Power supply line tower base displacement prediction and early warning method combining physical model and data-driven model

Abstract

The application discloses a power supply line tower foundation displacement prediction and early warning method combining a physical model and a data-driven model, aiming to solve the defects of time-consuming calculation of a pure physical model, difficulty in obtaining soil parameters, weak generalization of a pure data-driven model and easy violation of geotechnical mechanics rules. The method first generates a soil random field sample based on tower foundation geotechnical parameters, constructs an offline simulation feature library through finite element simulation and proxy model expansion and completes dimension reduction; online monitoring data are collected and parameter mapping is realized, similar working conditions are retrieved and fused to obtain comprehensive physical characteristics and safety factors; the integrated learning regression model is inputted with the fused characteristics to predict the tower foundation displacement; finally, the safety factor and the displacement prediction value are taken as double criteria to perform graded early warning. The application combines physical mechanism and data-driven, has strong engineering implementability, high prediction robustness, self-adaptive updating early warning and meets the online real-time early warning demand, effectively guaranteeing the safety of the power supply line tower foundation.

Description

Technical Field

[0001] This invention relates to the fields of power transmission line safety monitoring and artificial intelligence technology, specifically to a method for predicting and warning of power transmission line tower foundation displacement using a combination of physical model and data-driven model. Background Technology

[0002] The stability of power transmission line tower foundations is directly related to power grid safety. Tower foundation displacement is influenced by geological conditions, soil mechanical properties, meteorological and hydrological factors, and external loads, exhibiting complex nonlinear characteristics. Existing assessment methods have two typical limitations:

[0003] Pure physics modeling methods: Finite element-based numerical simulations have clear physical mechanisms and can calculate stress fields and safety factors under different working conditions. However, they are computationally time-consuming and cannot respond in real time to dynamically changing field environments; moreover, accurate soil constitutive parameters are difficult to obtain point-by-point on the engineering site.

[0004] Pure data-driven methods utilize historical data from on-site sensors to train machine learning models for trend prediction, resulting in high computational efficiency. However, in the early stages of tower foundation construction, when samples are scarce or rare extreme working conditions are encountered, the generalization ability of pure data models decreases significantly, and the prediction results may violate the basic laws of geotechnical mechanics.

[0005] Existing technologies lack a fusion mechanism that can efficiently inject high-fidelity simulation results of physical models as prior statistical features into data-driven models, while still ensuring engineering applicability under conditions of parameter uncertainty and imbalanced samples. Summary of the Invention

[0006] This invention aims to provide an engineering-feasible method for predicting and warning of tower foundation displacement using a combined physical model and a data-driven model. This method involves offline construction of a multi-condition simulation feature library and establishing a mapping model from field-monitorable parameters to simulation input parameters. The stress field characteristics of the physical simulation are compressed into low-dimensional statistical features, which are then fused with real-time monitoring data and input into an integrated learning model. Finally, a graded early warning system is achieved by combining the physical safety factor with data-predicted displacement.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for predicting and warning of power line tower foundation displacement using a combined physical model and a data-driven model, the method comprising the following steps:

[0009] Step S1: Generate random field samples of soil parameters based on the geological parameters of the tower base, design and sample multi-dimensional influence conditions, construct an initial offline simulation feature library through finite element fine simulation and surrogate model expansion, reduce the dimensionality of high-dimensional stress field vectors in the library by random projection and solidify the projection matrix, and obtain the final offline simulation feature library containing working condition parameters, low-dimensional stress field characteristics and stability safety factor.

[0010] Step S2: Collect on-site monitoring data online and complete the mapping of monitoring parameters to simulation parameters. Using the final offline simulation feature library as the basis for retrieval and fusion, obtain the comprehensive physical feature vector and comprehensive safety coefficient through similar working condition retrieval and inverse distance weighted fusion. Then, concatenate the monitoring data, comprehensive physical feature vector, and comprehensive safety coefficient into a fusion input feature vector.

[0011] Step S3: Input the fused input feature vector into the ensemble learning regression model after feature preprocessing, and output the predicted value of the future displacement of the tower base;

[0012] Step S4: Based on the comprehensive safety factor and the predicted future displacement of the tower base, a graded early warning is executed.

[0013] Preferably, in step S1, the Karhunen-Löw expansion is used to construct a three-dimensional random field for key soil parameters, and the working condition dimension includes rainfall intensity, wind speed, ice thickness, and groundwater depth.

[0014] Preferably, in step S1, after principal component analysis to reduce the dimensionality of the high-dimensional stress field vector obtained from the fine simulation, Gaussian process regression is used to train the surrogate model to complete the expansion of the simulation feature library.

[0015] Preferably, in step S1, the random projection matrix for dimensionality reduction by random projection is generated using a fixed pseudo-random seed, and the matrix elements follow a specified normal distribution and are permanently fixed after generation.

[0016] Preferably, in step S2, the mapping relationship between monitoring parameters and simulation parameters is calibrated through in-situ pumping tests.

[0017] Preferably, in step S2, during the retrieval and fusion process based on the final offline simulation feature library and through similar working condition retrieval, the formula for calculating the normalized Euclidean distance between the currently estimated working condition and the working condition in the simulation feature library is as follows:

[0018] ;

[0019] In the formula, For the first Normalized Euclidean distance between each simulated working condition and the current estimated working condition; This is the dimension index of the working condition vector; For the current estimated operating condition, the first Dimensional parameters; For the simulation feature library, the first The first working condition Dimensional parameters; This is the normalized scale for each dimension.

[0020] Preferably, in step S2, the inverse distance weighting method is used to fuse the K-nearest neighbor similar working condition features to obtain the comprehensive physical feature vector and the comprehensive safety factor.

[0021] Preferably, the ensemble learning regression model in step S3 is an extreme gradient boosting model, and the feature preprocessing includes standardization and principal component analysis operations.

[0022] Preferably, the graded early warning in step S4 is divided into attention level, warning level, and emergency level, which are determined by the displacement prediction value, displacement rate and safety factor threshold.

[0023] Preferably, the warning threshold used in step S4 for graded early warning adopts the sliding window statistical method and is periodically and adaptively updated based on historical monitoring data.

[0024] The present invention provides a method for predicting and warning of power line tower foundation displacement based on a combined physical model and a data-driven model, which achieves several technical advantages:

[0025] The project boasts strong feasibility: a hybrid library construction strategy combining refined simulation and surrogate model expansion, along with principal component analysis for dimensionality reduction modeling of high-dimensional stress field outputs, keeps offline computation within the project's tolerance. Furthermore, an in-situ test calibration mapping model has been established, bridging the gap between simulation and monitoring spaces, from field-monitorable physical quantities to simulation input parameters.

[0026] The fusion of physics and data is scientifically sound: the stress field obtained from physical simulation is compressed into low-dimensional statistical features through random projection, and then smoothly injected into the data-driven model using a K-nearest neighbor weighted approach. Random projection preserves the relative differences in stress fields under different working conditions while approximately maintaining the Euclidean distance, providing the machine learning model with discriminative statistical features. This fusion mechanism enables the model to exhibit significantly better predictive robustness than pure data models, especially in the early stages of construction when samples are scarce and under rare extreme working conditions.

[0027] The early warning logic is rigorous and adaptive: It adopts a tabular, dual-criteria hierarchical early warning mechanism, clearly distinguishing between two risk scenarios: abnormal deformation trends and deterioration of physical condition. This effectively avoids false alarms during the non-destructive creep period that may result from relying solely on data trends, and significantly reduces prediction uncertainty in rare operating conditions without historical samples. The thresholds are updated online adaptively through sliding window statistics, ensuring the continuous effectiveness of the early warning mechanism throughout the entire lifecycle of the tower base.

[0028] Computational efficiency meets online requirements: all finite element simulations, surrogate model training, stress field extraction, and random projection dimensionality reduction are completed offline. The online phase only requires K-nearest neighbor search, inverse distance weighted fusion, and XGBoost forward inference, meeting the response speed requirements of real-time on-site early warning. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method for predicting and warning the displacement of power line tower foundations based on the combined physical model and data-driven model of the present invention.

[0030] Figure 2 This is a schematic diagram of the K-nearest neighbor retrieval and inverse distance weighting of the present invention.

[0031] Figure 3 This is a curve comparing the measured and predicted values ​​of the tower base displacement according to the present invention.

[0032] Figure 4 This is a comparison chart of the prediction accuracy of different models in this invention. Detailed Implementation

[0033] The following detailed implementation of a power line tower foundation displacement prediction and early warning method based on a joint physical model and a data-driven model according to the present invention will be described in detail with reference to specific embodiments. The embodiments of the present invention are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0034] Example 1: Implementation of a method for predicting and warning the displacement of power line tower foundations using a combination of physical model and data-driven model.

[0035] Combined with appendix Figures 1-2 As shown, this invention provides a method for predicting and warning of power line tower foundation displacement using a combined physical model and a data-driven model, comprising the following steps:

[0036] Step S1: Construct a three-dimensional geotechnical mechanics simulation feature library for the tower base-foundation system offline.

[0037] Step S1 generates random field samples of soil parameters based on the geological parameters of the tower base, designs and samples multi-dimensional influence conditions, and constructs an initial offline simulation feature library through finite element fine simulation and surrogate model expansion. The high-dimensional stress field vectors in the library are randomly projected to reduce the dimensionality and solidify the projection matrix to obtain the final offline simulation feature library containing working condition parameters, low-dimensional stress field characteristics, and stability safety factor.

[0038] In step S1, the Karhunen-Löw expansion is used to construct a three-dimensional random field for key soil parameters. The working condition dimension includes rainfall intensity, wind speed, ice thickness, and groundwater depth.

[0039] In step S1, after principal component analysis to reduce the dimensionality of the high-dimensional stress field vector obtained from the fine simulation, Gaussian process regression is used to train the surrogate model to complete the expansion of the simulation feature library.

[0040] In step S1, the random projection matrix for dimensionality reduction is generated using a fixed pseudo-random seed. The matrix elements follow a specified normal distribution and are permanently fixed after generation. The specific process is as follows:

[0041] Step S1.1: Handling uncertainties in soil parameters.

[0042] Obtain the geological survey report for the tower base area and extract the prior distribution range of the soil's physical and mechanical parameters, including: natural unit weight. Cohesion internal friction angle Elastic modulus Permeability coefficient and Poisson's ratio To account for the spatial variability of the soil, the Karhunen-Loève (KL) expansion (i.e., the Karhunen-Loève expansion) was used to analyze the key parameter cohesion. internal friction angle Perform 3D random field modeling and generate The parameters of the set are used to implement the sample. The correlation length and variogram parameters of the random field are determined based on the soil statistical characteristics in the geological survey data or engineering experience values.

[0043] Step S1.2: Working condition combination design and sampling.

[0044] Identify the main external factors affecting the stability of the tower foundation: rainfall intensity Wind speed Ice thickness and groundwater level depth The Latin hypercube sampling method is used to generate [the data] within the aforementioned four-dimensional parameter space. A set of representative sampling points constitutes the operating condition set. ,in ( ).

[0045] Step S1.3: Feature extraction accelerated by finite element simulation and surrogate model.

[0046] To address the issue of high computational complexity in 3D fluid-structure interaction finite element simulation, this step employs a hybrid strategy of combining a small amount of detailed simulation with surrogate model expansion:

[0047] Detailed simulation samples: From In the full parameter combination, Latin hypercube sampling is used again to select... A representative sample, of which This represents the number of external operating condition combinations generated through Latin hypercube sampling; To determine the number of sample groups based on soil parameters generated using random field theory; To select the representative number of samples from the full parameter combination for performing a complete finite element detailed simulation, a three-dimensional model of the tower base-foundation was built using finite element software, and a complete numerical simulation was performed. The three-dimensional finite element model maintained a completely consistent mesh topology across all detailed simulation samples, ensuring a one-to-one correspondence between the node numbers of each sample in physical space. For each set of detailed simulations, the following two types of information were extracted:

[0048] Base node stress characteristics: Extracting the pre-set stress characteristics at the bottom of the tower foundation. Three-dimensional effective stress tensor of key nodes superscript Node indices are used. The stress tensors of all nodes are expanded and concatenated according to a uniform node index order to form a high-dimensional stress field description vector. .

[0049] in, The number of key nodes pre-defined at the bottom of the tower foundation; Let be the three-dimensional effective stress tensor of the i-th simulation sample at the n-th critical node; For the normal stress component in the x-direction of the effective stress tensor; For the normal stress component in the y-direction of the effective stress tensor; The z-direction normal stress component in the effective stress tensor; The shear stress component in the xy plane of the effective stress tensor; The shear stress component in the yz plane of the effective stress tensor; denoted as , where is the shear stress component in the zx plane of the effective stress tensor; n is the index number of the critical node, ranging from 1 to . ; This is the high-dimensional stress field description vector corresponding to the i-th simulation sample, with a dimension of 6. ; For a dimension of 6N node The real vector space; The overall stability safety factor of the slope or foundation is calculated by the strength reduction method for the i-th simulation sample.

[0050] Overall stability safety factor: The slope / foundation stability safety factor of this sample was calculated using the strength reduction method. .

[0051] Agent model extension: utilizing the above The surrogate model is trained using a set of refined simulation results. To reduce the complexity of high-dimensional output modeling and ensure the training stability of the surrogate model, the high-dimensional stress field vector is first processed. Principal component analysis (PCA) is performed for dimensionality reduction, retaining only the first L principal components as the output target of the surrogate model. L is the number of principal components retained after PCA dimensionality reduction, and L is much less than 6. The surrogate model employs Gaussian process regression. After training, this surrogate model is used to quickly generate the full dataset. The principal component scores of the stress field of the sample group are then used to reconstruct the complete stress field vector through inverse principal component analysis. and the corresponding safety factor .

[0052] Finally, an offline simulation feature library was constructed. .

[0053] The total number of full simulation samples is equal to M and The product; This is a collection of simulation feature libraries built offline, containing operating condition parameters, stress field vectors, and safety factors; Let be the combination of operating parameters corresponding to the i-th simulation sample.

[0054] Step S1.4: Stress field random projection dimensionality reduction.

[0055] To solve the problem of high-dimensional stress field vectors The difficulties of online computation and storage are addressed by performing dimensionality reduction through random projection in the offline stage:

[0056] ;

[0057] in, is a random projection transformation matrix, which is a randomly generated and permanently fixed matrix; The target low-dimensional feature dimension for dimensionality reduction of random projection is typically between 50 and 200. The matrix generation rule is as follows: use a fixed pseudo-random number generator seed (integer 42) to make the matrix elements independently follow a Gaussian distribution with a mean of 0 and a variance of 1 / D. Let be the low-dimensional stress field feature vector obtained by random projection dimensionality reduction of the i-th simulation sample.

[0058] Random projection, while approximately preserving the Euclidean distance, can retain the relative structural differences between stress fields under different working conditions, providing discriminative low-dimensional statistical features for subsequent machine learning models. The basis matrix of random projection is independent of the specific simulation data distribution; when new simulation samples need to be added to the feature library later (while maintaining the mesh topology), the pre-defined matrix can be directly applied to calculate the new features. There is no need to re-perform principal component decomposition of all data. If the geometry is fundamentally changed due to basis expansion or other modifications, the mesh is rebuilt and the corresponding fixed projection matrix is ​​generated. The old and new feature libraries run in parallel through version management. It is the corresponding low-dimensional stress field feature vector calculated when adding a new simulation sample to the feature library; It is the high-dimensional stress field description vector corresponding to the addition of a simulation sample to the feature library.

[0059] Step S2: Online phase multi-source data acquisition, parameter mapping and feature generation.

[0060] In step S2, online data collection of field monitoring is performed and the mapping of monitoring parameters to simulation parameters is completed. Based on the final offline simulation feature library, the comprehensive physical feature vector and comprehensive safety coefficient are obtained through similar working condition retrieval and inverse distance weighted fusion. The monitoring data, comprehensive physical feature vector, and comprehensive safety coefficient are then concatenated into a fusion input feature vector.

[0061] In step S2, the mapping relationship between monitoring parameters and simulation parameters is calibrated through in-situ pumping tests.

[0062] In step S2, the inverse distance weighting method is used to fuse the K-nearest neighbor similar working condition features to obtain the comprehensive physical feature vector and the comprehensive safety factor. The specific process is as follows:

[0063] Step S2.1: On-site monitoring data collection.

[0064] Deploy sensor arrays at the target tower base site with fixed sampling intervals. Real-time acquisition of the following physical quantities: GNSS three-dimensional displacement increment , , Soil volumetric water content pore water pressure Rainfall Wind speed and ambient temperature The original monitoring vector at the current time t is denoted as... .

[0065] Step S2.2: Establish a mapping model of "monitoring parameters → simulation parameters".

[0066] To connect the monitoring data space and the simulation working condition space, the following empirical mapping relationship based on in-situ test calibration at specific working points is established:

[0067] Groundwater level depth In the initial stage of system deployment, in-situ pumping tests were conducted near the tower base to establish the volumetric water content of the soil at the work site. pore water pressure The local calibration curve of the groundwater level depth, i.e., the fitting function During the online phase, this calibration function is used to estimate the depth of the groundwater level in real time.

[0068] Wind speed and icing: The observations from the micrometeorological monitoring device on the tower were directly used as the simulation input. and .

[0069] Rainfall intensity: directly using rain gauge observations. .

[0070] Therefore, the estimated working condition vector at the current moment is defined as .

[0071] Step S2.3: K-nearest neighbor-based similar working condition retrieval and feature weighted fusion.

[0072] like Figure 2 The diagram shown illustrates the K-nearest neighbor retrieval and inverse distance weighting of this invention. The K-nearest neighbor weighting method is used to smooth the discreteness of the working space.

[0073] 1. Calculate the current estimated operating condition With each working condition in the simulation feature library Normalized Euclidean distance:

[0074] ;

[0075] in, The normalized Euclidean distance between the current estimated working condition vector and the i-th historical working condition vector in the simulation feature library; This is the estimated working condition vector at the current moment; This is the i-th historical operating condition vector in the simulation feature library; This represents the j-th dimension component of the current estimated working condition vector; Let j be the j-th component of the i-th historical working condition vector; j is the dimension index of the working condition vector, with a value range of 1 to 4, corresponding to rainfall intensity, wind speed, ice thickness, and groundwater depth, respectively. The normalized scaling parameter corresponding to the j-th feature is taken as the upper limit of the statistical standard deviation of the corresponding variable in each historical working condition or the range of values ​​in the engineering design.

[0076] 2. Select the K working conditions with the smallest distance and denote them as the neighborhood set. .

[0077] K is the number of nearest neighbor historical operating conditions selected, and its value ranges from 3 to 5; Let K be the neighborhood set consisting of the historical conditions of the selected K nearest neighbors.

[0078] 3. Calculate the comprehensive physical characteristics at the current moment using inverse distance weighting:

[0079] ;

[0080] ;

[0081] in, Let be the weight value of the i-th historical condition in the inverse distance weighted fusion; The distance power parameter for inverse distance weighting has a value of 2; This is the low-dimensional stress field feature vector corresponding to the i-th historical working condition in the simulation feature library; This is the current time-integrated physical feature vector obtained after inverse distance weighted fusion; The stability safety factor is the i-th historical operating condition in the simulation feature library. The current moment's comprehensive safety coefficient is obtained after inverse distance weighted fusion.

[0082] Weighted features Weighted safety factor With real-time monitoring data Concatenate the vectors to form the final input vector of the model: . The final input vector of the model is composed of the original monitoring data vector, the comprehensive physical feature vector, and the comprehensive safety factor.

[0083] Step S3: Construct an ensemble learning online prediction model.

[0084] Step S3 inputs the fused input feature vector into the ensemble learning regression model that has undergone feature preprocessing, and outputs the predicted value of the future displacement of the tower base.

[0085] In step S3, the ensemble learning regression model is an extreme gradient boosting model, and feature preprocessing includes standardization and principal component analysis. The specific process is as follows:

[0086] XGBoost (eXtreme Gradient Boosting) is used as the core regression predictor, and a feature preprocessing module is connected in series at the input front end.

[0087] The feature preprocessing module includes the following operations, the specific combination of which is determined through cross-validation:

[0088] 1) Standardization: StandardScaler is used to scale numerical features to eliminate dimensional differences.

[0089] 2) Principal Component Analysis: Decorrelates and reduces the dimensionality of features to eliminate features influenced by physical simulation. To avoid multicollinearity that may be caused by the characteristics of the monitoring conditions, principal components whose cumulative variance contribution rate reaches a preset threshold (99%) are retained.

[0090] The preprocessed feature vector is denoted as... The prediction function of the XGBoost model is represented as an additive tree model:

[0091] ;

[0092] in, The total number of regression trees, For the first A tree of return, The future output of the model The predicted displacement value at the time step.

[0093] During model training, models constructed within historical time periods are used. The sample pairs serve as the training set, and the tree structure and leaf node weights are optimized by minimizing the mean squared error loss function.

[0094] Step S4: Prediction and graded early warning of tower base displacement time sequence.

[0095] Step S4 uses the comprehensive safety factor and the predicted future displacement of the tower base as dual criteria to perform graded early warning.

[0096] In step S4, the graded early warning is divided into attention level, warning level, and emergency level, which is determined by the displacement prediction value, displacement rate and safety factor threshold.

[0097] In step S4, the warning thresholds used for tiered early warning employ a sliding window statistical method, which is periodically and adaptively updated based on historical monitoring data. The specific process is as follows:

[0098] Step S4.1: Displacement time series prediction.

[0099] The input vector constructed at the current time Input into the trained model to obtain the future number. Step (i.e.) Displacement prediction value at time (time) . The number of steps to predict is set according to the required advance warning time.

[0100] Step S4.2: Dual-criteria graded early warning mechanism.

[0101] The early warning trigger condition is determined by the model-predicted displacement and the weighted safety factor based on physical simulation. The decision is made jointly. Thresholds at each level are based on the relevant clauses regarding foundation displacement limits and slope stability safety factors in the "Design Code for 110kV~750kV Overhead Transmission Lines" (GB 50545-2010) and the "Technical Code for Building Slope Engineering" (GB 50330-2013), combined with specific tower design parameters and geological survey reports. The logic for determining the warning level is shown in Table 1. Table 1 is the table for determining the triggering conditions for graded warnings.

[0102] Table 1

[0103]

[0104] The symbols in the table are defined as follows: , where is the displacement rate predicted by the model. This is the measured displacement value at the current moment; This is the upper limit of the normal displacement rate of the tower base, determined based on historical monitoring data. To determine the threshold for the attention level safety factor, take... ; To design allowable displacement values, The ultimate bearing displacement is determined according to the tower foundation design specifications and structural mechanics calculations.

[0105] The aforementioned thresholds are initially set during system deployment. During long-term operation, the displacement rate threshold is adjusted based on a sliding window statistical method. Periodic updates are performed (re-estimation is made quarterly using the most recent 90 days of normal operating condition monitoring data) to ensure that the early warning criteria are synchronized with the actual operating status of the tower base.

[0106] Example 2: Based on the power line tower foundation displacement prediction and early warning method of the combined physical model and data-driven model in Example 1, a simulation example is carried out using the tower foundation displacement prediction and early warning of a 110 kV transmission line angle steel tower as an example.

[0107] Combined with appendix Figures 3-4 As shown, Figure 3 This is a curve comparing the measured and predicted values ​​of the tower base displacement according to the present invention. Figure 4 This is a comparison chart of the prediction accuracy of different models in this invention.

[0108] 1. Project Overview and Offline Database Construction

[0109] The tower foundation is located in the hilly region of South China, with a surface layer of silty clay and an underlying layer of strongly weathered sandstone. The distribution range of soil parameters was determined based on the geological survey report: natural unit weight... Cohesion internal friction angle The correlation length of the random field is 15 m horizontally and 2 m vertically. It is generated using KL expansion. A three-dimensional random field is implemented. External operating conditions include rainfall intensity. Wind speed Ice thickness Groundwater level depth Latin hypercube sampling was used to generate... Each operating condition combination, with a total parameter combination of: Group.

[0110] Latin hypercube sampling was used again from the 6000 groups. A representative set of samples was used to build a three-dimensional finite element model of the tower base-soil structure using ABAQUS, and a complete fluid-structure interaction simulation was performed. The base was extracted. The stress tensors of key nodes are concatenated into a 480-dimensional stress vector. Principal component analysis (PCA) is used to reduce the dimensionality of the stress vector obtained from the detailed simulation, retaining the first L=30 principal components (cumulative variance contribution rate > 95%) as the output target of the surrogate model. The surrogate model is trained using Gaussian process regression, and then the principal component scores of the remaining 5800 samples are quickly generated. Finally, the complete stress field vector is reconstructed using inverse PCA. Set the target dimension of the random projection to D=80, generate a Gaussian random matrix (seed=42) and fix it, then proceed according to the formula. The low-dimensional features of all 6000 samples were calculated. The offline simulation feature library has been completed.

[0111] 2. Online monitoring and feature generation

[0112] Sensor arrays were deployed at the tower base, with sampling intervals... Hours. Before system deployment, a local calibration curve of water content versus water level depth was established through on-site pumping tests: (Buried depth, unit: m). During online operation, the estimated working conditions are calculated at each moment. Normalized scale Take: Rainfall Wind speed Ice thickness 20 mm, water level fluctuation 4 m. According to the formula... Calculate the distance and retrieve the K=4 nearest neighbor working conditions from the feature database. According to the formula and Inverse distance weighted fusion is obtained and According to the formula Combined with real-time monitoring data to form model input .

[0113] 3. Model Training and Prediction

[0114] Monitoring data for the tower base was collected over six months since its commissioning, resulting in approximately 4000 valid samples. The data was then divided into a training set (first four months) and a validation set (last two months) chronologically. Feature preprocessing employed StandardScaler and principal component analysis (preserving key features). Variance). XGBoost hyperparameters are determined through Bayesian optimization. (Learning rate 0.05, maximum depth 5). The model's root mean square error for 24-hour displacement prediction on the validation set is 1.2 mm, compared to the model without physical features. The accuracy of pure data XGBoost (root mean square error 2.5 mm) is significantly improved.

[0115] 4. Early warning threshold setting and operational effectiveness

[0116] According to the design data, the tower foundation is designed to withstand displacement. Limit displacement Upper limit of normal displacement rate Note the safety factor threshold. During the system's online operation, in a period of continuous heavy rainfall, the model predicted a displacement of 35 mm within the next 24 hours, while the weighted safety factor was... The displacement rate dropped to 1.4, triggering a warning level alert. Based on the alert information, maintenance personnel reinforced the tower base drainage facilities in advance, preventing further displacement deterioration. The system uses the most recent 90 days of normal operating condition monitoring data quarterly, employing a sliding window statistical method to analyze the displacement rate intervals. Update.

[0117] like Figure 3 and Figure 4 As shown, under simulated typical operating conditions including rainfall disturbance, the prediction results of the method of this invention show a high degree of consistency with the measured displacement. Compared with the pure data-driven model that does not incorporate physical features, its prediction error is significantly reduced. Statistically, the root mean square error of the method of this invention on a 24-hour prediction scale is approximately 1~1.5 mm, while that of the comparative method is approximately 2~3 mm, verifying the superior prediction accuracy of the proposed method under complex operating conditions.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0119] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting and warning of power line tower foundation displacement using a combined physical model and a data-driven model, characterized in that, The method includes the following steps: Step S1: Generate random field samples of soil parameters based on the geological parameters of the tower base, design and sample multi-dimensional influencing conditions, construct an initial offline simulation feature library through finite element fine simulation and surrogate model expansion, reduce the dimensionality of high-dimensional stress field vectors in the library by random projection and solidify the projection matrix to obtain the final offline simulation feature library containing working condition parameters, low-dimensional stress field characteristics and stability safety factor. Step S2: Collect on-site monitoring data online and complete the mapping of monitoring parameters to simulation parameters. Using the final offline simulation feature library as the basis for retrieval and fusion, obtain the comprehensive physical feature vector and comprehensive safety coefficient through similar working condition retrieval and inverse distance weighted fusion. Then, concatenate the monitoring data, comprehensive physical feature vector, and comprehensive safety coefficient into a fusion input feature vector. Step S3: Input the fused input feature vector into the ensemble learning regression model after feature preprocessing, and output the predicted value of the future displacement of the tower base; Step S4: Based on the comprehensive safety factor and the predicted future displacement of the tower base, a graded early warning is executed.

2. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model as described in claim 1, characterized in that, In step S1, the Karhunen-Löw expansion is used to construct a three-dimensional random field for key soil parameters. The working condition dimension includes rainfall intensity, wind speed, ice thickness, and groundwater depth.

3. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model as described in claim 1, characterized in that, In step S1, after principal component analysis to reduce the dimensionality of the high-dimensional stress field vector obtained from the fine simulation, Gaussian process regression is used to train the surrogate model to complete the expansion of the simulation feature library.

4. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model as described in claim 1, characterized in that, In step S1, the random projection matrix for dimensionality reduction is generated using a fixed pseudo-random seed. The matrix elements follow a specified normal distribution and are permanently fixed after generation.

5. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S2, the mapping relationship between monitoring parameters and simulation parameters is calibrated through in-situ pumping tests.

6. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S2, during the retrieval and fusion process based on the final offline simulation feature library and through similar working condition retrieval, the formula for calculating the normalized Euclidean distance between the current estimated working condition and the working condition in the simulation feature library is as follows: ； In the formula, For the first Normalized Euclidean distance between each simulated working condition and the current estimated working condition; This is the dimension index of the working condition vector; For the current estimated operating condition, the first Dimensional parameters; For the simulation feature library, the first The first working condition Dimensional parameters; This is the normalized scale for each dimension.

7. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S2, the inverse distance weighting method is used to fuse the K-nearest neighbor similar working condition features to obtain the comprehensive physical feature vector and the comprehensive safety factor.

8. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S3, the ensemble learning regression model is an extreme gradient boosting model, and the feature preprocessing includes standardization and principal component analysis.

9. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S4, the graded early warning is divided into three levels: attention level, warning level, and emergency level, which are determined by the displacement prediction value, displacement rate, and safety factor threshold.

10. The method for predicting and early warning of power line tower foundation displacement using a combined physical model and a data-driven model according to claim 1, characterized in that, In step S4, the warning thresholds used for graded early warning adopt the sliding window statistical method and are periodically and adaptively updated based on historical monitoring data.

Images