Power transmission and transformation tower foundation settlement trend prediction method and system based on generative adversarial network

Abstract

The application discloses a power transmission and transformation tower foundation settlement trend prediction method and system based on a generative adversarial network, belongs to the technical field of tower foundation settlement trend prediction, and comprises the following steps: firstly, collecting tower foundation settlement and multi-source environmental data, and constructing an environmental noise sequence through probability estimation and a Copula function; according to a historical prediction accuracy parameter, dynamically dividing historical settlement data and a measured sequence, and building a generative adversarial network with a long short-term memory network as a generator and a discriminator capable of discriminating time sequence statistical characteristics; inputting the historical data and random noise into the generator, and making the accuracy rate of the generator reach a standard through alternating adversarial training; finally, outputting a future settlement trend sequence by using the trained generator, combining a cumulative settlement amount and a settlement rate to evaluate the stability of the tower foundation, and realizing graded early warning.

Description

Technical Field

[0001] This invention relates to the field of tower foundation settlement trend prediction technology, and in particular to a method and system for predicting the settlement trend of power transmission and transformation tower foundations based on generative adversarial networks. Background Technology

[0002] Transmission tower foundations are the load-bearing core of transmission lines, and their stability directly affects the safety of the power grid. During operation, tower foundations are subjected to the combined effects of complex environments such as geological creep, hydrological changes, and external loads, making them prone to slow and uneven settlement. If the settlement evolution trend is not monitored in time, it can lead to serious accidents such as tower tilting, component fracture, or even tower collapse, severely threatening reliable power supply and public safety. Therefore, accurately predicting tower foundation settlement trends and proactively anticipating minute deformations and long-term evolution patterns is crucial for shifting the operation and maintenance model from reactive repair to proactive early warning. This has extremely important engineering value and practical significance for scientifically formulating maintenance strategies, preventing structural risks, and ensuring the safe operation of the power energy artery throughout its entire life cycle.

[0003] In the monitoring of settlement of power transmission and transformation tower foundations, the tower foundations are subjected to the coupled effects of multiple complex environmental factors such as geological creep, hydrological changes, and external loads over a long period of time. This results in settlement monitoring data generally suffering from problems such as sparse samples, large noise interference, and significant temporal nonlinear characteristics. Traditional prediction models have limited fitting accuracy for such data, large deviations in long-term settlement trend predictions, weak model generalization ability, and are prone to prediction distortion, making it difficult to meet the accuracy requirements for high-reliability and safe operation and maintenance of power facilities. Therefore, there is an urgent need for a prediction method that can accurately capture the implicit evolution law of tower foundation settlement under complex operating conditions and small sample conditions, and achieve reliable short-term numerical prediction and long-term trend extrapolation. Summary of the Invention

[0004] To address the limitations of existing technologies in predicting the settlement of power transmission tower foundations, such as sparse monitoring data samples, strong multi-source noise interference, and significant nonlinear characteristics of the settlement process, which lead to insufficient fitting accuracy and large deviations in long-term trend predictions using traditional models, this invention provides a method and system for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks. The technical solution is as follows: On the one hand, a method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks is provided. This method includes: S1, collecting settlement monitoring data and surrounding environmental data from historical monitoring periods of the power transmission tower foundations. The surrounding environmental data includes geological creep parameters, hydrological change parameters, and external load parameters. Based on the surrounding environmental data, a multi-source environmental disturbance randomness assessment is performed to obtain an environmental noise sequence, which serves as the random noise of the surrounding environment of the power transmission tower foundations; S2, dynamically dividing the settlement monitoring data based on historical prediction accuracy parameters to obtain measured data of historical settlement data and prediction sequence segments; S3, constructing a generative adversarial network, which includes a generator and a discriminator. The generator is used to fit the temporal variation law of settlement and outputs the future settlement corresponding to the period to be predicted. The trend prediction sequence is used by a discriminator to distinguish the features of real monitoring data and generator prediction data, so as to effectively expand and enhance the features of small sample data; S4, the generative adversarial network is trained in an alternating adversarial manner: historical settlement data and random noise are input into the generator, and the generator outputs a prediction sequence of future settlement trends; the measured data of the prediction sequence segment and the prediction sequence are input into the discriminator, and the generator is optimized through adversarial means to capture the dynamic evolution law of settlement data until the generator generates a settlement curve with an accuracy rate that reaches a preset threshold; S5, the latest collected real monitoring data is used as a new historical settlement data sequence, which is input into the trained generator, and the predicted sequence of the sequence segment to be analyzed is output. Based on the predicted sequence of the sequence segment to be analyzed, the stability of the tower base is assessed and graded early warning is performed.

[0005] On the other hand, a system for predicting the settlement trend of power transmission and transformation tower foundations based on generative adversarial networks (GANs) is provided. This system includes: a data acquisition module for collecting settlement monitoring data and surrounding environmental data from historical monitoring periods of the power transmission and transformation tower foundations; an environmental noise assessment module for performing multi-source environmental interference randomness assessment based on surrounding environmental data to obtain an environmental noise sequence; a data partitioning module for dynamically partitioning settlement monitoring data based on historical prediction accuracy parameters to obtain measured data of historical settlement data and predicted sequence segments; a model building module for constructing a GAN, which includes a generator and a discriminator, with the generator built using a long short-term memory network; an adversarial training module for inputting historical settlement data and random noise into the generator, outputting a predicted sequence, and then inputting the measured data of the predicted sequence segment and the predicted sequence into the discriminator for alternating adversarial training until the accuracy of the generated settlement curve reaches a preset threshold; a trend prediction module for inputting the latest collected real monitoring data into the trained generator, outputting a predicted sequence of the sequence segment to be analyzed; and a stability assessment and early warning module for performing tower foundation stability assessment and graded early warning based on the predicted sequence of the sequence segment to be analyzed.

[0006] On the other hand, a computer-readable storage medium is provided, which stores instructions that, when executed on a computer, cause the computer to execute a method for predicting the settlement trend of power transmission and transformation tower foundations based on generative adversarial networks.

[0007] Beneficial effects The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By integrating multi-source environmental parameters from geology, hydrology, and external loads, and combining kernel density estimation and Copula function, a multi-dimensional environmental noise sequence is accurately constructed. This quantifies the randomness and parameter correlation of environmental disturbances, thereby accurately simulating the complex environmental disturbance conditions of tower foundation settlement. This achieves full coverage and refined modeling of settlement influencing factors, effectively improving the adaptability of the prediction model to complex field environments.

[0008] 2. By dynamically dividing settlement monitoring data based on accuracy parameters such as historical prediction error and goodness of fit, and combining adversarial training mechanism to expand small sample data and enhance temporal features, the model training dataset structure is adaptively optimized and the dynamic evolution law of settlement time series is accurately captured. This solves the problems of poor small sample adaptability and insufficient extraction of temporal features in traditional prediction methods, and significantly improves the accuracy and stability of settlement trend prediction.

[0009] 3. By coupling the maximum cumulative settlement and the average settlement rate, a quantitative stability assessment index is constructed, and a multi-level standardized early warning threshold system is set up to achieve accurate quantitative assessment and hierarchical early warning of the tower foundation settlement status. This leads to the formation of a fully intelligent handling system from data collection and trend prediction to risk assessment and early warning output, ensuring the safe operation of power transmission and transformation tower foundations and the efficiency of power grid operation and maintenance management. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 Flowchart of the method for predicting the settlement trend of power transmission tower foundation based on generative adversarial networks provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the power transmission tower foundation settlement trend prediction system based on generative adversarial networks provided in the embodiments of this application. Detailed Implementation

[0012] The following provides explanations for some of the terms used in this application. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0013] The embodiments of this application involve at least one, including one or more; wherein, multiple means two or more. Furthermore, it should be understood that in the description of this specification, terms such as "first," "second," and "third" are used only for descriptive purposes and should not be construed as indicating relative importance or order. For example, "first device" and "second device" do not represent the degree of importance of the two or their order, but are merely for descriptive distinction. In the embodiments of this application, "and / or" merely describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone.

[0014] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0015] like Figure 1The diagram shows a flowchart of a method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks, provided in this application embodiment. The method includes the following steps: S1, collecting settlement monitoring data and surrounding environmental data from historical monitoring periods of the power transmission tower foundation. The settlement monitoring data includes single settlement displacement, cumulative settlement, and a settlement rate sequence calculated based on time difference at each monitoring point of the tower foundation. The settlement monitoring data is obtained by deploying static levels, displacement sensors, and global navigation satellite system receivers at key locations on the foundation platform, tower leg support points, and surrounding ground surfaces of the power transmission tower foundation. Continuous automated data collection is performed according to a preset sampling frequency, and time-series data for model input is generated after time synchronization and denoising preprocessing. The surrounding environmental data includes geological creep parameters, hydrological change parameters, and external load parameters. A multi-source environmental interference randomness assessment is performed based on the surrounding environmental data, constructing a probability distribution model of environmental noise and generating an environmental noise sequence as random noise in the surrounding environment of the power transmission tower foundation. S2, dynamically dividing the data based on historical prediction accuracy parameters... S3: Based on the settlement monitoring data, historical settlement data and measured data of the predicted sequence segment are obtained. S4: A generative adversarial network (GAN) is constructed, consisting of a generator and a discriminator. The generator, built using a long short-term memory network, fits the temporal variation of settlement and outputs a predicted sequence of future settlement trends corresponding to the period to be predicted. The discriminator distinguishes the features of the actual monitoring data from the generator's predicted data, effectively expanding and strengthening the features of small sample data. S5: The GAN is trained through alternating adversarial training: historical settlement data and random noise are input into the generator, which outputs a predicted sequence of future settlement trends. Measured data of the predicted sequence segment and the predicted sequence are input into the discriminator. Through adversarial optimization, the generator captures the dynamic evolution of settlement data until the accuracy of the generated settlement curve reaches a preset threshold. S6: The latest collected actual monitoring data is used as a new historical settlement data sequence and input into the trained generator, which outputs a predicted sequence of the segment to be analyzed. Based on the predicted sequence of the segment to be analyzed, the tower foundation stability is assessed and a graded early warning is issued.

[0016] Furthermore, the reference values ​​used in this invention serve as the benchmark scale for evaluating each indicator in the scheme, used to calculate the corresponding impact scores, covering four major categories: prediction accuracy, model training, safety assessment, and input and noise. The prediction accuracy category is determined based on the accuracy requirements of the power transmission and transformation settlement monitoring industry and the general qualified standards for time-series prediction; the model training category combines experience in generative adversarial networks (GANs) training with the characteristics of binary classification tasks; the safety assessment category is derived from the design code for overhead transmission line foundations and the general threshold for foundation stability in geotechnical engineering; and the input and noise category is determined based on the optimal input characteristics of long short-term memory networks and statistical results of historical environmental parameters. All reference values ​​have clear engineering or technical basis. The compensation factor is the weight adjustment coefficient for each indicator in the comprehensive evaluation and parameter update process, used to regulate the degree of influence of different indicators on the final result, covering five application scenarios: historical prediction accuracy evaluation, discrimination feature difference calculation, generator input processing, adversarial training parameter update, and tower foundation stability assessment. All compensation factors were calibrated using methods such as orthogonal experiments, ablation experiments, and effect comparison experiments, combined with engineering experience and task characteristics. The values ​​were set within a reasonable range according to the importance of the corresponding indicators. Core judgment indicators had higher weights, and auxiliary adjustment indicators were adapted to the functional requirements of the corresponding links, thereby ensuring the rationality and stability of the calculation logic of each link.

[0017] In this embodiment, the present invention overcomes the technical pain points of traditional tower foundation settlement prediction by integrating multi-source environmental disturbance data assessment, dynamic data partitioning mechanism, LSTM generative adversarial network adversarial training, and stability hierarchical early warning through an integrated technical solution. This solution relies solely on single settlement data, ignores random environmental disturbances, suffers from poor prediction accuracy with small samples, and has weak model generalization ability. The multi-source environmental noise assessment quantifies the randomness of three core external disturbances: geological, hydrological, and load-related, compensating for the inability of traditional prediction models to adapt to complex field environmental disturbances. The dynamic data partitioning can adapt to historical prediction accuracy. The modeling data range should be adjusted to address the poor model adaptability caused by fixed data sequences. By leveraging the LSTM-GAN network structure, the accuracy of settlement trend prediction in small-sample scenarios is significantly improved through generator fitting of temporal settlement patterns, discriminator enhancement of data features, and expansion of small-sample data. Finally, by combining real-time monitoring data, prediction and hierarchical early warning are achieved, enabling early prediction of potential settlement hazards in power transmission tower foundations. This effectively improves the safety, foresight, and intelligence of power tower foundation operation and maintenance, and avoids safety risks such as transmission line faults and tower tilting / collapse caused by tower foundation settlement imbalance.

[0018] Preferably, the steps for performing a multi-source environmental disturbance stochastic assessment based on surrounding environmental data to obtain the environmental noise sequence include: geological creep parameters, including the creep coefficient, internal friction angle, and cohesion of the soil and rock mass; wherein, the creep coefficient of the soil and rock mass characterizes the rate of deformation of the soil and rock mass under constant load with time, and the unit is usually 1 / year or Pa·s, reflecting the speed of soil and rock creep deformation. It can be performed according to the "Standard for Geotechnical Testing Methods" (GB / T50123-2019), after drilling core samples at the tower foundation, conducting indoor triaxial creep tests or direct shear creep tests, applying constant loads in stages and continuously recording the deformation at different times, and obtaining the parameters through fitting an empirical creep model; or... The internal friction angle, measured in degrees, is one of the core indicators of shear strength of soil and rock mass. It reflects the frictional resistance between soil and rock particles and can be directly measured by indoor direct shear tests or triaxial compression tests, following the same standard as GB / T50123-2019. Alternatively, it can be calculated using regional empirical formulas based on in-situ test results from standard penetration tests and static cone penetration tests. Cohesion, also a core indicator of shear strength of soil and rock mass, is measured in kPa and reflects the bond strength between soil and rock particles. It is consistent with the internal friction angle and can be measured by indoor shear tests or obtained through inversion based on in-situ test data. Hydrological change parameters... The data includes daily precipitation, soil moisture content, and groundwater level elevation. Daily precipitation is the cumulative daily rainfall in the area where the tower base is located, expressed in mm. This can be obtained by accessing daily observation data from the local national meteorological station; or by deploying automatic rain gauges around the tower base to automatically collect data on a daily scale and synchronize it to the monitoring system. Soil moisture content is the volumetric water content of the soil layer within the tower base's influence range, expressed as a percentage, reflecting the soil's wet and dry state. This can be obtained by burying frequency domain reflectance (FDR) or time domain reflectance (TDR) soil moisture sensors at different depths around the tower base and automatically collecting data according to a set monitoring cycle. Groundwater level elevation is the absolute elevation of the groundwater level near the tower base, expressed in meters, reflecting the groundwater level. The dynamic changes can be monitored by installing dedicated water level observation wells within the influence range of the tower foundation, and installing pressure-type or float-type water level gauges to automatically monitor the water level depth. The groundwater level elevation can be calculated by combining the absolute elevation of the wellhead. External load parameters include the vertical load of the tower foundation, horizontal wind load, and conductor tension. Among them, the vertical load of the tower foundation is the total vertical load applied to the top surface of the foundation, including the self-weight of the tower, the self-weight of the conductor and ground wire, the weight of the insulator, and the weight of ice accumulation, in kN. It can be calculated according to the standard value combination of permanent load and variable load in the "Load Code for Overhead Transmission Lines" (DL / T5551-2018). Alternatively, it can be obtained by burying earth pressure sensors at the bottom of the tower foundation for in-situ real-time monitoring.Horizontal wind load is the horizontal load generated by wind acting on the tower body and conductors, measured in kN. It can be calculated according to the wind load formula in the "Design Code for 110kV~750kV Overhead Transmission Lines" (GB50545-2010), by inputting parameters such as local basic wind pressure, tower shape coefficient, conductor windward area, and wind pressure height variation coefficient; or it can be calculated in real time according to the formula by combining the wind speed monitoring data at the tower base. Conductor tension is the axial tensile force of the overhead conductor, measured in kN. Its horizontal component is transmitted to the tower base. It can be directly monitored online by installing tension sensors at the insulator strings in the tension section; or it can be indirectly calculated according to the overhead line mechanics formula based on the conductor type, span, sag, and ambient temperature. Probability density estimation was performed on geological creep parameters, hydrological variation parameters, and external load parameters to obtain the marginal distribution of each surrounding environmental data. Specifically, this included: obtaining historical time-series values ​​of the soil and rock creep coefficient from the geological creep parameters, performing kernel density estimation on these historical time-series values ​​to obtain the marginal distribution of the soil and rock creep coefficient; obtaining historical time-series values ​​of the internal friction angle from the geological creep parameters, performing kernel density estimation on these historical time-series values ​​to obtain the marginal distribution of the internal friction angle; obtaining historical time-series values ​​of cohesion from the geological creep parameters, performing kernel density estimation on these historical time-series values ​​to obtain the marginal distribution of the cohesion; and obtaining historical time-series values ​​of daily precipitation, soil moisture content, and groundwater level elevation from the hydrological variation parameters, and performing probability density estimation on the historical time-series values ​​of daily precipitation, soil moisture content, and groundwater level elevation, respectively. Kernel density estimation was performed on historical time-series values ​​to obtain the marginal distributions of daily precipitation, soil moisture content, and groundwater level. Historical time-series values ​​of tower foundation vertical load, horizontal wind load, and conductor tension were obtained from the external load parameters. Kernel density estimation was then performed on these values ​​to obtain the marginal distributions of tower foundation vertical load, horizontal wind load, and conductor tension. The marginal distributions of soil creep coefficient, internal friction angle, cohesion, daily precipitation, soil moisture content, groundwater level, tower foundation vertical load, horizontal wind load, and conductor tension were collectively used as the marginal distributions of the surrounding environmental data. Based on the marginal distribution of surrounding environmental data, a multidimensional random disturbance sequence corresponding to the historical monitoring period is generated. This multidimensional random disturbance sequence is used as the environmental noise sequence under multi-source environmental interference. Specifically, this includes: obtaining the time span of the historical monitoring period as the generation length of the disturbance sequence; constructing a joint distribution characterizing the correlation between surrounding environmental data based on the marginal distribution of surrounding environmental data using a Copula function; selecting Gaussian Copula as the joint distribution construction function and using the maximum likelihood estimation method to estimate the parameters of the correlation coefficient matrix of the Copula function; and first calculating the Spearman rank correlation coefficient based on the historical time series samples of each environmental parameter. The matrix is ​​used to initialize the correlation structure of the Copula function, and then the optimal joint distribution model is obtained through iterative fitting. Random sampling adopts the Monte Carlo sampling method, sampling once per time step, and the sampling results are arranged in chronological order to obtain a multidimensional random disturbance sequence with the same length as the historical monitoring period. Random sampling is performed from the joint distribution, and the number of samplings is equal to the length of the disturbance sequence generation to obtain a multidimensional random sampled value sequence corresponding to the historical monitoring period. The multidimensional random sampled value sequence is used as a multidimensional random disturbance sequence under multi-source environmental interference. The multidimensional random disturbance sequence is standardized to obtain a standardized multidimensional random disturbance sequence, and the standardized multidimensional random disturbance sequence is used as an environmental noise sequence.

[0019] In this embodiment, the present invention precisely defines nine core environmental impact parameters in three categories: geological creep, hydrological changes, and external loads. It then uses kernel density estimation to obtain the independent marginal distributions of each parameter, and relies on the Copula function to construct a joint distribution of multiple parameters, completing random sampling and standardization processing to ultimately obtain a multidimensional environmental noise sequence that closely matches the actual scenario. This solves the technical shortcomings of traditional environmental disturbance simulations, which only use single random noise, ignore the correlation between multiple environmental parameters, and cannot realistically reproduce complex field disturbance conditions. Kernel density estimation does not require pre-setting parameter distribution types and can accurately adapt to the nonlinear distribution characteristics of various environmental parameters. The Copula function effectively characterizes the coupling relationship between different environmental factors, and standardization processing ensures the dimensional adaptability of noise data and settlement monitoring data. The final generated environmental noise can highly reproduce the real random disturbances during the tower foundation settlement process, providing realistic and reliable disturbance input support for subsequent generative adversarial network training, and significantly improving the environmental adaptability and realism of the model prediction results.

[0020] Preferably, the step of dynamically dividing settlement monitoring data based on historical prediction accuracy parameters to obtain measured data of historical settlement data and predicted sequence segments includes: obtaining historical prediction accuracy parameters, including mean absolute error, root mean square error, and trend fit goodness of fit. The mean absolute error is the average of the absolute deviations between historical predicted and measured values, in mm, reflecting the overall error level of the prediction. It can be automatically calculated by comparing point-by-point after each prediction cycle is completed and the measured settlement data for the corresponding period is obtained. The root mean square error is the square root of the mean square of the prediction error, in mm, and is more sensitive to large errors, reflecting the degree of error dispersion. The trend fit goodness of fit characterizes the degree of matching between the overall trend of the predicted sequence and the measured sequence, with a value range of 0 to 1. The specific calculation formula is as follows: In the formula, Indicates the goodness of fit of the trend. For the predicted settlement data at the i-th time monitoring point, For the measured settlement data at the i-th time monitoring point, The initial settlement data sequence length is calculated as the arithmetic mean of all measured settlement data within the monitoring period, where n is the total number of time monitoring points within the monitoring period. The length of the historical settlement data sequence is obtained as the initial settlement data sequence length. A historical prediction accuracy impact index is generated based on historical prediction accuracy parameters. If the historical prediction accuracy impact index is less than a preset threshold, the initial settlement data sequence length is increased by a preset first proportion to obtain the final settlement data sequence length. If the historical prediction accuracy impact index is greater than or equal to the preset threshold, the initial settlement data sequence length is decreased by a preset second proportion to obtain the final settlement data sequence length. Within the time range covered by the historical monitoring cycle, settlement monitoring data with a length equal to the settlement data sequence length is extracted from the starting time and used as historical settlement data. The remaining data not extracted during the historical monitoring cycle is used as the measured data for the prediction sequence segment.

[0021] The preset threshold value is 0.75, which is calibrated according to the general prediction accuracy requirements of the power transmission and transformation engineering settlement monitoring industry; the preset first proportion is 15%, and the preset second proportion is 10%. The proportion coefficient is determined based on the non-stationary characteristics of the tower foundation settlement time series and the optimal input length range of the LSTM model; the initial settlement data sequence length is taken as 70% of the total length of the historical monitoring cycle, and the minimum length is not less than 30 monitoring cycles to ensure the fit of the time series characteristics; all reference values ​​adopt the industry benchmark value for settlement prediction of tower foundations of the same region and geological type, and the compensation factor is calibrated by the orthogonal test method, with a value range of [0.6, 1.0].

[0022] In this embodiment, the present invention introduces three core prediction accuracy indicators—mean absolute error, root mean square error, and trend fit goodness—to quantify the indicators affecting historical prediction accuracy. Based on the comparison between the indicator values ​​and preset thresholds, the length of the settlement data sequence is adaptively scaled to intelligently separate historical data for modeling from measured data for verification. Through dynamic data matching, when prediction accuracy is low, the modeling data is expanded and temporal features are enriched to improve model fitting ability. When prediction accuracy meets the standard, the data is simplified and redundant information is eliminated to reduce computational cost during model training. Simultaneously, the historical data used for model training and the measured data used for verification are accurately separated, ensuring the scientific rigor and rationality of model training and verification.

[0023] Preferably, the specific steps for generating historical prediction accuracy impact indicators based on historical prediction accuracy parameters include: for the negative indicator, mean absolute error, dividing the mean absolute error reference value by the current historical mean absolute error and normalizing it to obtain the mean absolute error impact score; for the negative indicator, root mean square error, dividing the root mean square error reference value by the current historical root mean square error and normalizing it to obtain the root mean square error impact score; the higher the impact score of the above negative indicators, the higher the prediction accuracy; for the positive indicator, trend fit goodness of fit, dividing the current historical trend fit goodness of fit by the trend fit goodness of fit reference value and normalizing it to obtain the trend fit goodness of fit impact score. The mean absolute error compensation factor and the mean absolute error influence score are interactively processed to obtain the mean absolute error compensation value. The mean absolute error influence score is obtained by analyzing the proportion of the mean absolute error reference value to the current historical mean absolute error. The root mean square error compensation factor and the root mean square error influence score are interactively processed to obtain the root mean square error compensation value. The root mean square error influence score is obtained by analyzing the proportion of the root mean square error reference value to the current historical root mean square error. All influence scores are positive indicators; higher values ​​correspond to higher prediction accuracy. The trend fit goodness compensation factor and the trend fit goodness influence score are interactively processed to obtain the trend fit goodness compensation value. The trend fit goodness influence score is obtained by analyzing the proportion of the trend fit goodness reference value to the current historical trend fit goodness. The mean absolute error compensation value, the root mean square error compensation value, and the trend fit goodness compensation value are coupled, i.e., normalized and then summed, to obtain the historical prediction accuracy influence index.

[0024] The interactive processing involves a weighted multiplication of the compensation factor and the influence score. The compensation factor is a pre-defined proportionality coefficient based on prior knowledge, and the influence score is a normalized dimensionless value. The compensation value is obtained by multiplying the two. For example, when calculating the mean absolute error (MAE) compensation value, assuming the pre-defined MAE compensation factor k is 0.8 based on prior knowledge; if the current historical MAE is 0.12, and the pre-defined MAE reference value is 0.15, then the ratio of the reference value to the current value is used: 0.15 / 0.12 = 1.25. After normalization, the MAE influence score s is 1.0 (or mapped to a set interval according to normalization rules). In this case, the interactive processing performs a weighted multiplication operation: compensation value = k * s. Multiplying the compensation factor 0.8 by the influence score 1.0 yields a final MAE compensation value of 0.80, which reflects the quantitative compensation level of the current prediction accuracy. When the error value increases to 0.18, the reference value / current value = 0.15 / 0.18 ≈ 0.83, which reduces the impact on the score. The compensation value decreases accordingly, which conforms to the positive logic that the smaller the error, the higher the score.

[0025] In this embodiment, the present invention obtains dimensionless influence scores by performing a proportional analysis of the consistent direction of three types of prediction accuracy indicators. Combined with a preset prior compensation factor, a weighted interactive operation is performed to obtain the compensation values ​​for each indicator. Finally, normalization coupling yields a quantified historical prediction accuracy influence index. This multi-indicator weighted coupling quantification method balances prediction numerical error and trend fitting effect. The compensation factor is set based on prior industry knowledge, aligning with the engineering characteristics of tower foundation settlement prediction. Dimensionless processing eliminates interference from differences in the dimensions of different indicators, ultimately achieving a precise, comprehensive, and quantitative assessment of historical prediction accuracy. This provides a precise and reliable quantitative basis for subsequent dynamic data sequence partitioning, improving the accuracy and scientific rigor of data partitioning.

[0026] Preferably, the steps for the discriminator to distinguish the features of real monitoring data and generator predicted data include: extracting the real time-series statistical features of the measured data of the predicted sequence segment and the generated time-series statistical features of the predicted sequence output by the generator. The real time-series statistical features include the real mean, real standard deviation, and real autocorrelation coefficient. The generated time-series statistical features include the generated mean, generated standard deviation, and generated autocorrelation coefficient, with the autocorrelation coefficient being the average of the autocorrelation coefficients at lag 1 and lag 3. The real mean and generated mean are then compared (divided) to obtain the mean deviation impact score. The real standard deviation and generated standard deviation are then compared to obtain the standard deviation impact score. The real autocorrelation coefficient and generated autocorrelation coefficient are then compared to obtain the autocorrelation deviation impact score. The mean deviation compensation factor and the mean deviation impact score are then processed interactively to obtain the mean deviation compensation value. The standard deviation deviation compensation factor and the standard deviation impact score are then processed interactively to obtain the standard deviation deviation compensation value. The autocorrelation deviation compensation factor and the autocorrelation deviation impact score are then processed interactively to obtain the autocorrelation deviation compensation value. The autocorrelation bias compensation value is obtained; the mean bias compensation value, standard deviation bias compensation value, and autocorrelation bias compensation value are coupled to obtain the discrimination feature difference value. The discrimination feature difference value is used to drive the discriminator to distinguish between real monitoring data and generator predicted data. The smaller the discrimination feature difference value, the higher the feature consistency between the generated data and the real data. If the discrimination feature difference value does not exceed the preset discrimination difference threshold, the generator predicted data is marked as valid predicted data that conforms to the real settlement evolution law; if the discrimination feature difference value exceeds the preset discrimination difference threshold, the generator predicted data is marked as invalid predicted data that deviates from the real settlement evolution law.

[0027] The discriminator employs a bidirectional long short-term memory network architecture connecting fully connected layers, outputting the probability value that a sample belongs to real data. This probability value is used to construct the adversarial training loss function. Simultaneously, the discriminant feature difference value is used as a direct criterion for determining the validity of generated data, achieving explicit discrimination at the statistical feature level. This forms a two-layer complementary discriminant logic of deep network feature discrimination and explicit statistical difference judgment. The autocorrelation coefficient is extracted using the mean of lag 1 and lag 3 as the calculation basis to cover short-term and medium-term time-series dependence characteristics. The preset discriminant difference threshold is 0.2, calibrated based on the industry-acceptable deviation range of settlement time-series characteristics.

[0028] In this embodiment, the present invention extracts three core time-series statistical features—mean, standard deviation, and autocorrelation coefficient—from real monitoring data and model prediction data. Through proportion analysis, weighted compensation, and feature coupling, it obtains quantitative discriminant feature difference values, clearly defining that a difference value not exceeding a threshold is valid. The invention characterizes differences from multiple dimensions, including data amplitude, dispersion, and temporal correlation, and uses a preset threshold to determine the validity of the prediction data. Simultaneously, it clarifies the functional boundaries and complementary relationships between the network probability output and the statistical difference values. This multi-dimensional compensation and coupling quantitative judgment method avoids the one-sidedness of single-feature discrimination, accurately identifying whether the generator's prediction data conforms to the actual settlement evolution pattern. This effectively strengthens the discriminator's feature recognition capability, forcing the generator to learn more accurate settlement time-series features, and further improving the authenticity and effectiveness of the prediction sequence.

[0029] Preferably, the step of inputting historical settlement data and random noise into a generator to output a predicted sequence of future settlement trends includes: performing a ratio analysis of the length of the historical settlement data sequence with a sequence length reference value to obtain a sequence length influence score; performing interactive processing of the sequence length compensation factor and the sequence length influence score to obtain a sequence length compensation value; obtaining the perturbation variance of the environmental noise sequence, performing a ratio analysis of the perturbation variance of random noise with a perturbation variance reference value to obtain a noise intensity influence score; performing interactive processing of the noise intensity compensation factor and the noise intensity influence score to obtain a noise intensity compensation value; performing a sliding window truncation on the historical settlement data according to the sequence length compensation value to obtain a generator input feature sequence; scaling the random noise amplitude according to the noise intensity compensation value to obtain a scaled noise sequence; aligning the generator input feature sequence and the scaled noise sequence in the time dimension, concatenating them in the feature dimension, inputting them into the Long Short-Term Memory (LSTM) network in the generator, and having the LSM network fit the settlement time-series change pattern to output a predicted sequence of future settlement trends corresponding to the period to be predicted.

[0030] In this embodiment, the present invention adaptively completes the sliding window truncation of historical settlement data and the scaling of random noise amplitude by quantifying the influence scores and compensation values ​​of two core input parameters: sequence length and noise intensity. Finally, the spliced ​​features are input into the LSTM network to generate the prediction sequence. By dynamically adjusting the length of the input data window, the core temporal features of settlement data at different time periods can be accurately captured. The adaptive scaling of noise amplitude can match the environmental interference intensity under different working conditions. The feature dimension splicing input method realizes the deep integration of settlement temporal features and environmental interference features, allowing the LSTM network to more comprehensively and accurately fit the dynamic evolution law of tower foundation settlement, significantly improving the accuracy and adaptability of future settlement trend prediction sequences.

[0031] Preferably, the step of enabling the generator to capture the dynamic evolution of settlement data through adversarial optimization until the accuracy of the generated settlement curve reaches a preset threshold includes: analyzing the proportion of the accuracy of the current generator's generated prediction sequence to an accuracy reference value to obtain an accuracy impact score; interactively processing the accuracy compensation factor and the accuracy impact score to obtain an accuracy compensation value; if the accuracy compensation value is greater than or equal to the preset threshold, then the current generator is used as a trained generator; if the accuracy compensation value is less than the preset threshold, then obtaining the discriminant loss value of the discriminator for the prediction sequence and the generator's generation loss value for the generated prediction sequence, wherein the generator loss is a weighted sum of mean squared error loss and adversarial loss, and the discriminator loss is a binary cross-entropy loss; parameter updates use the Adam optimizer, updating the discriminator twice and then the generator once in each training round. Next, the discriminant loss compensation factor and the discriminant loss influence score are interactively processed to obtain the discriminant parameter update amount. The discriminant loss influence score is obtained by analyzing the proportion of the current discriminant loss value and the discriminant loss reference value. The generator loss compensation factor and the generator loss influence score are interactively processed to obtain the generator parameter update amount. The generator loss influence score is obtained by analyzing the proportion of the current generator loss value and the generator loss reference value. The discriminant network weights are adjusted using the discriminant parameter update amount, and the generator network weights are adjusted using the generator parameter update amount to complete one round of adversarial training. Then, the process returns to the step of inputting historical settlement data and random noise into the generator until the accuracy compensation value reaches the preset threshold.

[0032] The generator loss function is a weighted sum of mean squared error loss and adversarial loss, while the discriminator loss function uses binary cross-entropy loss. Parameter updates utilize the Adam optimizer, with the generator learning rate set to 0.001 and the discriminator learning rate set to 0.0005 to ensure the stability of adversarial training. Accuracy is measured using the coefficient of determination R between the predicted and measured sequences. 2 The calculations are performed with an accuracy reference value of 0.85, which corresponds to the industry's acceptable accuracy standard for settlement prediction. In each round of training, the discriminator is updated twice and the generator is updated once to maintain the adversarial balance.

[0033] In this embodiment, the present invention determines whether the model training meets the standard by quantifying the prediction accuracy compensation value. For models that do not meet the standard, the network weights are adaptively updated based on the quantified compensation values ​​of the discriminant loss and the generation loss, and multiple rounds of iterative adversarial training are completed until the model accuracy meets the standard. With prediction accuracy as the core criterion, the practicality of the final model's prediction is guaranteed. By updating the discriminator and generator parameters through bidirectional quantization, the network training bias is accurately corrected, and the model convergence is accelerated. At the same time, the generator continuously learns the dynamic evolution law of sedimentation data, and the discriminator continuously optimizes its feature discrimination ability, realizing refined iteration of adversarial training. Finally, a high-precision and high-stability optimal prediction model is obtained, which completely solves the technical problems of insufficient fitting and large prediction bias in traditional models.

[0034] Preferably, the steps for assessing and classifying the stability of the tower foundation based on the predicted sequence of the sequence segment to be analyzed include: extracting the maximum cumulative settlement and the average settlement rate from the predicted sequence of the sequence segment to be analyzed; performing a ratio analysis between the maximum cumulative settlement and a cumulative settlement reference value to obtain a cumulative settlement influence score; performing a ratio analysis between the average settlement rate and a settlement rate reference value to obtain a settlement rate influence score; performing interactive processing between the cumulative settlement compensation factor and the cumulative settlement influence score to obtain a cumulative settlement assessment value; and performing interactive processing between the settlement rate compensation factor and the settlement rate influence score to obtain a settlement rate assessment value. The cumulative settlement assessment value and the settlement rate assessment value are superimposed to obtain the tower base stability index; based on the comparison between the tower base stability index and the multi-level warning thresholds, the warning level is determined and a graded warning signal is output; when the tower base stability index is less than the first warning threshold, a normal state signal is issued; when the tower base stability index is between the first warning threshold and the second warning threshold, a attention warning signal is issued; when the tower base stability index is between the second warning threshold and the third warning threshold, a warning signal is issued; when the tower base stability index is greater than or equal to the third warning threshold, a danger warning signal is issued.

[0035] The first warning threshold is set at 0.4, the second at 0.7, and the third at 0.9. These thresholds are determined based on the correspondence between settlement limits and tower foundation safety levels in the "Technical Specification for Foundation Design of Overhead Transmission Lines." Both the cumulative settlement compensation factor and the settlement rate compensation factor are set at 0.5, with equal weight, jointly characterizing the safety status of tower foundation settlement. The maximum cumulative settlement is taken as the cumulative settlement value at the end of the prediction sequence, and the average settlement rate is taken as the average settlement value per unit monitoring period within the prediction sequence.

[0036] In this embodiment, the present invention extracts two core evaluation indicators—maximum cumulative settlement and average settlement rate—from the predicted sequence, and obtains the tower foundation stability index through quantitative compensation coupling. It then uses multi-level thresholds to achieve four-level graded early warning output. The stability index, coupled with multiple indicators, comprehensively and objectively reflects the actual operating status of the tower foundation. The four-level graded early warning mechanism has clear division standards and distinct levels, accurately distinguishing between normal, caution, warning, and danger operating conditions. This provides maintenance personnel with precise and differentiated maintenance guidelines, enabling early detection, early warning, and early handling of tower foundation settlement hazards. This significantly improves the safety and maintenance efficiency of power transmission and transformation tower foundation operation, ensuring the long-term stable operation of the power transmission system.

[0037] like Figure 2 The diagram shows the structure of a power transmission tower foundation settlement trend prediction system based on a generative adversarial network (GAN) provided in this application embodiment. The system includes: a data acquisition module for collecting settlement monitoring data and surrounding environmental data from historical monitoring periods of the power transmission tower foundation; an environmental noise assessment module for performing multi-source environmental interference randomness assessment based on the surrounding environmental data to obtain an environmental noise sequence; a data partitioning module for dynamically partitioning the settlement monitoring data based on historical prediction accuracy parameters to obtain measured data of historical settlement data and predicted sequence segments; a model building module for constructing a generative adversarial network, which includes a generator and a discriminator, with the generator built using a long short-term memory network; an adversarial training module for inputting historical settlement data and random noise into the generator, outputting a predicted sequence, and then inputting the measured data of the predicted sequence segment and the predicted sequence into the discriminator for alternating adversarial training until the accuracy of the generated settlement curve reaches a preset threshold; a trend prediction module for inputting the latest collected real monitoring data into the trained generator, outputting a predicted sequence of the sequence segment to be analyzed; and a stability assessment and early warning module for performing tower foundation stability assessment and graded early warning based on the predicted sequence of the sequence segment to be analyzed.

[0038] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for predicting the settlement trend of power transmission and transformation tower foundations based on generative adversarial networks.

[0039] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)). Where there is no conflict, the solutions in the above embodiments can be used in combination.

[0040] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope and intent of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks, characterized in that, Includes the following steps: S1. Collect settlement monitoring data and surrounding environmental data of the power transmission and transformation tower base during the historical monitoring period. The surrounding environmental data includes geological creep parameters, hydrological change parameters and external load parameters. Based on the surrounding environmental data, perform a multi-source environmental interference randomness assessment to obtain an environmental noise sequence, which is used as the random noise of the surrounding environment of the power transmission and transformation tower base. S2, dynamically divide the settlement monitoring data based on the historical prediction accuracy parameters to obtain the measured data of historical settlement data and prediction sequence segments; S3, Construct a generative adversarial network, which includes a generator and a discriminator; the generator is used to fit the temporal variation law of settlement and output the future settlement trend prediction sequence corresponding to the period to be predicted; the discriminator is used to distinguish the features of real monitoring data and generator prediction data, so as to achieve effective expansion and feature enhancement of small sample data. S4, perform alternating adversarial training on the generative adversarial network: input historical settlement data and random noise into the generator, and output a predicted sequence of future settlement trends; input the measured data of the predicted sequence segment and the predicted sequence into the discriminator, and through adversarial optimization, enable the generator to capture the dynamic evolution law of settlement data until the generator generates a settlement curve with an accuracy rate that reaches a preset threshold. S5 takes the latest collected real monitoring data as a new historical settlement data sequence, inputs it into the trained generator, outputs the predicted sequence of the sequence segment to be analyzed, and performs tower foundation stability assessment and graded early warning based on the predicted sequence of the sequence segment to be analyzed.

2. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 1, characterized in that: The step of performing a multi-source environmental interference stochastic assessment based on the surrounding environmental data to obtain an environmental noise sequence includes: The geological creep parameters include the creep coefficient of the soil and rock mass, the internal friction angle, and the cohesion; the hydrological variation parameters include daily precipitation, soil moisture content, and groundwater level; the external load parameters include the vertical load on the tower foundation, the horizontal wind load, and the conductor tension. Probability density estimation is performed on the geological creep parameters, hydrological variation parameters, and external force load parameters to obtain the marginal distribution of each surrounding environmental data. Based on the edge distribution of the surrounding environmental data, a multidimensional random disturbance sequence corresponding to the historical monitoring period is generated, and the multidimensional random disturbance sequence is used as the environmental noise sequence under the multi-source environmental interference.

3. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 1, characterized in that: The steps for dynamically dividing settlement monitoring data based on historical prediction accuracy parameters to obtain measured data of historical settlement data and prediction sequence segments include: Obtain the historical prediction accuracy parameters, including mean absolute error, root mean square error, and trend fit goodness; obtain the length of the historical settlement data sequence as the initial settlement data sequence length; Generate historical prediction accuracy impact indicators based on the historical prediction accuracy parameters; If the historical prediction accuracy impact index is less than the preset division threshold, then the initial settlement data sequence length is increased by a preset first ratio to obtain the settlement data sequence length; If the historical prediction accuracy impact index is greater than or equal to the preset division threshold, then the initial settlement data sequence length is reduced by a preset second ratio to obtain the settlement data sequence length. Within the time range covered by the historical monitoring period, settlement monitoring data with a length equal to the length of the settlement data sequence is extracted from the starting time and used as the historical settlement data; the remaining data that was not extracted in the historical monitoring period is used as the measured data of the predicted sequence segment.

4. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 3, characterized in that: The specific steps for generating historical prediction accuracy impact indicators based on the historical prediction accuracy parameters include: The mean absolute error compensation factor and the mean absolute error influence score are interactively processed to obtain the mean absolute error compensation value. The mean absolute error influence score is obtained by analyzing the proportion of the mean absolute error reference value and the current historical mean absolute error. The root mean square error compensation factor and the root mean square error influence score are interactively processed to obtain the root mean square error compensation value. The root mean square error influence score is obtained by analyzing the proportion of the root mean square error reference value and the current historical root mean square error. The trend fit goodness compensation factor and the trend fit goodness influence score are interactively processed to obtain the trend fit goodness compensation value. The trend fit goodness influence score is obtained by analyzing the proportion of the trend fit goodness reference value and the current historical trend fit goodness. The average absolute error compensation value, root mean square error compensation value, and trend fit goodness compensation value are coupled to obtain the historical prediction accuracy impact index.

5. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 1, characterized in that: The steps by which the discriminator distinguishes the features of real monitoring data from generator prediction data include: Extract the true time-series statistical features of the measured data of the predicted sequence segment, and the generated time-series statistical features of the predicted sequence output by the generator. The true time-series statistical features include the true mean, the true standard deviation, and the true autocorrelation coefficient. The generated time-series statistical features include the generated mean, the generated standard deviation, and the generated autocorrelation coefficient. The influence scores of mean deviation, mean and standard deviation are analyzed by comparing the proportions of the true mean and the generated mean; the influence scores of standard deviation, standard deviation, and autocorrelation coefficient are analyzed by comparing the proportions of the true autocorrelation coefficient and the generated autocorrelation coefficient. The mean deviation compensation factor and the mean deviation influence score are interactively processed to obtain the mean deviation compensation value; the standard deviation deviation compensation factor and the standard deviation influence score are interactively processed to obtain the standard deviation deviation compensation value; the autocorrelation deviation compensation factor and the autocorrelation deviation influence score are interactively processed to obtain the autocorrelation deviation compensation value. The mean deviation compensation value, standard deviation deviation compensation value and autocorrelation deviation compensation value are coupled to obtain the discriminant feature difference value, which is used to drive the discriminator to distinguish between real monitoring data and generator predicted data. If the discriminant feature difference value does not exceed the preset discriminant difference threshold, the generator prediction data is marked as valid prediction data that conforms to the actual settlement evolution law; if the discriminant feature difference value exceeds the preset discriminant difference threshold, the generator prediction data is marked as invalid prediction data that deviates from the actual settlement evolution law.

6. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 4, characterized in that: The step of inputting historical settlement data and random noise into the generator and outputting a predicted sequence of future settlement trends includes: The length of the historical settlement data sequence is compared with the sequence length reference value to obtain the sequence length influence score; the sequence length compensation factor and the sequence length influence score are interactively processed to obtain the sequence length compensation value. The disturbance variance of the random noise is compared with the disturbance variance reference value to obtain the noise intensity influence score; the noise intensity compensation factor and the noise intensity influence score are processed interactively to obtain the noise intensity compensation value. The historical settlement data is truncated using a sliding window based on the sequence length compensation value to obtain the generator input feature sequence; the random noise is scaled based on the noise intensity compensation value to obtain a scaled noise sequence. The generator input feature sequence and the scaled noise sequence are concatenated along the feature dimension and then input into the Long Short-Term Memory (LSTM) network in the generator. The LSM network fits the settlement time series variation pattern and outputs a future settlement trend prediction sequence corresponding to the time period to be predicted.

7. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 5, characterized in that: The step of using adversarial optimization to enable the generator to capture the dynamic evolution of settlement data until the accuracy of the generated settlement curve reaches a preset threshold includes: The accuracy of the predicted sequence generated by the current generator is compared with the accuracy reference value to obtain the accuracy impact score; the accuracy compensation factor and the accuracy impact score are interacted to obtain the accuracy compensation value. If the accuracy compensation value is greater than or equal to the preset threshold, then the current generator is used as the trained generator. If the accuracy compensation value is less than a preset threshold, then the discriminant loss value of the discriminator for the predicted sequence and the generation loss value of the generator for generating the predicted sequence are obtained; the discriminant loss compensation factor and the discriminant loss influence score are interactively processed to obtain the discriminator parameter update amount, wherein the discriminant loss influence score is obtained by proportional analysis of the current discriminant loss value and the discriminant loss reference value; the generation loss compensation factor and the generation loss influence score are interactively processed to obtain the generator parameter update amount, wherein the generation loss influence score is obtained by proportional analysis of the current generation loss value and the generation loss reference value; The discriminator network weights are adjusted using the discriminator parameter update amount, and the generator network weights are adjusted using the generator parameter update amount to complete one round of adversarial training. Then, the process returns to the step of inputting historical settlement data and random noise into the generator until the accuracy compensation value reaches a preset threshold.

8. The method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in claim 1, characterized in that: The steps for assessing tower base stability and providing graded early warning based on the predicted sequence of the sequence segment to be analyzed include: Extract the maximum cumulative settlement and average settlement rate from the predicted sequence of the sequence segment to be analyzed; The cumulative settlement influence score is obtained by comparing the maximum cumulative settlement with the cumulative settlement reference value; the settlement rate influence score is obtained by comparing the average settlement rate with the settlement rate reference value. The cumulative settlement compensation factor and the cumulative settlement influence score are interactively processed to obtain the cumulative settlement assessment value; the settlement rate compensation factor and the settlement rate influence score are interactively processed to obtain the settlement rate assessment value. The cumulative settlement assessment value and the settlement rate assessment value are superimposed to obtain the tower foundation stability index; Based on the comparison between the tower base stability index and the multi-level warning thresholds, the warning level is determined and a graded warning signal is output; when the tower base stability index is less than the first warning threshold, a normal state signal is issued; when the tower base stability index is between the first warning threshold and the second warning threshold, a attention warning signal is issued; when the tower base stability index is between the second warning threshold and the third warning threshold, a warning signal is issued; when the tower base stability index is greater than or equal to the third warning threshold, a danger warning signal is issued.

9. A system applying the method for predicting the settlement trend of power transmission tower foundations based on generative adversarial networks as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect settlement monitoring data and surrounding environmental data of the power transmission and transformation tower foundation during the historical monitoring period; An environmental noise assessment module is used to perform a multi-source environmental interference stochasticity assessment based on the surrounding environmental data to obtain an environmental noise sequence. The data segmentation module is used to dynamically segment settlement monitoring data based on historical prediction accuracy parameters, and obtain measured data of historical settlement data and prediction sequence segments; The model building module is used to build a generative adversarial network, which includes a generator and a discriminator. The generator is built using a long short-term memory network. The adversarial training module is used to input historical settlement data and the random noise into the generator, output a prediction sequence, and then use the measured data of the prediction sequence segment and the prediction sequence into the discriminator for alternating adversarial training until the accuracy of the generated settlement curve reaches a preset threshold. The trend prediction module is used to input the latest collected real monitoring data into the trained generator and output the predicted sequence of the sequence segment to be analyzed. The stability assessment and early warning module is used to perform tower base stability assessment and graded early warning based on the predicted sequence of the sequence segment to be analyzed.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for predicting the settlement trend of power transmission and transformation tower foundations based on generative adversarial networks as described in any one of claims 1-9.

Images