Earthwork calculation method and device for wind power plant road construction and storage medium

Abstract

The invention relates to the technical field of intelligent calculation of earth-rock work amount, and discloses an earth-rock calculation method and device for wind power plant road construction and a storage medium. Comprising the following steps: S1, acquiring topographic point cloud, engineering geology and meteorological time series data of a wind power plant construction area through sensing equipment; s2, constructing a topographic map structure, generating an adjacent matrix and extracting space-time diagram convolution features; s3, constructing and solving a dynamic differential model based on the features, and predicting the volume change of earthwork; s4, improving the solving efficiency by adopting a mixed precision heterogeneous calculation strategy; and S5, introducing a geomechanical constraint optimization prediction result, and outputting a construction area earth-rock distribution diagram. According to the method, the spatial structure characteristics of the complex terrain of the wind power plant are extracted by adopting the graph convolutional neural network, a high-fidelity modeling effect on the terrain form of the road construction area is realized, and the defect that terrain detail information is difficult to express is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent calculation of earthwork engineering quantities, and in particular to a method, device and storage medium for calculating earthwork quantities for wind farm road construction. Background Art

[0002] Earthwork calculation is a crucial step in wind farm construction. Road excavation and landfill during construction directly impact project costs, schedule, and safety. Therefore, accurately and in real time calculating earthwork volume changes becomes a critical issue in construction decision-making. With the expansion of wind farm construction and the increasing complexity of terrain, traditional earthwork calculation methods are no longer sufficient. To better support wind farm construction planning, we are developing a new method for calculating earthwork volume.

[0003] Existing technologies for calculating earthwork volume for wind farm road construction often rely on traditional numerical methods or image processing techniques based on elevation data. These methods primarily rely on existing topographic maps, digital elevation models (DEMs), and simplified linear calculation models to provide a preliminary estimate of the terrain. These models can generate basic earthwork data and assist in planning construction routes. These methods can provide a reference for wind farm construction, particularly in areas with relatively regular or uniform terrain, enabling relatively rapid calculation of earthwork volume.

[0004] However, existing technologies have exposed many shortcomings in practical applications. First, traditional methods ignore the influence of dynamic factors such as soil density and slope, and cannot accurately reflect the changes in earthwork volume during construction. In complex terrain environments, changes in soil density and slope directly affect the difficulty of excavation and landfill, and existing methods cannot handle these nonlinear changes, resulting in large calculation errors. Second, existing calculation methods have excessively high demands on computing resources. Especially in large-scale construction projects, the calculation speed and efficiency of traditional methods cannot meet the needs of real-time prediction, causing delays or waste of resources during construction. Finally, existing technologies generally ignore physical constraints, especially geological constraints. Changes in earthwork volume should conform to certain geological laws. Traditional methods fail to introduce these constraints into the calculation, resulting in model outputs that do not meet actual construction conditions. To this end, those skilled in the art have proposed a method, device, and storage medium for calculating earthwork for wind farm road construction to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a method, device and storage medium for calculating earthwork for wind farm road construction, which solves the problems in the existing technology of difficulty in coping with dynamic changes in complex terrain, insufficient prediction accuracy and lack of physical constraints.

[0006] To achieve the above object, the present application is implemented by the following technical solutions: A wind farm road construction earthwork calculation method comprises the following steps: S1, collecting multi-source data of the wind farm construction area through a perception device, including terrain point cloud data, engineering geological data and meteorological time series data; S2, constructing a terrain graph structure based on the multi-source data, generating a terrain adjacency matrix and performing spatio-temporal graph convolution feature extraction; S3, based on the spatio-temporal graph convolution feature, constructing a dynamic differential model of earthwork volume change, and solving by a neural differential equation; S4, using a hybrid precision heterogeneous computing strategy, efficiently solving the dynamic differential model to obtain the prediction result of the earthwork volume; S5, combining the geomechanics constraint condition, optimizing the prediction result, and outputting the earthwork distribution map of the wind farm road construction area.

[0007] Preferably, the step S1 comprises: The multi-source data includes terrain point cloud data collected by a laser radar device, geological data obtained by an exploration device, and meteorological data collected by a meteorological sensor; The collected data is standardized, including normalizing the terrain point cloud data, structuring the geological data, and time serializing the meteorological data; The processed data is encoded as a structured tensor as the input of the subsequent model.

[0008] Preferably, the step S2 comprises: Based on the terrain point cloud data, a graph structure is constructed, and the adjacency matrix of the graph structure is defined by calculating the spatial distance and the slope gradient; The spatio-temporal features are extracted by multi-order spectral domain graph convolution, and the graph convolution operation extracts high-dimensional features of the nodes for subsequent modeling; the adjacency matrix A ij satisfies the following relationship: Where: p i , p j is the three-dimensional coordinates of the i, j nodes, σ is the parameter of the Gaussian kernel function, is the slope gradient between nodes.

[0009] Preferably, the step S3 comprises: Based on the graph convolution feature, a differential model of earthwork volume change is constructed, and the change of earthwork volume is jointly affected by the slope gradient and the construction equipment efficiency; The model is solved by neural differential equations to obtain the time-evolving earthwork volume change value; The volume change rate satisfies the following relationship: Where: V is the volume of earth and stone, is the slope gradient, E(t) is the amount of change in equipment effectiveness over time, and α(t) and β(t) are time functions generated based on graph convolution features and equipment data.

[0010] Preferably, the time functions α(t) and β(t) are calculated by the following formulas: α(t)=softplus(W α ·H t ),β(t)=sigmoid(W β ·[H t ||E(t)]); Where: W α ,W β is the trainable weight matrix, H t is the expression of graph convolution features at time t, and [·||·] represents the feature concatenation operation.

[0011] Preferably, step S4 includes: Dynamically determine the current calculation accuracy based on the sparsity of graph convolution features; When the feature sparsity is greater than 0.7, 16-bit floating point numbers are used for calculation; When feature sparsity is between 0.3 and 0.7, 32-bit floating point numbers are used for calculation; When the feature sparsity is less than 0.3, 64-bit floating point numbers are used for calculation; The tasks are divided by the GPU and CPU of the computing platform to perform graph convolution operations and dynamic equation solving tasks respectively.

[0012] Preferably, step S5 includes: A geological constraint term is introduced into the model loss function to physically regularize the earthwork volume gradient. Optimize volume prediction results by combining node soil density and terrain slope angle information; Output the earthwork distribution map of the wind farm road construction area; The formula for the geological constraint term is as follows: in: is the loss value of the geological constraint term; λ is the regularization term coefficient, ρ i is the soil density of the ith node, g is the acceleration of gravity, θ i is the slope angle, is the volume gradient of the node; N is the total number of graph nodes involved in the calculation.

[0013] Preferably, the step S5 further comprises: Generate a 3D earthwork volume distribution map, where the earthwork volume of each voxel is processed by combining convolutional features and volume estimation values ​​into a multi-layer perceptron. The voxel volume distribution map is generated by the following formula: in: Represents the volume of earth and stone at the three-dimensional coordinate position (x, y, z), H k is the k-th layer graph convolution feature, V k is the volume prediction value, w k is the fusion weight of the convolution result; MLP([·]) means concatenating the graph features and the predicted volume and inputting them into the multi-layer perceptron network for feature transformation and fusion processing; K is the total number of layers of the graph convolution network.

[0014] A device for calculating earthwork for road construction in a wind farm, comprising the following modules: Acquisition module, used to collect terrain point cloud data, geological data and meteorological data; Graph modeling module, used to construct topographic map structure and perform graph convolution feature extraction; Dynamic modeling module, used to build a dynamic differential model of earthwork volume changes and solve neural differential equations; Computation scheduling module, used to execute mixed-precision heterogeneous computing strategies; Optimization module, used to introduce geological constraints to optimize prediction results; The output module is used to generate and output the earthwork distribution map of the wind farm road construction.

[0015] A storage medium stores a computer program, which, when executed by a processor, is used to implement the above-mentioned earthwork calculation method for wind farm road construction.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This paper uses a graph convolutional neural network to extract the spatial structural characteristics of the complex terrain of the wind farm, achieving high-fidelity modeling of the terrain morphology of the road construction area. Compared with the existing solutions that only build terrain models based on DEM elevation or linear interpolation, it breaks through the calculation error problem caused by local distortion of elevation data and effectively solves the problem of difficulty in expressing terrain details.

[0017] 2. By introducing a neural differential equation to dynamically model the earthwork volume change, the present application has the ability to time-series predict the volume change trend in the construction phase. Compared with the traditional estimation technology based on static profile or section deduction method, it is more suitable for the construction situation under nonlinear and multiple disturbance conditions, and avoids the problems of insufficient dynamic deduction accuracy and response lag of the original scheme.

[0018] 3. The present application adds a regularization constraint term based on geophysical properties in model training, effectively improving the physical consistency of the prediction results. In existing methods, the effects of density and slope and other geological variables are generally ignored, resulting in drift or abnormal fluctuations in predicted values. The present application is a structural optimization for this defect that deviates from the physical boundary.

[0019] 4. The present application uses a mixed precision heterogeneous computer mechanism to complete the model solving process, so that large-scale wind farm data can be efficiently processed under limited resources. Traditional earthwork calculation tools often rely on high-precision full-process floating-point operations, which can easily form a calculation bottleneck. However, the present scheme successfully avoids the problem of response delay when high-performance computing resources are scarce. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The present application is a method flowchart; Figure 2 The present application is a device architecture schematic diagram. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] Please refer to the drawings of the present application Figure 1 The present application provides a wind farm road construction earthwork calculation method, comprising the following steps: S1, collect multi-source data of the wind farm construction area through sensing devices, including terrain point cloud data, engineering geological data and meteorological time series data; Specifically, in the wind farm road construction earthwork calculation method of the present application, first, multi-source data of the wind farm construction area is collected. This step is the starting point of the entire calculation method, and the purpose is to obtain various basic data required for subsequent modeling and calculation. Generally, multi-source data includes terrain point cloud data, engineering geological data and meteorological time series data, etc. These data provide necessary spatial information, geological information and meteorological conditions for further earthwork calculation.

[0023] In one possible implementation, the multi-source data are collected through different sensing devices. Specifically, terrain point cloud data can be obtained through a laser radar device (LIDAR), which accurately measures the surface height and shape of the construction area by emitting a laser beam and receiving the reflected signal. Engineering geological data can be obtained through exploration equipment (such as drilling equipment) to record important geological information such as the soil type, soil thickness, and rock and soil layers in the area. Meteorological data can be collected through meteorological sensors installed in the wind farm area, which usually includes information such as temperature, humidity, wind speed, and wind direction. These data help to consider the impact of weather factors on changes in earthwork volume during construction.

[0024] In this embodiment, the terrain point cloud data collected by the LiDAR equipment can be formatted through post-processing and converted into a standard set of three-dimensional coordinate points. After collection, the geological data undergoes structured processing to encode the characteristics of each geological layer into numerical values ​​suitable for calculation, such as soil density and permeability. Meteorological data is then converted into time series to reflect dynamic changes in the weather during construction.

[0025] Specifically, in this embodiment, this processed data serves as input to subsequent models, forming a structured tensor. A tensor is a multidimensional array structure that can effectively represent different types of data and facilitates input into subsequent graph convolutional networks. Once structured, the tensor can be used to extract spatiotemporal features through graph neural networks, thereby supporting subsequent earthwork volume calculations.

[0026] As an option, the standardization process for multi-source data includes normalizing topographic point cloud data to eliminate bias caused by differences in measurement accuracy. Furthermore, geological data can be converted into numerical forms related to spatial features through structured encoding to facilitate subsequent graph convolution operations. Meteorological data can be converted into time series format for further smoothing to remove outliers and improve data usability and accuracy.

[0027] In a possible implementation, the terrain point cloud data normalization in the above-mentioned normalization process can be performed using the following formula: Where: X is the original terrain data; μ is the mean of the data; σ is the standard deviation; X norm The data are normalized.

[0028] This process can effectively reduce the impact of different acquisition devices and measurement environments on data accuracy.

[0029] In some embodiments, the structured encoding of geological data can employ numerical methods based on soil characteristics. For example, the physical properties of different soil types and rock layers (such as density and moisture) are converted into numerical parameters and mapped to specific value ranges using pre-defined rules. Furthermore, the time series processing of meteorological data may include interpolation of wind speed, temperature, humidity, and other data collected by sensors to fill in missing data and smooth the time series data, thereby ensuring the continuity and consistency of meteorological data.

[0030] In addition, after the multi-source data in this embodiment is standardized and structured, the structured tensor generated will contain various forms of input data and be used for graph convolution in subsequent steps. This tensor not only covers information such as spatial coordinates, geological features, and meteorological data, but is also compatible with different types of data, facilitating subsequent multimodal learning and feature extraction.

[0031] S2: Constructing a topographic map structure based on multi-source data, generating a topographic adjacency matrix, and performing spatiotemporal graph convolution feature extraction. Specifically, in the present invention's earthwork calculation method for wind farm road construction, step S2 primarily involves constructing a corresponding topographic map structure based on the multi-source data collected in the previous step, further generating a topographic adjacency matrix, and performing spatiotemporal graph convolution feature extraction. This step converts the original topographic point cloud data and other auxiliary data into a structured data format suitable for graph convolutional network processing, providing a data foundation for the subsequent construction of an earthwork volume change model.

[0032] Typically, this step involves multiple data sources and processing steps to ensure that the constructed topographic map reflects the topographic characteristics of the construction area while also accounting for the impact of geological conditions and meteorological factors on the construction process. During the graph construction process, the raw data must first be effectively structured to generate a graphical adjacency matrix that represents the spatial relationships between topographic nodes. Graph convolution is then performed on this basis to extract deep spatiotemporal features for subsequent model solving.

[0033] In this embodiment, a topographic map structure is first constructed based on the topographic point cloud data collected in step S1. Specifically, the topographic point cloud data is obtained by a lidar device and contains three-dimensional coordinate information of the wind farm area. On this basis, the topographic map structure can be represented as an undirected graph, in which each node represents a specific geographical location, and the characteristics of the node include the spatial coordinates of the location and the terrain attributes related to it (such as elevation, slope, etc.). The edges in the graph represent the adjacency relationship between nodes, and the adjacency matrix is ​​defined by the spatial distance between nodes or other geographical features (such as slope gradient).

[0034] Specifically, the terrain adjacency matrix A ijIt is defined based on the combination of spatial distance and slope gradient. The adjacency matrix can calculate the similarity between nodes through the Gaussian kernel function to determine which nodes are adjacent in the graph. For example, the similarity A between node i and node j is ij It can be expressed by the following formula: Where: p i ,p j are the three-dimensional coordinates of the i-th and j-th nodes; σ is the parameter of the Gaussian kernel function; is the slope gradient between node i and node j. This formula can combine the effects of spatial distance and slope gradient to generate an adjacency relationship that conforms to the terrain characteristics.

[0035] Alternatively, the adjacency matrix can be generated using other algorithms, such as those based on topological relationships or watershed delineation, selecting nodes with stronger spatial associations as adjacent nodes. Depending on the specific conditions of the construction area, the adjacency matrix construction method may vary, but it must ensure that it accurately describes the spatial distribution of the terrain.

[0036] In one possible implementation, after constructing the graph structure, graph convolution operations further extract spatiotemporal features. Graph convolutional neural networks (GCNs) can perform convolution operations on the graph structure, extracting high-dimensional features for each node. By stacking multiple layers of graph convolution layers, long-range dependencies between nodes can be effectively captured, and the feature information of each node can be fused for subsequent earthwork volume prediction.

[0037] In some embodiments, the graph convolution operation can be defined as follows: in: is the normalized adjacency matrix; H l is the node feature matrix of the lth layer; W l is the weight matrix of the lth layer; σ is the activation function. The output feature matrix H of each layer l+1 It will serve as the input of the next layer of convolution operation, and finally obtain the spatiotemporal feature representation of each node.

[0038] Specifically, the core of the graph convolution operation is to update node features through the weighted sum of the adjacency matrix. This process can effectively fuse the feature information of adjacent nodes in the graph, so that the representation of each node not only contains its own features, but also takes into account the relevant information with adjacent nodes, thereby providing richer input data for model construction in subsequent steps.

[0039] In one possible implementation, the spatiotemporal graph convolution operation can also be combined with other feature inputs, such as geological and meteorological data. At each node in the graph structure, corresponding geological characteristics (such as soil density and permeability) and meteorological data (such as wind speed and temperature) can be used as additional features to further enhance the graph convolutional network's ability to perceive terrain and meteorological changes. In this way, rather than relying solely on terrain data, the overall model can more comprehensively reflect the changes in earthwork in the wind farm area.

[0040] In some embodiments, after graph convolutional feature extraction, feature dimensionality reduction or merging can be further performed through a fully connected layer or pooling layer to provide a more concise feature representation for the subsequent earthwork volume change model. Ultimately, the high-dimensional spatiotemporal features extracted by the graph convolutional network will serve as input data for subsequent modeling and enter the next step of solving dynamic differential equations.

[0041] S3, based on the convolutional features of the spatiotemporal graph, constructs a dynamic differential model of earthwork volume changes and solves it through neural differential equations; Specifically, in the earthwork calculation method for wind farm road construction described in this invention, the primary task of step S3 is to construct a dynamic differential model of earthwork volume changes based on the graph convolution features extracted in the previous steps and solve it using a neural differential equation. This step is the core of the entire calculation process, predicting the dynamic changes in earthwork volume during wind farm road construction based on the spatiotemporal features provided by the graph convolutional network. This process allows for a precise estimation of earthwork volume changes at each moment during construction, providing support for construction decision-making.

[0042] Generally speaking, changes in earthwork volume are not only affected by topographical variations but are also closely related to the efficiency of the equipment used during construction and the construction methods used. Therefore, the model in step S3 needs to consider the combined effects of slope gradient and construction equipment efficiency on earthwork volume changes. The spatiotemporal features provided by graph convolutional networks play a crucial role in this process, integrating multiple factors such as topography and weather to accurately and dynamically predict earthwork volume.

[0043] In this embodiment, when constructing a differential model for earthwork volume change, it is first necessary to combine the spatiotemporal features extracted by the graph convolutional network with other influencing factors during the construction process (such as the efficiency of construction equipment). By dynamically analyzing the slope gradient of the terrain and the efficiency of construction equipment over time, a differential equation describing the change in earthwork volume is constructed. Specifically, the rate of change of earthwork volume can be expressed as follows: Where: V is the volume of earth and rock; is the slope gradient; E(t) is the amount of change in equipment effectiveness over time; α(t) and β(t) are time functions generated based on graph convolution features and equipment data, respectively representing the degree of influence of terrain and equipment effectiveness on volume change.

[0044] Specifically, the slope gradient It is one of the key factors affecting the volume change of earthwork, especially in wind farm construction, where the ups and downs of the terrain directly determine the amount of earthwork to be excavated and filled. The slope gradient reflects the slope of the terrain and is usually calculated as follows: Wherein: z is the height value; x and y are the coordinates of the horizontal coordinate axis. In some embodiments, the calculation of the slope gradient can also be based on other factors, such as local elevation changes, or a higher-order numerical method can be used for more precise calculation.

[0045] Alternatively, in the differential equation for the rate of change of earthwork volume, the time-varying variation in equipment efficiency, E(t), is considered a key factor influencing construction efficiency. Equipment efficiency is typically determined by the operating status and operating efficiency of construction machinery, as well as the synergy between equipment. The effect of time-varying equipment efficiency on earthwork volume can be described by the following formula: Among them: the type of equipment and the mode of operation affect the equipment's operating capacity, while the construction schedule is closely related to the progress of construction.

[0046] Through the joint modeling of these factors, the role of equipment efficiency in the change of earthwork volume can be accurately reflected.

[0047] In one possible implementation, α(t) and β(t) are dynamically generated time functions based on graph convolution features and equipment data. Specifically, these time functions can be dynamically adjusted based on changes in graph convolution features, reflecting the time-varying effects of terrain and equipment performance. These time functions can be learned using neural network models. For example, by using temporal modeling methods such as LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit), the effects of terrain changes and equipment performance changes on earthwork volume can be learned.

[0048] In some embodiments, the generation process of the time functions α(t) and β(t) can be performed as follows: α(t)=softplus(W α ·H t ),β(t)=sigmoid(W β ·[H t ||E(t)]); Where: Wα and W β is the trainable weight matrix, H t is the representation of the graph convolution feature at time t, and [·||·] represents the feature concatenation operation.

[0049] These time functions learn the relationship between graph convolution features and equipment performance, enabling the model to automatically adjust its prediction ability for earthwork volume changes according to different time states.

[0050] Specifically, the neural differential equation solving method employed in step S3 utilizes a neural network to approximate the solution of differential equations. This method flexibly adjusts the model output based on input data and parameters, thereby dynamically predicting earthwork volume changes. Compared to traditional numerical solutions, the neural differential equation method offers higher accuracy and greater adaptability, particularly when dealing with complex, nonlinear earthwork volume changes, providing more accurate predictions.

[0051] S4, adopts a mixed precision heterogeneous computing strategy to efficiently solve the dynamic differential model and obtain the predicted results of earthwork volume; Specifically, in the earthwork calculation method for wind farm road construction presented in this invention, step S4 employs a mixed-precision heterogeneous computing strategy to efficiently solve the previously constructed dynamic differential model, thereby obtaining a predicted earthwork volume. This step, building on the previous graph convolution feature extraction and earthwork volume change model, further improves computational efficiency, ensuring that large-scale data processing can be completed within a reasonable timeframe.

[0052] In general, efficient computational strategies are key to ensuring that earthwork calculation methods can operate in real time during actual wind farm construction. Traditional numerical calculation methods often face the problem of excessive computational effort, especially when dealing with complex spatiotemporal data, where computational complexity increases significantly. To overcome this issue, this step employs a mixed-precision heterogeneous computing strategy, optimizing computational performance by selecting different precisions for different computational tasks.

[0053] The core concept of the mixed-precision heterogeneous computing strategy in this embodiment is to dynamically adjust computational precision based on the sparsity of the graph convolutional network's output features. Specifically, when the graph convolutional network's output features are sparse, lower-precision computation is used; when the features are dense, higher-precision computation is used. This approach significantly reduces computing resource consumption and improves computational efficiency.

[0054] Specifically, when solving the earthwork volume change model, the sparsity of the features output by the graph convolutional network must first be analyzed. When sparsity is high, most values ​​are zero or close to zero, so 16-bit floating-point numbers can be used for calculations. When sparsity is low, the values ​​within the features are more widely distributed, requiring higher precision, so 32-bit or 64-bit floating-point numbers are used. This method of dynamically adjusting calculation precision based on data features significantly reduces computational costs while maintaining accuracy.

[0055] Alternatively, the specific accuracy selection can be determined based on the sparsity metric. When the sparsity of the graph convolutional features is greater than 0.7, 16-bit floating-point numbers are used for calculation; when the sparsity is between 0.3 and 0.7, 32-bit floating-point numbers are used; when the sparsity is less than 0.3, 64-bit floating-point numbers are used. The accuracy selection criteria can be adjusted based on the specific application scenario to ensure a balance between computational efficiency and result accuracy.

[0056] One possible implementation, to further improve computing performance, employs a heterogeneous computing architecture, where different computing tasks are assigned to the GPU and CPU. In this architecture, the GPU handles large-scale parallel computing tasks, particularly graph convolution operations and spatiotemporal feature extraction, while the CPU handles more complex tasks, such as solving dynamic differential equations. This optimized computing architecture fully leverages the strengths of various hardware types, significantly accelerating the model solution process.

[0057] Specifically, heterogeneous computing can be managed through the GPU acceleration module in the deep learning framework. During the graph convolution calculation phase, the GPU uses its parallel processing capabilities to quickly calculate the features of each node in the graph. This parallelism can significantly improve calculation speed, especially in multi-layer graph convolutional networks. During the differential equation solution phase, since this process involves relatively complex numerical calculations, the CPU performs high-precision calculations to ensure model accuracy and stability.

[0058] In some embodiments, the GPU and CPU of a heterogeneous computing platform work together to maximize the advantages of computing resources. For example, in the graph convolution layer, the GPU is primarily responsible for processing the adjacency matrix and node feature calculations, while in the differential equation solution process, the CPU is responsible for more accurate numerical approximation and iterative calculations to ensure accurate and reliable earthwork volume prediction results.

[0059] As a possible implementation, the division of computing tasks can be dynamically adjusted based on the system's real-time feedback on the computing load. When the load is high, the system can shift some computing tasks to idle computing units, thereby avoiding computing bottlenecks and further improving computing efficiency.

[0060] Specifically, during this step, by combining calculations of varying precision with a heterogeneous computing architecture, computation time can be significantly shortened, efficiency improved, and the accuracy and stability of the results ensured. The collaborative operation of the GPU and CPU on a heterogeneous computing platform not only enhances computational flexibility but also adapts to large-scale data processing needs, thereby ensuring the real-time and reliability of the earthwork calculation method for wind farm road construction.

[0061] S5, combined with geomechanical constraints, optimizes the prediction results and outputs the earthwork distribution map of the wind farm road construction area.

[0062] Specifically, in a method for calculating earthwork for wind farm road construction proposed in the present invention, after completing the efficient solution of the dynamic differential model, step S5 is mainly responsible for the physical consistency correction and visual expression of the earthwork volume prediction results. Based on the aforementioned graph convolution feature extraction, differential modeling, neural differential solution and mixed precision calculation, this step physically regularizes the model output by introducing geological constraints, thereby improving the geological credibility of the volume prediction, and optimizing the local prediction error in combination with parameters such as node soil density and slope angle. In addition, based on the output of the graph neural network and the multi-layer perceptron structure, a three-dimensional earthwork volume distribution map for the construction area is generated, providing intuitive support for construction path design and engineering quantity estimation.

[0063] In this example, to ensure that the prediction results are consistent with the actual geological structure, a geological constraint term is introduced into the loss function of the model training. This constraint term is designed based on the consistency between the earthwork volume gradient and the gravity distribution. It is used to constrain the direction of volume change to be consistent with the direction of the actual terrain slope and adjust the amplitude of the change through geophysical parameters. The specific formula is as follows: in: is the loss value of the geological constraint term; λ is the regularization term coefficient, which is usually set based on training stability through experience or cross-validation; represents the volume gradient corresponding to the i-th node; ρ i is the soil density at the node; g is the gravitational acceleration constant; θ i is the slope angle at that location; N represents the total number of graph nodes involved in the calculation.

[0064] Generally, the direction of the volume gradient should be consistent with the slope direction, and the magnitude of the change is limited by the local soil density and slope angle. By introducing this regularization term, it is possible to effectively suppress non-physical changes that may occur during the prediction process, such as abnormally large volume changes in high-density soil areas.

[0065] Specifically, the volume gradient The slope angle θ can be estimated approximately by the difference in volume change rate between adjacent voxels in the differential model solution. i It can be obtained by calculating the spatial derivative of the terrain elevation map or digital terrain model (DTM). The density parameter ρ i It is obtained through drilling, geological exploration or historical engineering data.

[0066] In a possible implementation, the soil density ρ of the node is further combined i and slope angle θ i The output volume prediction value V of the model i Perform fine-grained optimization. The optimization strategy adjusts the boundary conditions of the predicted value under physical constraints to make the volume change fluctuate within the actual possible soil and rock distribution range, thereby enhancing the engineering usability of the model. The optimization formula is expressed as: Where: V i is the predicted volume value after constrained optimization; η is the learning rate parameter, which is used to control the step size of the gradient update; is the gradient of the loss term with respect to the volume prediction.

[0067] As an option, some areas can be set to dynamically adjust the weight strategy. When the local slope changes drastically or the density jumps, the penalty weight of the physical constraint term is increased, thereby improving the model's adaptability to extreme terrain. This strategy introduces the node perception factor c i Implementation, that is: Where: c i It represents the penalty intensity for the i-th node, which can be a linear function of the slope angle or a nonlinear combination, such as c i =exp(θ i )or

[0068] In this embodiment, after completing the aforementioned physical constraint optimization, the volume prediction results are further visualized to generate a three-dimensional earthwork distribution map for the wind farm road construction area. This distribution map is constructed using a voxel grid to represent the earthwork volume at each spatial location, facilitating regional construction planning by construction units. The generation method is as follows: in: is the predicted volume value at the three-dimensional coordinate (x, y, z); H k V represents the features extracted by the k-th layer graph convolutional network; k is the volume prediction result of the same layer; w kis the fusion weight of the k-th layer output in the final prediction; MLP([·]) means concatenating the graph features and the predicted volume and inputting them into the multi-layer perceptron network for feature transformation and fusion processing; K is the total number of layers in the graph convolutional network.

[0069] Specifically, during the voxel map generation phase, different levels of map features and volume estimation provide information representation capabilities at different scales. Fusion of outputs from different layers helps balance detail capture with global consistency. A multilayer perceptron, which can include two or three nonlinear network layers, outputs the earthwork volume of each voxel, which is then used to construct an overall distribution map.

[0070] In some embodiments, the distribution map can be used in subsequent road design simulation modules. By comparing the volume distribution of different construction paths, it can quickly pre-select road planning solutions and evaluate the workload. Furthermore, to accommodate applications at different resolutions, the spatial resolution of the voxel distribution map can be adjusted dynamically based on user settings or automatically according to terrain complexity.

[0071] The earthwork calculation device for wind farm road construction described below and the earthwork calculation method for wind farm road construction described above may refer to each other.

[0072] Please see the attached Figure 2 , a wind farm road construction earthwork calculation device, including the following modules: Acquisition module, used to collect terrain point cloud data, geological data and meteorological data; Graph modeling module, used to construct topographic map structure and perform graph convolution feature extraction; Dynamic modeling module, used to build a dynamic differential model of earthwork volume changes and solve neural differential equations; Computation scheduling module, used to execute mixed-precision heterogeneous computing strategies; Optimization module, used to introduce geological constraints to optimize prediction results; Output module, used to generate and output the earthwork distribution map of the wind farm road construction The device of this embodiment can be used to execute the above method embodiment, and its principles and technical effects are similar, so they will not be repeated here.

[0073] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, is used to implement the above-mentioned earthwork calculation method for wind farm road construction.

[0074] The computer-readable storage medium may be in the form of, but not limited to, any one or more of the following combinations: Read-only memory (ROM); random access memory (RAM); hard disk, solid-state drive (SSD) or mobile hard disk; flash storage device (such as USB flash drive, SD card, eMMC); optical disk (CD-ROM, DVD); cloud virtual storage resources (such as object storage, distributed file system).

[0075] The examples of this specific embodiment are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, any equivalent changes made based on the structure, shape, and principle of this application should be included in the scope of protection of this application.

Claims

1. A method for calculating earthwork for wind farm road construction, characterized in that: The following steps are involved: S1, collects multi-source data of the wind farm construction area through sensing equipment, including terrain point cloud data, engineering geological data and meteorological time series data; S2, constructing a topographic map structure based on the multi-source data, generating a topographic adjacency matrix and performing spatiotemporal graph convolution feature extraction; S3, based on the spatiotemporal graph convolution features, constructing a dynamic differential model of earthwork volume change and solving it through neural differential equations; S4, using a mixed precision heterogeneous computing strategy to efficiently solve the dynamic differential model and obtain the predicted result of the earthwork volume; S5, optimizing the prediction results in combination with geomechanical constraints, and outputting a distribution map of earthwork in the wind farm road construction area.

2. A method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S1 comprises: The multi-source data includes terrain point cloud data collected by lidar equipment, geological data acquired by exploration equipment, and meteorological data collected by meteorological sensors; Standardizing the collected data, including normalizing the terrain point cloud data, performing structured coding on the geological data, and performing time series processing on the meteorological data; The processed data is encoded into structured tensors as input to subsequent models.

3. A method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S2 comprises: Constructing a graph structure based on the terrain point cloud data, wherein the adjacency matrix of the graph structure is defined by calculating spatial distance and slope gradient; Extracting spatiotemporal features through multi-order spectral domain graph convolution. The graph convolution operation extracts high-dimensional features of nodes for subsequent modeling; The adjacency matrix A ij The following relations are satisfied: Where: p i ,p j is the three-dimensional coordinate of the i-th and j-th nodes, σ is the parameter of the Gaussian kernel function, is the slope gradient between nodes.

4. A method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S3 comprises: Based on the graph convolution features, a differential model of earthwork volume change is constructed. The change of earthwork volume is affected by both slope gradient and construction equipment efficiency. The model is solved by neural differential equations to obtain the time-evolving earthwork volume change value; the volume change rate satisfies the following relationship: Where: V is the volume of earth and stone, is the slope gradient, E(t) is the amount of change in equipment effectiveness over time, and α(t) and β(t) are time functions generated based on graph convolution features and equipment data.

5. A method for calculating earthwork for wind farm road construction according to claim 4, characterized in that: The time functions α(t) and β(t) are calculated by the following formulas: α(t)=softplus(W α ·H t ),β(t)=sigmoid(W β ·[H t ||E(t)]); Where: W α ,W β is the trainable weight matrix, H t is the expression of graph convolution features at time t, and [·||·] represents the feature concatenation operation.

6. A method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S4 comprises: Dynamically determine the current calculation accuracy based on the sparsity of graph convolution features; When the feature sparsity is greater than 0.7, 16-bit floating point numbers are used for calculation; When feature sparsity is between 0.3 and 0.7, 32-bit floating point numbers are used for calculation; When the feature sparsity is less than 0.3, 64-bit floating point numbers are used for calculation; The tasks are divided by the GPU and CPU of the computing platform to perform graph convolution operations and dynamic equation solving tasks respectively.

7. A method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S5 comprises: A geological constraint term is introduced into the model loss function to physically regularize the earthwork volume gradient. Optimize volume prediction results by combining node soil density and terrain slope angle information; Output the earthwork distribution map of the wind farm road construction area; The formula for the geological constraint term is as follows: in: is the loss value of the geological constraint term; λ is the regularization term coefficient, ρ i is the soil density of the ith node, g is the acceleration of gravity, θ i is the slope angle, is the volume gradient of the node; N is the total number of graph nodes involved in the calculation.

8. The method for calculating earthwork for wind farm road construction according to claim 1, characterized in that: The step S5 further comprises: Generate a 3D earthwork volume distribution map, where the earthwork volume of each voxel is processed by combining convolutional features and volume estimation values ​​into a multi-layer perceptron. The voxel volume distribution map is generated by the following formula: in: Represents the volume of earth and stone at the three-dimensional coordinate position (x, y, z), H k is the k-th layer graph convolution feature, V k is the volume prediction value, w k is the fusion weight of the convolution result; MLP([·]) means concatenating the graph features and the predicted volume and inputting them into the multi-layer perceptron network for feature transformation and fusion processing; K is the total number of layers of the graph convolution network.

9. A device for calculating earthwork for road construction in a wind farm, according to a method for calculating earthwork for road construction in a wind farm according to any one of claims 1 to 8, characterized in that: Includes the following modules: Acquisition module, used to collect terrain point cloud data, geological data and meteorological data; Graph modeling module, used to construct topographic map structure and perform graph convolution feature extraction; Dynamic modeling module, used to build a dynamic differential model of earthwork volume changes and solve neural differential equations; Computation scheduling module, used to execute mixed-precision heterogeneous computing strategies; Optimization module, used to introduce geological constraints to optimize prediction results; The output module is used to generate and output the earthwork distribution map of the wind farm road construction.

10. A storage medium, characterized in that: A computer program is stored, and when the program is executed by a processor, it is used to implement the earthwork calculation method for wind farm road construction as described in any one of claims 1 to 8.

Images