Extra-high voltage iron tower footing soil thickness prediction method

Abstract

The invention discloses an extra-high voltage iron tower footing soil thickness prediction method, and relates to the technical field of extra-high voltage power grid engineering. The method comprises the following steps: two-dimensional characteristic quantitative extraction; environment index quantification sub-step: acquiring climate parameters including annual average precipitation, rainfall seasonal fluctuation coefficient and annual average wind speed; according to the method, the limitation of single data is broken through by constructing the two-dimensional features of the environmental indexes and the geological features, the cooperative influence of the environment and the geology is quantified, the multi-scale model is combined with the attention weight to adapt to the complex terrain, the soil thickness prediction precision is improved, redundant exploration work is reduced through differentiated drilling point distribution and dynamic iteration adjustment, and the prediction efficiency is improved. Meanwhile, the model has the regional self-adaptive capability, stores optimal parameters of different regions, shortens the subsequent prediction period, provides technical support for rapid and safe site selection of the tower footing of the extra-high voltage iron tower, and guarantees the engineering safety and economy from the source.

Description

Technical Field

[0001] This invention belongs to the field of ultra-high voltage power grid engineering technology, and in particular relates to a method for predicting the soil thickness of ultra-high voltage tower foundations. Background Technology

[0002] The safety of ultra-high voltage (UHV) transmission tower foundation site selection relies on accurate soil thickness data. Soil thickness is a key basis for calculating engineering indicators such as tower foundation bearing capacity and pull-out resistance. Therefore, soil thickness prediction during the tower foundation site selection stage is a core technical requirement, while simultaneously considering both exploration efficiency and cost control to adapt to the characteristics of UHV line construction. Current traditional soil thickness prediction methods in the industry rely solely on single-type data from geology or the environment for data acquisition, employ single-scale algorithms in their models, and rely on uniform and dense drilling for result correction during the validation phase, without forming a systematic collaborative optimization design.

[0003] Traditional methods fail to adequately integrate the synergistic effects of the environment and geology, relying solely on single-dimensional data to comprehensively reflect the formation patterns of soil thickness. Furthermore, their models exhibit weak generalization capabilities, failing to adapt to the spatiotemporal heterogeneity of soil thickness in complex terrains such as steep mountain slopes and river terraces, leading to a sharp drop in accuracy. On the other hand, exploration costs remain high. Traditional intensive drilling methods do not differentiate between areas of varying prediction accuracy, redundantly drilling areas that already meet accuracy requirements, and lack dynamic adjustment mechanisms. When measured errors exceed thresholds, re-exploration is necessary, further increasing the cost burden.

[0004] To address these issues, we provide a method for predicting the soil thickness of ultra-high voltage (UHV) tower foundations. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the soil thickness of ultra-high voltage transmission tower foundations, which solves the problems of existing prediction methods that rely only on single-dimensional data, do not consider the synergistic effects of environment and geology, have weak model generalization ability, and experience a sharp drop in accuracy in complex terrain, ultimately leading to high exploration costs.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution.

[0007] This invention relates to a method for predicting the soil thickness of ultra-high voltage transmission tower foundations, comprising the following steps:

[0008] Two-dimensional feature quantization extraction:

[0009] Environmental index quantification sub-steps: collect climate parameters, including average annual precipitation, seasonal fluctuation coefficient of precipitation, and average annual wind speed, and calculate the climate dynamic index CDI;

[0010] Geological characteristic quantification sub-step: Collect geological parameters of the site selection area, including rock type and rock weathering degree, and calculate the lithological weathering index LWI;

[0011] The feature fusion construction sub-step involves associating the climate dynamics index (CDI) obtained from the environmental index quantification sub-step with the lithological weathering index (LWI) obtained from the geological feature quantification sub-step, with a weight of 1:1, to form a fusion feature vector of environmental index and geological feature.

[0012] Multi-scale model weighted fusion:

[0013] Fine-scale prediction sub-step: The random forest algorithm is used to capture the soil thickness characteristics of the local area by using a single grid cell within the selected area as the prediction unit.

[0014] Mesoscale correction sub-step: The spatiotemporal gated cyclic cell algorithm is adopted, with 10km×10km as the mesoscale cell. The fine-scale prediction results of all grid cells in the cell are input into the algorithm to correct the spatiotemporal heterogeneity in the region.

[0015] Coarse-scale optimization sub-step: Using a graph neural network algorithm, the entire site selection study area is taken as the coarse-scale range, each meso-scale unit is taken as a node of the graph neural network, the geological correlation features between nodes are input into the algorithm, and the cross-regional soil thickness optimization results are output to improve the overall prediction consistency.

[0016] The present invention is further configured such that the multi-scale model weighted fusion includes an attention-weighted fusion sub-step: three levels of model weights are assigned according to the terrain complexity of the selected area. For complex terrains such as steep mountain slopes and river terraces, the weight of the fine-scale model is 0.6, the weight of the mesoscale model is 0.3, and the weight of the coarse-scale model is 0.1. For gentle plain terrains, the weight of the fine-scale model is 0.3, the weight of the mesoscale model is 0.4, and the weight of the coarse-scale model is 0.3. Based on the above weights, the output results from the fine-scale prediction sub-step to the coarse-scale optimization sub-step are weighted and calculated to obtain the final predicted soil thickness value H. pred .

[0017] The present invention further specifies that the method includes an intelligent verification iterative optimization step:

[0018] Differentiated Drill Point Layout Sub-step: Based on the soil thickness prediction value H obtained from the attention-weighted fusion sub-step pred Calculate the prediction error probability for each region. For regions with a prediction error probability greater than 30%, calculate the error every 5 km. 2The density of one drilling point is set as a mandatory drilling point. For areas with a prediction error probability of less than or equal to 30%, a cluster sampling method is used to group grid cells with similar fusion feature vectors of environmental index and geological characteristics into one class, and one representative grid cell is randomly selected from each class to set up a drilling point.

[0019] Error calculation and judgment sub-step: Obtain the measured soil thickness H through drilling. act Formula E abs =|H pred -H act Calculate the absolute error, if the absolute error E abs If the error is greater than 0.2m, proceed to the model iterative optimization sub-step for model optimization. If the absolute error E abs If the value is less than or equal to 0.2m, then the current model parameters are retained;

[0020] Model Iterative Optimization Sub-step: The newly added drilling data from the drilling point differentiation sub-step are allocated to the training and validation sets in a 7:3 ratio. The three-level models of Random Forest, ST-GRU, and GNN from the multi-scale model weighted fusion step are retrained, and the absolute error E is automatically identified. abs For errors greater than 0.2m, such as environmental index deviations and omissions of geological features, adjust the weights of CDI and LWI in the fused feature vector of environmental index and geological features.

[0021] The closed-loop formation sub-step involves repeating the two-dimensional feature quantization extraction step to the intelligent verification iterative optimization step. Based on the iteratively optimized model, the soil thickness is re-predicted, drilling points are deployed, and the error is verified until the absolute error E of all areas within the selected site area is reached. abs All are less than or equal to 0.2m, enabling the prediction of soil thickness at the base of ultra-high voltage transmission towers.

[0022] The present invention is further configured such that the environmental index in the environmental index quantification sub-step also includes the hydrological connectivity index (HCI), which is calculated using the formula HCI = 0.4 × D. r +0.3×K r +0.3×(1 / L r The calculation is performed, where the river network density is denoted as D. r The groundwater recharge coefficient, denoted as K, is calculated based on the digital elevation model (DEM). r The length of the surface runoff path is denoted as L, determined by the lithology of the aquifer. r The Hydrological Connectivity Index (HCI) ranges from 0 to 4.

[0023] The present invention further specifies that the environmental index in the environmental index quantification sub-step includes an Ecological Soil Conservation Index (ESI), which is calculated using the formula ESI = 0.5 × FVC + 0.3 × (R...).d / 100)+0.2×(L t / 10) Calculation, where vegetation cover is denoted as FVC, with a value range of 0-1, obtained through remote sensing inversion, and root depth is denoted as R. d The thickness of the litter, obtained through remote sensing inversion, is denoted as L. t The average value of the on-site samples is taken, and the Ecological Soil Conservation Index (ESI) ranges from 0 to 1.

[0024] The present invention further specifies that the geological features in the geological feature quantification sub-step also include the tectonic influence index TII, which is expressed by the formula TII = 0.6 × F. d +0.4×F g The calculation is performed, where the fault density is denoted as F. d The calculation method is the ratio of fault length to the area of ​​the study area, and the degree of fold development is denoted as F. g The calculation method is the ratio of the number of folds to the area of ​​the study area, and the structural influence index (TII) ranges from 0 to 3.

[0025] The present invention further specifies that the geological features in the geological feature quantification sub-step also include the Topographic Micromorphological Index (TMI), which is calculated using the formula TMI = 0.2 × slope normalized value + 0.2 × aspect normalized value + 0.3 × slope coefficient + ...

[0026] 0.3×(R s / 10) Calculation, where the slope standardization value ranges from 0 to 1, with 0° corresponding to 0 and 90° corresponding to 1; the aspect standardization value ranges from 0 to 1, with sunny slopes corresponding to 1 and shady slopes corresponding to 0; the slope coefficient is assigned according to the slope shape, with convex slopes assigned 0.3, level slopes assigned 0.6, and concave slopes assigned 0.9; and surface roughness is denoted as R. s , where is the elevation standard deviation of the grid cell in the digital elevation model, and the terrain micromorphology index (TMI) ranges from 0 to 1.

[0027] The present invention is further configured such that the training samples of the random forest algorithm in the fine-scale prediction sub-step are derived from the fusion feature vector of environmental index and geological features of the surveyed area of ​​the UHV tower foundation and the measured soil thickness data of the corresponding area; the time step of the ST-GRU algorithm in the mesoscale correction sub-step is determined according to the climate seasonal change cycle of the site selection area, wherein the ST-GRU time step in the arid Northwest region is set to 12 and the ST-GRU time step in the humid South China region is set to 8; and the neighborhood range of the GNN algorithm in the coarse-scale optimization sub-step is determined according to the geological structure distribution characteristics of the site selection area, wherein the GNN neighborhood range in the arid Northwest region is set to 5km and the GNN neighborhood range in the humid South China region is set to 3km.

[0028] The present invention is further configured such that the clustering sampling method of the differential layout sub-step of drilling points has a similarity threshold of 80%, that is, grid cells whose sum of the absolute values ​​of the differences of each parameter of the fused feature vector is less than or equal to 20% are classified into the same class.

[0029] The present invention is further configured such that, when the model iterative optimization sub-step adjusts the weights of the fused feature vector of environmental index and geological features, if the source of error is CDI bias, the weight of CDI is increased to 1.2 times the original weight; if the source of error is LWI bias, the weight of LWI is increased to 1.2 times the original weight.

[0030] The present invention has the following beneficial effects.

[0031] This invention overcomes the limitations of single data by constructing a dual-dimensional feature system of environmental index and geological characteristics, quantifying the synergistic impact of environment and geology. The multi-scale model, combined with attention weights, adapts to complex terrain, improving the accuracy of soil thickness prediction. Through differentiated drilling site layout and dynamic iterative adjustment, redundant exploration work is reduced, significantly lowering exploration costs. At the same time, the model has regional adaptive capabilities, storing optimal parameters for different regions, shortening the subsequent prediction cycle, and providing technical support for the rapid and safe site selection of UHV tower foundations, ensuring the safety and economy of the project from the source. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0033] Figure 1 This is the main flowchart of a method for predicting the soil thickness of an ultra-high voltage transmission tower foundation.

[0034] Figure 2 This is a flowchart of the two-dimensional feature quantization extraction process in a method for predicting the soil thickness of ultra-high voltage tower foundations.

[0035] Figure 3 This is a flowchart of the multi-scale model weighted fusion process in a method for predicting the soil thickness of ultra-high voltage tower foundations.

[0036] Figure 4 This is a flowchart of intelligent verification and iterative optimization in a method for predicting the soil thickness of ultra-high voltage transmission tower foundations.

[0037] Figure 5 This is a flowchart of the model iterative optimization process in a method for predicting the soil thickness of an ultra-high voltage transmission tower foundation. Detailed Implementation

[0038] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0039] Please see Figure 1-5 This invention relates to a method for predicting the soil thickness of ultra-high voltage transmission tower foundations, comprising the following steps:

[0040] Step 1: Two-dimensional feature quantization extraction:

[0041] Environmental index quantification sub-step: For the site selection area of ​​UHV tower foundations, collect climate-related parameters, including average annual precipitation, seasonal fluctuation coefficient of precipitation, and average annual wind speed, where average annual precipitation is denoted as P. a The unit is mm. The seasonal fluctuation coefficient of precipitation is denoted as C, which is calculated as the ratio of the standard deviation of monthly average precipitation to the annual average precipitation. The annual average wind speed is denoted as V. a The unit is m / s, and the formula is CDI = 0.5 × (P) a / 1000)+0.3×C+0.2×(V a / 5) Calculate the Climate Dynamics Index (CDI) to quantify the dynamic impact of climate on soil erosion or deposition. The CDI value ranges from 0 to 5.

[0042] Geological characteristic quantification sub-step: Collect relevant geological parameters of the site selection area, including rock type and rock weathering degree. Assign a basic weathering coefficient according to rock type: 0.8 for igneous rocks, 0.6 for sedimentary rocks, and 0.7 for metamorphic rocks. Assign a weathering degree coefficient according to rock weathering grade: 1.0 for full weathering grade, 0.8 for strong weathering, 0.5 for moderate weathering, and 0.2 for slight weathering. Calculate the lithological weathering index LWI using the formula LWI = basic weathering coefficient × weathering degree coefficient to quantify the potential of rock weathering into soil. The LWI value range is 0-1.

[0043] The feature fusion construction sub-step involves associating the climate dynamic index (CDI) obtained from the environmental index quantification sub-step with the lithological weathering index (LWI) obtained from the geological feature quantification sub-step, with a weight of 1:1, to form a fusion feature vector of environmental index and geological feature. This fusion feature vector serves as the core input data for subsequent model calculations.

[0044] Step 2: Weighted fusion of multi-scale models:

[0045] Fine-scale prediction sub-step: The random forest algorithm is used to take a single grid cell in the selected area as the prediction unit. The environmental index and geological characteristics of the grid cell are fused into a feature vector and input into the algorithm. The single-point soil thickness prediction result of the grid cell is output to capture the soil thickness characteristics of local areas, such as slope differences and local lithological changes.

[0046] Mesoscale correction sub-step: The spatiotemporal gated cyclic cell algorithm is adopted, with 10km×10km as the mesoscale cell. The fine-scale prediction results of all grid cells in the cell are input into the algorithm, and the spatiotemporal correlation characteristics between cells are incorporated, such as the CDI gradient law and LWI continuity of adjacent grid cells. The soil thickness correction results of the mesoscale cell are output to correct the spatiotemporal heterogeneity in the region, such as the soil thickness gradient trend of river valley terraces.

[0047] Coarse-scale optimization sub-step: The graph neural network algorithm is adopted, taking the entire site study area as the coarse-scale range and each mesoscale unit as a node of the graph neural network. The geological correlation characteristics between nodes, such as the influence of fault zones on the LWI of adjacent mesoscale units, are input into the algorithm, and the output is the cross-regional soil thickness optimization result, which is used to improve the overall prediction consistency.

[0048] Attention-weighted fusion sub-step: Three-level model weights are assigned based on the terrain complexity of the selected area. For complex terrains such as steep mountain slopes and river terraces, the weights are 0.6 for the fine-scale model, 0.3 for the mesoscale model, and 0.1 for the coarse-scale model. For gentle plains, the weights are 0.3 for the fine-scale model, 0.4 for the mesoscale model, and 0.3 for the coarse-scale model. Based on these weights, the outputs from the fine-scale prediction sub-step to the coarse-scale optimization sub-step are weighted and calculated to obtain the final predicted soil thickness value H. pred ;

[0049] Step 3: Intelligent verification iterative optimization:

[0050] Differentiated Drill Point Layout Sub-step: Based on the soil thickness prediction value H obtained from the attention-weighted fusion sub-step pred The prediction error probability of each region is calculated. For regions with a prediction error probability greater than 30%, such as tectonic fault zones and abrupt topographic changes, mandatory drilling points are set at a density of one drilling point every 5 km. For regions with a prediction error probability less than or equal to 30%, such as gentle plains, a cluster sampling method is used to group grid units with similar fusion feature vectors of environmental indices and geological characteristics into one category, and one representative grid unit is randomly selected from each category to set up drilling points.

[0051] Error calculation and judgment sub-step: Obtain the measured soil thickness H through drilling. act Formula E abs =|H pred -H act Calculate the absolute error, if the absolute error E abs If the error is greater than 0.2m, proceed to the model iterative optimization sub-step for model optimization. If the absolute error E abs If the value is less than or equal to 0.2m, then the current model parameters are retained;

[0052] Model Iterative Optimization Sub-step: The newly added drilling data from the drilling point differentiation sub-step are allocated to the training and validation sets in a 7:3 ratio. The three-level models of Random Forest, ST-GRU, and GNN from the multi-scale model weighted fusion step are retrained, and the absolute error E is automatically identified. abs For errors greater than 0.2m, such as environmental index deviations and omissions of geological features, adjust the weights of CDI and LWI in the fused feature vector of environmental index and geological features.

[0053] The closed-loop formation sub-step involves repeating the two-dimensional feature quantization extraction step to the intelligent verification iterative optimization step. Based on the iteratively optimized model, the soil thickness is re-predicted, drilling points are deployed, and the error is verified until the absolute error E of all areas within the selected site area is reached. abs All are less than or equal to 0.2m, enabling the prediction of soil thickness at the base of ultra-high voltage transmission towers.

[0054] Example 1

[0055] The scene depicts steep mountain slopes in the cold temperate zone. The slopes in this area are mostly between 25° and 40°. The average annual precipitation is low and concentrated in the summer. Wind erosion is significant. The lithology is mainly igneous rock, with small faults distributed locally.

[0056] The implementation process begins with the extraction of two-dimensional features and the calculation of the Climate Dynamics Index (CDI). The formula given by this method is CDI = 0.5 × (P...). a / 1000)+0.3×C+0.2×(V a / 5), where P a The average annual precipitation for this scenario is 320 mm, C is the seasonal fluctuation coefficient of precipitation (0.65), and V... a Assuming an average annual wind speed of 4.2 m / s, substituting the parameters into the formula yields CDI = 0.5 × (320 / 1000) + 0.3 × 0.65 + 0.2 × (4.2 / 5) = 0.16 + 0.195 + 0.168 = 0.523. This value falls within the 0-5 range defined by the method, reflecting the environmental impact dominated by wind erosion. The lithological weathering index LWI is then calculated using the formula LWI = basic weathering coefficient × weathering degree coefficient. The method specifies a basic weathering coefficient of 0.8 for igneous rocks. Field investigations have determined that the igneous rocks in this area are strongly weathered, and the method specifies a strong weathering degree coefficient of 0.8. Substituting this into the formula, we get LWI = 0.8 × 0.8 = 0.64, which conforms to the 0-1 value range defined by the method. This quantifies the weathering potential of igneous rocks into soil. Finally, the tectonic influence index TII is calculated using the formula TII = 0.6 × F d +0.4×F g F d With a fault density of 0.12 km / km⁵, F gAssuming a fold development degree of 0.06 folds / km5, substituting the values, we get TII = 0.6 × 0.12 + 0.4 × 0.06 = 0.072 + 0.024 = 0.096, which falls within the method-defined range of 0-3, indicating weak tectonic disturbance. The topographic micromorphological index (TMI) is calculated using the formula: TMI = 0.2 × slope-standardized value + 0.2 × aspect-standardized value + 0.3 × slope coefficient + 0.3 × (R) s / 10), the slope normalization value is 0.42, the aspect normalization value is 0.85, the slope coefficient is 0.3, and R s Assuming a surface roughness of 3.2m, substituting the values, we get TMI = 0.2 × 0.42 + 0.2 × 0.85 + 0.3 × 0.3 + 0.3 × (3.2 / 10) = 0.084 + 0.17 + 0.09 + 0.096 = 0.44, which falls within the 0-1 range defined by the method. This indicates that the terrain provides relatively weak support for soil deposition. Subsequently, CDI and LWI are correlated with a 1:1 weight as required by the method to form a fused feature vector of environmental index and geological features, which serves as the input for subsequent models.

[0057] In the multi-scale model weighted fusion stage, the fine-scale model uses the random forest algorithm specified by the method. After inputting the fusion feature vector, it outputs the predicted soil thickness at a single point, such as 0.82m for convex slope areas and 0.43m for fault zones, capturing the thickness differences caused by local topography and structure. The mesoscale model uses the ST-GRU algorithm specified by the method, setting the time step to 12 according to the seasonal variation cycle of the cold temperate zone. It inputs the fine-scale results within a 10km×10km cell, while incorporating the CDI gradient law between cells to correct the thickness deviation of the windward slope caused by wind erosion. The coarse-scale model uses the GNN algorithm required by the method, setting the neighborhood range to 5km according to the regional geological structure distribution. It uses the mesoscale cells as nodes, inputs the LWI continuity features between nodes, and optimizes the consistency of cross-regional predictions. Finally, according to the weight allocation rule of complex terrain in the method, namely, steep slopes are considered complex terrain, with fine-scale 0.6, mesoscale 0.3, and coarse-scale 0.1, the results of the three-level model are weighted and calculated to obtain the final predicted soil thickness value.

[0058] In the intelligent verification and iterative optimization phase, the prediction error probability for each region is first calculated according to the method. For fault zones with an error probability >30%, drilling points are set at a density of one mandatory drilling point every 5 km², as specified in the method. For regions with an error probability ≤30%, the clustering sampling method specified in the method is used to group grid cells into one class based on an 80% similarity of the fused feature vectors. One representative cell is randomly selected from each class to set up a drilling point. The initial drilling volume is reduced by 34% compared to the traditional method, significantly reducing redundant costs. Subsequently, the measured soil thickness H is obtained through drilling. act According to formula E abs =|H pred -H act|Calculate the absolute error for all drilling points E abs All values ​​are ≤0.19m, requiring no additional data points. Finally, the newly added drilling data are distributed to the training and validation sets in the required 7:3 ratio, and the three-level model is retrained. The weight of TMI in the fused feature vector is slightly adjusted. After iteration, the average prediction error is 0.17m, which meets the accuracy requirements of the method. The cost of single-line exploration is reduced by 32%, which is suitable for the wind erosion and complex terrain features of this scenario.

[0059] Example 2

[0060] For the scenario of river valley terraces in humid areas, the terraces in this region are 500m-1km wide, with abundant annual precipitation, significant hydrological effects, and lithology dominated by sedimentary rocks and dense vegetation cover.

[0061] During implementation, the two-dimensional features were first extracted according to the method's workflow: Regarding the environmental index, the CDI was calculated to be 1.005 using the formula in the method, with parameters including annual average precipitation of 1550 mm, C 0.42, and V... a 2.6 m / s, according to the given HCI calculation formula, HCI = 0.4 × D r +0.3×K r +0.3×(1 / L r The calculated HCI is 0.824, with a river network density of 0.85 km / km. 2 K r 0.9, L r 1.4km, according to the ESI formula in the method, ESI=0.5×FVC+0.3×(R) d / 100)+0.2×(L t / 10) The calculated ESI is 0.856, with parameters FVC 0.88 and R d 82cm, L t 8.5cm.

[0062] Regarding geological characteristics, LWI was calculated to be 0.3 according to the formula in the method, with parameters of basic weathering coefficient of sedimentary rock of 0.6 and moderate weathering coefficient of 0.5, and TMI of 0.519 was calculated. These parameters were then fused into a feature vector with a weight of 1:1.

[0063] In the multi-scale fusion stage, the fine-scale random forest predicts the bottom of the terrace at 1.35m and the top at 0.92m, the mesoscale ST-GRU sets the time step to 8 according to the precipitation cycle of the humid region, and the coarse-scale GNN neighborhood range is 3km. The fusion is performed according to the weight of complex terrain.

[0064] During verification, mandatory points were set up in the transition zone where the error exceeded 30%, and the central cluster sampling was used. The measured error was ≤0.18m, and the mean after iteration was 0.15m, reducing the cost by 29%, which meets the method's adaptation requirements for humid scenarios.

[0065] Example 3

[0066] This method is used for temperate, gently sloping plains with a slope of <5°, a mild climate, and predominantly sedimentary rocks. During dual-dimensional feature extraction, the formulas in this method yielded CDI 0.59, ESI 0.57, LWI 0.6, sedimentary rock weathering coefficient 1.0, TII 0.052, and TMI 0.378, which were then fused into a feature vector. In the multi-scale fusion stage, the fine-scale prediction thickness was 1.0-1.2m, the mesoscale ST-GRU time step was 10, and the coarse-scale GNN neighborhood range was 4km. The method uses the gentle terrain weights (0.3 for fine-scale, 0.4 for mesoscale, and 0.3 for coarse-scale) for fusion. During validation, 90% of the regions were clustered for sampling, with only paleochannels used as mandatory points, reducing drilling by 46%. The measured error was ≤0.14m, and the mean after iteration was 0.13m, meeting the method's design goal of low-cost and efficient prediction of gentle terrain.

[0067] Example 4

[0068] Adapted to subtropical peak-cluster depression scenarios, with metamorphic rocks as the main lithology and dissolution characteristics, the dual-dimensional feature extraction method calculates CDI 0.899, HCI 0.504, ESI 0.505, LWI 0.56, basic weathering coefficient of metamorphic rocks 0.7, strong weathering coefficient 0.8, TII 0.164, and TMI 0.545 according to the formula in this method. These are then fused into feature vectors. During multi-scale fusion, the fine-scale prediction of depressions is 0.72m and peak-clusters is 0.32m. The time step of the medium-scale ST-GRU is 9, and the neighborhood range of the coarse-scale GNN is 4.5km. The fusion is weighted according to the complex terrain. During the verification, mandatory points are set at the edge of the depression, and peak-cluster clustering is used for sampling. The drilling volume is reduced by 32%, the measured error is ≤0.19m, and the mean after iteration is 0.18m, which solves the problem of prediction bias for special lithologies.

[0069] Example 5

[0070] Applied to a cold-temperate mixed coniferous and broad-leaved forest scenario, characterized by significant winter freeze-thaw cycles and deep-rooted mixed vegetation, the two-dimensional feature extraction yielded CDI 0.72, ESI 0.816, FVC 0.78, and R... d 72cm, LWI 0.35, weathering coefficient in metamorphic rock 0.5, TMI 0.453, fused into feature vectors, fine-scale correction of pseudo-thickness caused by freeze-thaw during multi-scale fusion, medium-scale ST-GRU time step 9, coarse-scale GNN neighborhood range 4km, fused according to medium-complexity terrain weights i.e. fine 0.4, medium 0.4, coarse 0.2, mandatory points were set up in the strong freeze-thaw area of ​​the slope top during verification, forest area clustering sampling, drilling volume reduced by 39%, measured error ≤0.17m, mean after iteration 0.16m, adapting to the dynamic prediction needs of freeze-thaw scenarios in the adaptation method.

[0071] Example 6

[0072] For the hilly area in the subtropical typhoon-affected zone, which experiences an average of 2-3 typhoons per year and is dominated by sedimentary rocks, the two-dimensional feature extraction method calculates CDI as 1.379 and V using the formula in this method. a 5.2 m / s, HCI 0.813, ESI 0.848, LWI 0.48, sedimentary rock strong weathering coefficient 0.8, TII 0.124, TMI 0.572, were fused into feature vectors. During multi-scale fusion, fine-scale corrections were made for windward slope erosion caused by typhoons. The time step of the mesoscale ST-GRU was 8, and the neighborhood range of the coarse-scale GNN was 3 km. Fusion was performed according to the weight of complex terrain. During the verification, mandatory points were set on the windward slope, and cluster sampling was performed on the leeward slope. The drilling volume was reduced by 33%, the measured error was ≤0.15m, the mean after iteration was 0.14m, and the deviation after the typhoon was still ≤0.2m, which met the requirements of the method for the stability of typhoon scenario prediction.

[0073] The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation described herein. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. A method for predicting the soil thickness of an ultra-high voltage transmission tower foundation, characterized in that... This includes the following steps: Two-dimensional feature quantization extraction: Environmental index quantification sub-steps: collect climate parameters, including average annual precipitation, seasonal fluctuation coefficient of precipitation, and average annual wind speed, and calculate the climate dynamic index CDI; Geological characteristic quantification sub-step: Collect geological parameters of the site selection area, including rock type and rock weathering degree, and calculate the lithological weathering index LWI; The feature fusion construction sub-step involves associating the climate dynamics index (CDI) obtained from the environmental index quantification sub-step with the lithological weathering index (LWI) obtained from the geological feature quantification sub-step, with a weight of 1:1, to form a fusion feature vector of environmental index and geological feature. Multi-scale model weighted fusion: Fine-scale prediction sub-step: The random forest algorithm is used to capture the soil thickness characteristics of the local area by using a single grid cell within the selected area as the prediction unit. Mesoscale correction sub-step: The spatiotemporal gated cyclic unit algorithm is adopted, with 10km×10km as the mesoscale unit. The fine-scale prediction results within the unit are input into the algorithm to correct the spatiotemporal heterogeneity within the region. Coarse-scale optimization sub-step: Using a graph neural network algorithm, the entire site selection study area is taken as the coarse-scale range, each meso-scale unit is taken as a node of the graph neural network, the geological correlation features between nodes are input into the algorithm, and the cross-regional soil thickness optimization results are output to improve the overall prediction consistency.

2. The method for predicting soil thickness at the foundation of an ultra-high voltage transmission tower according to claim 1, characterized in that: The multi-scale model weighted fusion also includes an attention-weighted fusion sub-step: The weights of the three-level model are assigned based on the topographic complexity of the site selection area. For complex terrains such as steep mountain slopes and river terraces, the weight of the fine-scale model is 0.6, the weight of the mesoscale model is 0.3, and the weight of the coarse-scale model is 0.

1. For gentle plain terrains, the weights of the fine-scale model are 0.3, the weight of the mesoscale model is 0.4, and the weight of the coarse-scale model is 0.

3. Based on these weights, the outputs of the fine-scale prediction sub-step to the coarse-scale optimization sub-step are weighted and calculated to obtain the final predicted soil thickness value H. pred .

3. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The method also includes an intelligent verification iterative optimization step: Differentiated Drill Point Layout Sub-step: Based on the soil thickness prediction value H obtained from the attention-weighted fusion sub-step pred Calculate the prediction error probability for each region. For regions with a prediction error probability greater than 30%, set up mandatory drilling points at a density of 1 drilling point per 5km5. For regions with a prediction error probability less than or equal to 30%, use cluster sampling method to group grid units with similar fusion feature vectors of environmental index and geological characteristics into one category, and randomly select one representative grid unit from each category to set up drilling points. Error calculation and judgment sub-step: Obtain the measured soil thickness H through drilling. act Formula E abs =|H pred -H act Calculate the absolute error, if the absolute error E abs If the error is greater than 0.2m, proceed to the model iterative optimization sub-step for model optimization. If the absolute error E abs If the value is less than or equal to 0.2m, then the current model parameters are retained; Model Iterative Optimization Sub-step: The newly added drilling data from the drilling point differentiation sub-step are allocated to the training and validation sets in a 7:3 ratio. The three-level models of Random Forest, ST-GRU, and GNN from the multi-scale model weighted fusion step are retrained, and the absolute error E is automatically identified. abs For errors greater than 0.2m, such as environmental index deviations and omissions of geological features, adjust the weights of CDI and LWI in the fused feature vector of environmental index and geological features. The closed-loop formation sub-step involves repeating the two-dimensional feature quantization extraction step to the intelligent verification iterative optimization step. Based on the iteratively optimized model, the soil thickness is re-predicted, drilling points are deployed, and the error is verified until the absolute error E of all areas within the selected site area is reached. abs All are less than or equal to 0.2m, enabling the prediction of soil thickness at the base of ultra-high voltage transmission towers.

4. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The environmental index quantification sub-step also includes the hydrological connectivity index (HCI), which is calculated using the formula HCI = 0.4 × d. r +0.3×K r +0.3×(1 / L r The calculation is performed, where the river network density is denoted as D. r The groundwater recharge coefficient, denoted as K, is calculated based on the digital elevation model (DEM). r The length of the surface runoff path is denoted as L, determined by the lithology of the aquifer. r The Hydrological Connectivity Index (HCI) ranges from 0 to 4.

5. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The environmental index quantification sub-step also includes the Ecological Soil Conservation Index (ESI), which is calculated using the formula ESI = 0.5 × FVC + 0.3 × (R). d / 100)+0.2×(L t / 10) Calculation, where vegetation cover is denoted as FVC, with a value range of 0-1, obtained through remote sensing inversion, and root depth is denoted as R. d The thickness of the litter, obtained through remote sensing inversion, is denoted as L. t The average value of the on-site samples is taken, and the Ecological Soil Conservation Index (ESI) ranges from 0 to 1.

6. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The geological features quantification sub-step also includes the tectonic influence index (TII), which is calculated using the formula TII = 0.6 × F. d +0.4×F g The calculation is performed, where the fault density is denoted as F. d The calculation method is the ratio of fault length to the area of ​​the study area, and the degree of fold development is denoted as F. g The calculation method is the ratio of the number of folds to the area of ​​the study area, and the structural influence index (TII) ranges from 0 to 3.

7. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The geological features quantification sub-step also includes the Topographic Micromorphological Index (TMI), which is calculated using the formula: TMI = 0.2 × slope-standardized value + 0.2 × aspect-standardized value + 0.3 × slope coefficient + 0.3 × (R) s / 10) Calculation, where the slope standardization value ranges from 0 to 1, with 0° corresponding to 0 and 90° corresponding to 1; the aspect standardization value ranges from 0 to 1, with sunny slopes corresponding to 1 and shady slopes corresponding to 0; the slope coefficient is assigned according to the slope shape, with convex slopes assigned 0.3, level slopes assigned 0.6, and concave slopes assigned 0.9; and surface roughness is denoted as R. s , where is the elevation standard deviation of the grid cell in the digital elevation model, and the terrain micromorphology index (TMI) ranges from 0 to 1.

8. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 1, characterized in that: The random forest algorithm in the fine-scale prediction sub-step uses environmental index and geological feature fusion vectors from the surveyed areas of the UHV tower foundations and the corresponding measured soil thickness data. The ST-GRU algorithm in the mesoscale correction sub-step has a time step determined according to the seasonal climate change cycle of the site selection area, with the ST-GRU time step set to 12 in the arid Northwest region and 8 in the humid South China region. The GNN algorithm in the coarse-scale optimization sub-step has a neighborhood range determined according to the geological structure distribution characteristics of the site selection area, with the GNN neighborhood range set to 5km in the arid Northwest region and 3km in the humid South China region.

9. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 3, characterized in that: The clustering sampling method for the differential layout of drilling points has a similarity threshold of 80%, meaning that grid cells whose sum of the absolute values ​​of the differences between the parameters of the fused feature vectors is less than or equal to 20% are classified into the same class.

10. The method for predicting the soil thickness of an ultra-high voltage transmission tower foundation according to claim 3, characterized in that: When adjusting the weights of the fused feature vector of environmental index and geological features in the model iterative optimization sub-step, if the error source is CDI bias, the weight of CDI is increased to 1.2 times the original weight; if the error source is LWI bias, the weight of LWI is increased to 1.2 times the original weight.

Images