial distribution feature F1 and the corresponding temperature attribute value, which can capture the potential correlation between wind speed and temperature; after the operation is completed, the second spatial distribution feature F2 that integrates location and temperature information is finally generated.

[0100] In practical applications, the second feature extraction layer can perform the following operations: multiply each feature map in the first spatial distribution feature F1 element-wise with a weight matrix generated based on the temperature attribute value, and then process it through an activation function to finally output the second spatial distribution feature F2 with a dimension of 64.

[0101] Step 1034: The second spatial distribution features are transformed through the output layer of the convolutional neural network to output the initial wind field feature vector.

[0102] In step 1034, the output layer is the final structure of the convolutional neural network, responsible for mapping the complex features extracted from the previous layers into the final required output format.

[0103] In this embodiment of the application, the output layer of the convolutional neural network receives the second spatial distribution feature F2 and flattens the multidimensional feature data F2 into a one-dimensional long vector; then, a fully connected layer performs a linear transformation and compression on the long vector to adjust its dimension to a preset length; finally, a fixed-length initial wind field feature vector V1 representing the initial state of the wind field in the current spatiotemporal subset S1 is output.

[0104] In practical applications, this output layer can flatten the second spatial distribution feature F2 with a dimension of 64 into a one-dimensional vector with a length of 4096, and then pass it through a fully connected layer with 256 neuron units to finally output an initial wind field feature vector V1 with a length of 256.

[0105] In step 103, this application uses a convolutional neural network to automatically extract and generate feature vectors that can characterize the initial state of the wind field at different times from the coupled data, providing structured feature inputs for subsequent capture of the dynamic changes of the wind field.

[0106] Step 104: Based on the initial wind field feature vectors at different times, and combined with the long short-term memory network, generate the target wind field feature vector.

[0107] Among them, the target wind field feature vector is a one-dimensional feature array that has undergone time-series processing and can comprehensively reflect the evolution law of the wind field over a period of time.

[0108] In this embodiment, step 104 includes the following process, such as... Figure 2 As shown:

[0109] Step 1041: Sort the initial wind field feature vectors at different times in chronological order to obtain the input feature set.

[0110] In step 1041, the input feature set is an ordered list, in which the elements are multiple initial wind field feature vectors arranged in chronological order of their corresponding acquisition times.

[0111] In this embodiment of the application, firstly, according to the time labels corresponding to the multiple initial wind field feature vectors generated in step 103, these vectors are arranged in order from early to late time, and then the arranged vectors are combined into an ordered list to form the input feature set.

[0112] In practical applications, assume that step 103 generates three initial wind field feature vectors corresponding to times T1, T2, and T3, with vector values ​​V1, V2, and V3 respectively; after arranging them in time order T1, T2, and T3, the input feature set [V1, V2, V3] is obtained.

[0113] Step 1042: Input the input feature set into the memory unit of the long short-term memory network. The memory unit calculates the current unit state data based on the input feature set and the historical feature information stored in the memory unit.

[0114] The current cell state data is a vector generated by the state update module. This vector is obtained by fusing the historical state filtered by the forgetting weight and the candidate state weighted by the update weight. It combines historical information and current new information and represents the complete internal state of the memory cell after the update at the current moment.

[0115] Step 1042 may specifically include the following steps:

[0116] A1: The memory unit receives the input feature vector at the current time from the input feature set, and inputs the input feature vector and the unit state data at the previous time to the first control module of the memory unit.

[0117] In step A1, the memory unit is the core structure in the long short-term memory network responsible for storing and updating historical information. The first control module is a functional part inside the memory unit, which is used to determine how much past information needs to be forgotten. Historical feature information can be represented as the unit state data at the previous moment. This data is stored and transmitted inside the memory unit and carries all historical feature information up to the previous moment.

[0118] In this embodiment, the memory unit reads each feature vector in the input feature set in sequence; when processing the feature vector at the current moment, the memory unit inputs the feature vector together with the unit state data obtained after processing the data at the previous moment stored in itself into the first control module inside.

[0119] In practical applications, when processing the second feature vector V2 in the input feature set [V1, V2, V3], the memory unit inputs V2 together with the unit state data C1 stored after processing V1 into the first control module.

[0120] A2: The first control module calculates the forgetting weight value based on the input feature vector and the unit state data of the previous time step, and performs a dot product operation between the forgetting weight value and the unit state data of the previous time step to obtain the first intermediate state data.

[0121] In step A2, the forgetting weight value is a numerical vector ranging from 0 to 1. Each element in the vector represents the degree to which the historical information at the corresponding position needs to be retained, where 0 represents complete forgetting and 1 represents complete retention. The first intermediate state data is the intermediate result obtained by multiplying the forgetting weight value by the unit state data at the previous time step.

[0122] In this embodiment, the first control module calculates the received input feature vector and the unit state data of the previous time step. During the calculation, a specific function can be used to add the input feature vector and the unit state data of the previous time step after a linear transformation, and then process them through a nonlinear function to finally generate the forgetting weight value Ft. Then, the forgetting weight value Ft is multiplied element-wise with the unit state data Ct-1 of the previous time step to obtain the first intermediate state data.

[0123] In practical applications, assuming the previous unit state data C1 is [0.5, 0.8], and the forgetting weight value Ft calculated by the first control module for the input feature vector V2 is [0.9, 0.6], then the first intermediate state data is [0.5×0.9, 0.8×0.6], which is [0.45, 0.48].

[0124] A3: Input the current input feature vector and the previous unit state data into the second control module and the third calculation module in the memory unit. The second control module calculates the updated weight value, and the third calculation module calculates the candidate state data.

[0125] In step A3, the second control module is a structure in the memory unit used to determine how much current information needs to be added, and its output update weight value is a numerical vector with a value range between 0 and 1; the third calculation module is a structure used to generate new state information to be added based on the current input, and its output candidate state data is a vector.

[0126] The update weight value is a numerical vector generated by the second control module, with values ​​ranging from 0 to 1. Each element in the vector determines what proportion of the candidate state data at the corresponding position needs to be updated or added to the new unit state, where 0 indicates no addition and 1 indicates full addition.

[0127] Candidate state data is a vector generated by the third calculation module. This vector is calculated based on the input feature vector at the current time step and the unit state data at the previous time step. This data represents the candidate information contained in the current input that may become the components of the new state.

[0128] In this embodiment, the current input feature vector and the previous unit state data are simultaneously input to the second control module and the third calculation module. The second control module calculates the updated weight value It in a manner similar to that of the first control module but with different parameters. The third calculation module calculates the candidate state data Ct based on the same input using another function. This embodiment does not specifically limit the design of the parameters corresponding to these two calculation modules; they can be set according to actual conditions.

[0129] In practical applications, the updated weight value It calculated by the second control module based on V2 and C1 can be [0.7, 0.3]; the candidate state data Ct calculated by the third calculation module based on V2 and C1 may be [0.2, 0.9].

[0130] A4: The state update module of the memory unit performs a dot product operation on the updated weight value and the candidate state data to obtain the second intermediate state data, and adds the first intermediate state data and the second intermediate state data to obtain the unit state data at the current time.

[0131] In step A4, the state update module is the structure in the memory unit responsible for finally synthesizing the new state, and the second intermediate state data is the result of the dot product of the updated weight value and the candidate state data.

[0132] In this embodiment, the state update module first multiplies the update weight value It with the candidate state data Ct element by element to obtain the second intermediate state data; then it adds the first intermediate state data obtained in step A2 to this second intermediate state data element by element, and the resulting sum vector is the current unit state data Ct.

[0133] In practical applications, the updated weight value It[0.7, 0.3] is multiplied by the candidate state data Ct[0.2, 0.9] to obtain the second intermediate state data [0.14, 0.27]; this data is then added to the first intermediate state data [0.45, 0.48] to obtain the current cell state data Ct as [0.59, 0.75].

[0134] Step 1043: Based on the unit state data, generate the feature output for the current moment.

[0135] In step 1043, the feature output at the current moment is the information calculated based on the current internal state and passed from the memory unit to other parts of the external network.

[0136] In this embodiment, after calculating the cell state data Ct at the current time, the memory unit performs a nonlinear transformation on the cell state data Ct through an activation function, and then multiplies it by an output weight matrix to finally generate the feature output ht at the current time.

[0137] In practical applications, assuming the unit state data Ct is [0.59, 0.75], after passing through a hyperbolic tangent activation function and weight matrix transformation, the feature output ht can be generated as [0.56, 0.65].

[0138] Step 1044: Merge and transform the feature outputs from multiple consecutive time points to generate the target wind field feature vector.

[0139] In this embodiment, the feature outputs corresponding to each time step in the input feature set are concatenated into a long combined vector in chronological order; then, a fully connected layer is used to linearly transform the combined vector to compress its dimension to a preset length, and finally the target wind field feature vector is generated.

[0140] In practical applications, assuming the feature outputs at three time points are h1[0.5, 0.6], h2[0.56, 0.65], and h3[0.6, 0.7], they are concatenated to obtain [0.5, 0.6, 0.56, 0.65, 0.6, 0.7]. This is then transformed into a vector of length 3 [0.8, 1.1, 0.9] through a fully connected layer. This vector is the target wind field feature vector.

[0141] In step 104, this application uses a long short-term memory network to model the temporal wind field characteristics, and fuses the discrete features at different times to generate a comprehensive feature vector that can characterize the dynamic evolution of the wind field, providing deep feature information with time dependence for the final wind field inversion.

[0142] Step 105: Input the target wind field feature vector into the wind field inversion model, process the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data, wherein the wind field inversion model is constructed based on a high-order Taylor expansion polynomial.

[0143] Among them, the wind field inversion model is a mathematical model used to map characteristic information to specific wind field physical parameters; the higher-order Taylor expansion polynomial is a mathematical tool that uses power series containing higher-order derivative terms to approximate a complex functional relationship.

[0144] Three-dimensional wind field data is a dataset containing three-dimensional wind speed vector information for all spatial points; turbulence feature data is a dataset containing turbulence intensity information for all spatial points.

[0145] In this embodiment, step 105 includes the following process:

[0146] Step 1051: Input the target wind field feature vector into the input module of the wind field inversion model, analyze the target wind field feature vector, and separate the first input parameter and the second input parameter.

[0147] In step 1051, the input module is the data preprocessing part of the wind field inversion model; the first input parameter and the second input parameter are two sets of sub-data separated from the target wind field feature vector, which are used for subsequent calculation of three-dimensional wind speed and turbulence features, respectively.

[0148] In this embodiment of the application, firstly, the input module receives the target wind field feature vector, and then splits the vector into two independent data segments according to a predetermined parameter mapping rule; one part of the data segment is used as the first input parameter, and the other part of the data segment is used as the second input parameter.

[0149] In practical applications, assume that the target wind field feature vector is a vector of length 10 [0.8, 1.1, 0.9, 0.3, 0.6, 0.2, 0.5, 0.7, 0.4, 0.1]; according to the mapping rule, the first 6 elements are used as the first input parameter, i.e. [0.8, 1.1, 0.9, 0.3, 0.6, 0.2], and the last 4 elements are used as the second input parameter, i.e. [0.5, 0.7, 0.4, 0.1].

[0150] Step 1052: The first input parameters are processed through the residual network module of the wind field inversion model to obtain the first deep feature parameters.

[0151] In step 1052, the residual network module is a data processing network with a "short-circuit connection" structure, which can effectively train deeper networks and alleviate the gradient vanishing problem; the first deep feature parameters are data representations that contain more abstract and high-level information after being processed by the residual network module.

[0152] In this embodiment, the first input parameter is input into the residual network module. The module first performs nonlinear transformation on the input data through multiple convolutional layers and activation layers to extract preliminary features. Then, it adds the input data and the transformed features element by element through its unique short-circuit connection structure to achieve deep fusion and transfer of features. After processing by multiple such residual blocks, the final output data is the first deep feature parameter.

[0153] In practical applications, the residual network module can contain 3 residual blocks; after the first input parameter passes through the first residual block, a feature map is output; the feature map is then processed by the second residual block and added to its own input feature; similarly, after being processed by the third residual block, the final output first deep feature parameter can be a feature vector with a dimension of 256.

[0154] Step 1053: The second input parameters are weighted and fused using the attention mechanism module of the wind field inversion model to obtain the second deep feature parameters.

[0155] In step 1053, the attention mechanism module is a structure that can dynamically allocate different processing weights according to the relative importance of each part of the data; the second deep feature parameter is the feature data obtained after weighted fusion processing by the attention mechanism module.

[0156] In this embodiment, the second input parameter is input to the attention mechanism module. The module first calculates the correlation score between each data element in the second input parameter to generate an attention weight matrix. Then, the weight matrix is ​​multiplied by the original input data to enhance important features and suppress unimportant features. Finally, the weighted features are linearly transformed and aggregated to output the second deep feature parameters.

[0157] In practical applications, the second input parameter is a vector [0.5, 0.7, 0.4, 0.1]; the attention mechanism module may calculate the importance weights of each element as [0.2, 0.5, 0.2, 0.1]; then multiply the weights by the original vector to obtain [0.1, 0.35, 0.08, 0.01]; after passing through a fully connected layer, the second deep feature parameter is obtained as a vector of length 64.

[0158] Step 1054: Calculate the first deep feature parameters using the first physical relationship function of the first calculation module in the wind field inversion model to generate the first calculation result for each spatial point.

[0159] In step 1054, the first calculation module is the part of the wind field inversion model specifically used to calculate three-dimensional wind field data; the first physical relationship function is a specific calculation function constructed based on a higher-order Taylor expansion polynomial, which defines the mapping relationship from the first deep feature parameter to the three-dimensional wind speed; the first calculation result is a vector that contains the wind speed components of the spatial point in the three-dimensional direction.

[0160] In this embodiment of the application, the first deep feature parameter is input into the first calculation module. The module calls the first physical relationship function stored internally. For example, the specific expression of the function can be defined by a higher-order Taylor expansion polynomial. The function calculates with each element in the first deep feature parameter as the independent variable and finally outputs a vector containing three values. This vector is the first calculation result of the spatial point, where the three values ​​represent the magnitude of the wind speed in the three coordinate directions.

[0161] In practical applications, assume the first physical relation function has the following form: , where F represents the first calculation result, x and y are elements in the first deep feature parameters, and a, b, c, d, e, and f are polynomial coefficients.

[0162] Step 1055: Calculate the second deep feature parameters using the second physical relationship function of the second calculation module in the wind field inversion model to generate the second calculation result for each spatial point.

[0163] In step 1055, the second calculation module is the part of the wind field inversion model specifically used to calculate turbulence characteristic data; the second physical relation function is another calculation function that is constructed based on a higher-order Taylor expansion polynomial and is different from the first physical relation function. This function defines the mapping relationship from the second deep characteristic parameters to the turbulence characteristics; the second calculation result is a numerical value or vector that characterizes the turbulence intensity or energy distribution characteristics of the airflow at that spatial point.

[0164] The specific mathematical forms of the "first physical relation function" and the "second physical relation function" are constructed based on higher-order Taylor expansion polynomials. Here, the higher-order Taylor expansion polynomials serve as the basic mathematical model for constructing these physical relation functions, providing a specific, higher-order nonlinear computational framework for the functions. This enables these functions to handle complex mapping relationships between feature inputs and physical quantities. It should be noted that the embodiments of this application do not specifically limit the specific expressions of the above two physical relation functions.

[0165] In this embodiment, the second deep feature parameter is input into the second calculation module. The module calls the internally stored second physical relationship function, which is also defined by a higher-order Taylor expansion polynomial, but with different coefficients. The function calculates using the elements in the second deep feature parameter as independent variables and finally outputs a value to describe the turbulence intensity. This value is the second calculation result for the spatial point.

[0166] In practical applications, assume the second physical relation function has the following form: Where T is the second calculation result, u and v are elements in the second deep characteristic parameters, and p, q, r, s, and t are polynomial coefficients; then, after substituting the specific characteristic values, the second calculation result is 0.15, which can be a dimensionless value representing the turbulence intensity.

[0167] Step 1056: Summarize the first calculation results of all spatial points to form three-dimensional wind field data, and summarize the second calculation results of all spatial points to form turbulence characteristic data.

[0168] In this embodiment of the application, the first calculation results of all spatial points are organized into a list or matrix according to their spatial location to form the three-dimensional wind field data; at the same time, the second calculation results of all spatial points are organized in the same spatial order to form the turbulence characteristic data.

[0169] In practical applications, assuming there are 3 spatial points, their first calculation results are [3.2, -1.5, 0.8], [2.9, -1.2, 0.7], and [3.5, -1.8, 0.9], respectively. The three-dimensional wind field data formed after summarizing is the set of these three vectors; their second calculation results are 0.15, 0.13, and 0.18, respectively. The turbulence characteristic data formed after summarizing is the set of these three values.

[0170] In step 105, this application uses an inversion model to efficiently and accurately convert the deep parameters extracted from complex features into two-dimensional wind field and turbulence data with clear physical meaning, thus completing the final mapping from abstract features to specific wind field states.

[0171] Step 106: Based on the three-dimensional wind field data and the turbulence feature data, construct a near-surface wind field model of the target terrain area using an adaptive interpolation algorithm.

[0172] The wind field model is a dataset that covers the near-ground space of the target terrain area and contains three-dimensional wind speed and turbulence information at each spatial location.

[0173] In this embodiment, step 106 includes the following process:

[0174] Step 1061: Determine the spatial grid of the area to be modeled based on the digital elevation model of the target terrain area.

[0175] In step 1061, the digital elevation model is a digital map describing the distribution of surface elevation, and the spatial grid of the area to be modeled refers to a regular two-dimensional or three-dimensional cell array formed by dividing the geographical area covered by the digital elevation model according to certain latitude and longitude intervals or equal intervals.

[0176] The area to be modeled is determined directly based on the boundaries and geographical extent of the digital elevation model data of the target terrain area. That is, the area to be modeled is the geographical area covered by the digital elevation model and where the wind field model needs to be built. The target terrain area is a geographical concept, a macroscopic and task-oriented description. The area to be modeled specifically refers to the array of cells composed of regular spatial grids, which is divided based on the above geographical extent in the specific modeling steps. It is the specific object of subsequent spatial interpolation calculations.

[0177] In this embodiment of the application, the digital elevation model file of the target terrain area is first read, which defines the planar range of the area and the elevation of each location; then, according to the preset grid resolution, for example, at intervals of 100 meters, multiple square grid cells are divided in the planar range, and the set of positions of these grid cells on the horizontal plane is the spatial grid of the area to be modeled.

[0178] In practical applications, assuming the target terrain area is a rectangle with a planar range of 1000 meters east-west and 800 meters north-south, then with 100-meter intervals, 10 columns of grids will be divided in the east-west direction and 8 rows of grids in the north-south direction, forming a total of 80 spatial grids.

[0179] Step 1062: Use the spatial coordinates of each spatial point in the three-dimensional wind field data as the first interpolation node, and use the spatial coordinates of each spatial point in the turbulence feature data as the second interpolation node.

[0180] In step 1062, the first interpolation node is a spatial location point used for interpolating and calculating the three-dimensional wind field data, and its position is defined by the spatial coordinates recorded by each data point in the three-dimensional wind field data; the second interpolation node is a spatial location point used for interpolating and calculating the turbulence feature data, and its position is defined by the spatial coordinates recorded by each data point in the turbulence feature data.

[0181] In this embodiment of the application, the spatial coordinates corresponding to each data point are extracted from the three-dimensional wind field data, and the set of these coordinate points is marked as the first interpolation node; at the same time, the spatial coordinates corresponding to each data point are extracted from the turbulence feature data, and the set of these coordinate points is marked as the second interpolation node.

[0182] In practical applications, the three-dimensional wind field data contains data from three spatial points with coordinates of point A [100, 200], point B [300, 400], and point C [500, 600]. The first interpolation node is these three points. The turbulence feature data also contains data from these three points with the same coordinates, so the second interpolation node is also these three points.

[0183] Step 1063: Based on the spatial density distribution of the first interpolation node and the second interpolation node, adaptively select the interpolation calculation method for each spatial grid.

[0184] In step 1063, spatial density distribution refers to the spatial aggregation or dispersion of the first interpolation node or the second interpolation node. The interpolation calculation method includes, but is not limited to, inverse distance weighting method, Kriging method, etc.

[0185] In this embodiment of the application, for a spatial grid, if the number of first interpolation nodes within a certain radius is greater than a set threshold, the area is determined to be densely populated, and the computationally efficient inverse distance weighting method is selected as the interpolation method for its three-dimensional wind speed; if the number of nodes is less than or equal to the threshold, the area is determined to be sparsely populated, and the Kriging method, which considers spatial variability, is selected as the interpolation method; for the interpolation of turbulence characteristics, the same logic is used to select the corresponding interpolation method for each grid based on the density distribution of the second interpolation nodes.

[0186] In practical applications, the threshold radius is set to 200 meters and the threshold number of nodes is 5. Then, for a spatial grid with coordinates [150, 250], if 4 first interpolation nodes are found within its 200-meter radius, which is less than 5, the Kriging method is selected for its three-dimensional wind speed interpolation. At the same time, if 4 second interpolation nodes are found within this range, the Kriging method is also selected for its turbulence characteristic interpolation.

[0187] Step 1064: Calculate the three-dimensional wind speed vector and turbulence intensity information corresponding to each spatial grid according to the interpolation calculation method.

[0188] In step 1064, the three-dimensional wind speed vector is data that includes wind speed components in three directions and is assigned to each spatial grid, while the turbulence intensity information is a value that is assigned to each spatial grid and characterizes the degree of airflow pulsation.

[0189] In this embodiment of the application, the actual calculation process is performed according to the specific interpolation calculation method selected for each spatial grid in step 1063: For each spatial grid, the three-dimensional wind speed vector of the grid center point is obtained by using the coordinates of the first interpolation node around it and the corresponding three-dimensional wind speed data according to the formula of the selected interpolation method; at the same time, the turbulence intensity information of the grid center point is obtained by using the coordinates of the second interpolation node around it and the corresponding turbulence characteristic data according to the formula of the selected interpolation method.

[0190] In practical applications, taking a grid using the inverse distance weighting method as an example, it has three first interpolation nodes around it, and their three-dimensional wind speed vectors are respectively... Their distances from the center of the grid are respectively Then the three-dimensional wind speed vector of the grid. The calculation process is as follows: first, calculate the weight of the reciprocal of each distance; the weight... Then the weights are normalized to obtain... The wind speed vector of this grid The calculation process for turbulence intensity information is similar, using the corresponding turbulence values.

[0191] Step 1065: Based on the three-dimensional wind speed vector and turbulence intensity information of each spatial grid, construct a wind field model of the near-surface of the target terrain area.

[0192] In this embodiment of the application, a data record is created for each spatial grid, which contains three fields: grid number or center point coordinates, three-dimensional wind speed vector, and turbulence intensity information; then, the data records of all grids are aggregated to form a wind field model covering the entire target terrain area near the ground.

[0193] In practical applications, for each of the 80 spatial grids, a data record with the following structure is generated: {Grid number: G01, center point coordinates: [50, 50], three-dimensional wind speed vector: [2.5, -1.1, 0.3], turbulence intensity: 0.12}; finally, all 80 such records are stored in order in a database or file to complete the construction of the wind field model.

[0194] In step 106, this application adaptively expands the wind field and turbulence data at discrete points into a continuous, gridded model covering the entire modeling area, thereby providing a more complete and intuitive refined distribution map of the near-surface wind field.

[0195] Figure 3 A schematic diagram of a ground-based lidar near-surface wind field modeling system based on data fusion, provided in an embodiment of this application, is shown below. Figure 3 As shown, the system includes:

[0196] The scanning module 31 is used to scan the target terrain area using at least one ground-based lidar through multiple scanning methods to obtain an observation dataset of the target terrain area.

[0197] The acquisition module 32 is used to acquire surface temperature data that is spatiotemporally matched with each three-dimensional point cloud data in the observation dataset, and to fuse the surface temperature data with the corresponding three-dimensional point cloud data to generate a coupled dataset.

[0198] The first generation module 33 is used to generate initial wind field feature vectors at different times based on the coupled dataset.

[0199] The second generation module 34 is used to generate a target wind field feature vector based on the initial wind field feature vectors at different times and in combination with a long short-term memory network.

[0200] The input module 35 is used to input the target wind field feature vector into the wind field inversion model, and process the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data. The wind field inversion model is constructed based on a high-order Taylor expansion polynomial.

[0201] The construction module 36 is used to construct a near-surface wind field model of the target terrain area based on the three-dimensional wind field data and the turbulence feature data through an adaptive interpolation algorithm.

[0202] The ground-based lidar near-surface wind field modeling system based on data fusion in this application embodiment is used to implement the aforementioned ground-based lidar near-surface wind field modeling method based on data fusion. Therefore, the specific implementation of the ground-based lidar near-surface wind field modeling system based on data fusion can be found in the embodiment section of the ground-based lidar near-surface wind field modeling method based on data fusion mentioned above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0203] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described data fusion-based ground-based lidar near-surface wind field modeling methods.

[0204] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for modeling near-surface wind fields using ground-based lidar based on data fusion.

[0205] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0206] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the ground-based lidar near-surface wind field modeling method based on data fusion.

[0207] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0208] The foregoing has provided a detailed description of a ground-based lidar near-surface wind field modeling method and system based on data fusion, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for modeling near-surface wind fields using ground-based lidar based on data fusion, characterized in that, include: The target terrain area is scanned using at least one ground-based lidar with multiple scanning methods to obtain an observation dataset of the target terrain area; Obtain land surface temperature data that is spatiotemporally matched with each 3D point cloud data in the observation dataset, and fuse the land surface temperature data with the corresponding 3D point cloud data to generate a coupled dataset; Based on the coupled dataset, initial wind field feature vectors at different times are generated; Based on the initial wind field feature vectors at different times, and combined with a long short-term memory network, a target wind field feature vector is generated. The target wind field feature vector is input into the wind field inversion model, and the target wind field feature vector is processed by the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data. The wind field inversion model is constructed based on a high-order Taylor expansion polynomial. Based on the three-dimensional wind field data and the turbulence feature data, a near-surface wind field model of the target terrain area is constructed using an adaptive interpolation algorithm; The step of generating a target wind field feature vector based on the initial wind field feature vectors at different times, combined with a long short-term memory network, includes: The initial wind field feature vectors at different times are sorted in chronological order to obtain the input feature set; The input feature set is input into the memory unit of the long short-term memory network, and the memory unit calculates the unit state data at the current moment based on the input feature set and the historical feature information stored in the memory unit. Based on the unit state data, generate the feature output for the current moment; The feature outputs from multiple consecutive time points are merged and transformed to generate the target wind field feature vector. The memory unit calculates the current unit state data based on the input feature set and the historical feature information stored in the memory unit, including: The memory unit receives the current input feature vector from the input feature set and inputs the input feature vector and the unit state data from the previous time to the first control module of the memory unit; The first control module calculates the forgetting weight value based on the input feature vector and the unit state data of the previous time step, and performs a dot product operation between the forgetting weight value and the unit state data of the previous time step to obtain the first intermediate state data. The input feature vector at the current moment and the unit state data at the previous moment are input to the second control module and the third calculation module in the memory unit. The second control module calculates the updated weight value, and the third calculation module calculates the candidate state data. The state update module of the memory unit performs a dot product operation on the updated weight value and the candidate state data to obtain the second intermediate state data, and then adds the first intermediate state data and the second intermediate state data to obtain the unit state data at the current time.

2. The method according to claim 1, characterized in that, The step of inputting the target wind field feature vector into the wind field inversion model, and processing the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data includes: The target wind field feature vector is input into the input module of the wind field inversion model, and the target wind field feature vector is analyzed to separate the first input parameter and the second input parameter. The first input parameters are processed through the residual network module of the wind field inversion model to obtain the first deep feature parameters. The second input parameters are weighted and fused using the attention mechanism module of the wind field inversion model to obtain the second deep feature parameters. The first physical relationship function of the first calculation module in the wind field inversion model is used to calculate the first deep feature parameters and generate the first calculation result for each spatial point. The second deep feature parameters are calculated using the second physical relationship function of the second calculation module in the wind field inversion model to generate the second calculation result for each spatial point. The first calculation results of all spatial points are summarized to form three-dimensional wind field data, and the second calculation results of all spatial points are summarized to form turbulence characteristic data.

3. The method according to claim 1, characterized in that, The step of constructing a near-surface wind field model of the target terrain region based on the three-dimensional wind field data and the turbulence feature data using an adaptive interpolation algorithm includes: Based on the digital elevation model of the target terrain region, determine the spatial grid of the area to be modeled; The spatial coordinates of each spatial point in the three-dimensional wind field data are used as the first interpolation node, and the spatial coordinates of each spatial point in the turbulence feature data are used as the second interpolation node. Based on the spatial density distribution of the first interpolation node and the second interpolation node, the interpolation calculation method for each spatial grid is adaptively selected; Based on the interpolation calculation method, the three-dimensional wind speed vector and turbulence intensity information corresponding to each spatial grid are calculated respectively; Based on the three-dimensional wind speed vector and turbulence intensity information of each spatial grid, a wind field model of the near-surface of the target terrain region is constructed.

4. The method according to claim 1, characterized in that, The step of fusing the surface temperature data with the corresponding 3D point cloud data to generate a coupled dataset includes: Each 3D point cloud data in the observation dataset is labeled as a spatial point, and the spatial coordinates of each spatial point are determined; Based on the spatial coordinates of each spatial point, the corresponding temperature value is extracted from the surface temperature data, and based on the temperature value, a temperature attribute is assigned to each spatial point to form a spatial point set. The spatial coordinates and temperature attributes of the spatial points in the three-dimensional point cloud data are jointly encoded to generate a coupled dataset.

5. The method according to claim 1, characterized in that, The generation of initial wind field feature vectors at different times based on the coupled dataset includes: Extract the spatiotemporal subset corresponding to each time step in the coupled dataset, and perform the following steps for each spatiotemporal subset: The spatiotemporal subset is input into a convolutional neural network, and the first feature extraction layer of the convolutional neural network generates the first spatial distribution feature of the spatial point set based on the spatial coordinates and three-dimensional point cloud data of the spatiotemporal subset. The second feature extraction layer of the convolutional neural network is used to associate the first spatial distribution feature with the temperature attribute of the spatiotemporal subset, and the associated data is processed to generate the second spatial distribution feature. The output layer of the convolutional neural network transforms the second spatial distribution features to output an initial wind field feature vector.

6. A ground-based lidar near-surface wind field modeling system based on data fusion, characterized in that, include: The scanning module is used to scan the target terrain area using at least one ground-based lidar through multiple scanning methods to obtain an observation dataset of the target terrain area; The acquisition module is used to acquire surface temperature data that is spatiotemporally matched with each three-dimensional point cloud data in the observation dataset, and to fuse the surface temperature data with the corresponding three-dimensional point cloud data to generate a coupled dataset. The first generation module is used to generate initial wind field feature vectors at different times based on the coupled dataset; The second generation module is used to generate a target wind field feature vector based on the initial wind field feature vectors at different times, combined with a long short-term memory network. The input module is used to input the target wind field feature vector into the wind field inversion model, and process the target wind field feature vector through the wind field inversion model to obtain three-dimensional wind field data and turbulence feature data. The wind field inversion model is constructed based on a high-order Taylor expansion polynomial. The construction module is used to construct a near-surface wind field model of the target terrain area based on the three-dimensional wind field data and the turbulence feature data using an adaptive interpolation algorithm. The step of generating a target wind field feature vector based on the initial wind field feature vectors at different times, combined with a long short-term memory network, includes: The initial wind field feature vectors at different times are sorted in chronological order to obtain the input feature set; The input feature set is input into the memory unit of the long short-term memory network, and the memory unit calculates the unit state data at the current moment based on the input feature set and the historical feature information stored in the memory unit. Based on the unit state data, generate the feature output for the current moment; The feature outputs from multiple consecutive time points are merged and transformed to generate the target wind field feature vector. The memory unit calculates the current unit state data based on the input feature set and the historical feature information stored in the memory unit, including: The memory unit receives the current input feature vector from the input feature set and inputs the input feature vector and the unit state data from the previous time to the first control module of the memory unit; The first control module calculates the forgetting weight value based on the input feature vector and the unit state data of the previous time step, and performs a dot product operation between the forgetting weight value and the unit state data of the previous time step to obtain the first intermediate state data. The input feature vector at the current moment and the unit state data at the previous moment are input to the second control module and the third calculation module in the memory unit. The second control module calculates the updated weight value, and the third calculation module calculates the candidate state data. The state update module of the memory unit performs a dot product operation on the updated weight value and the candidate state data to obtain the second intermediate state data, and then adds the first intermediate state data and the second intermediate state data to obtain the unit state data at the current time.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the ground-based lidar near-surface wind field modeling method based on data fusion as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the ground-based lidar near-surface wind field modeling method based on data fusion as described in any one of claims 1 to 5.

Images