A Deformation Prediction Method for Cross-Sea Bridges Based on STL-ARIMA-Meteorological Coupled Model

Abstract

This invention provides a method for predicting the deformation of cross-sea bridges based on an STL-ARIMA-meteorological coupled model, relating to the field of bridge monitoring technology. The method includes: collecting deformation and meteorological data of the cross-sea bridge and fully aligning the deformation and meteorological data on a time scale; constructing deformation and meteorological features based on the deformation and meteorological data of the cross-sea bridge, respectively; optimizing the seasonal cycle and trend window of the STL algorithm based on the deformation data, and decomposing the cumulative settlement time series into a trend term, a seasonal term, and a residual term; modeling the residual term obtained from the STL decomposition with meteorological factors using random forest regression to generate meteorological influence components, and then calculating the meteorological correction residual; superimposing the trend term obtained from the STL decomposition with the meteorological correction residual to generate a trend-correction residual, which serves as the ARIMAX modeling target; expanding the search space using the auto_arima algorithm to drive ARIMAX modeling with meteorological features; and fusing the ARIMAX model output, historical seasonal term replication, and meteorological feature prediction results to achieve multi-scale deformation reconstruction.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring technology, and in particular to a method for predicting the deformation of cross-sea bridges based on the STL-ARIMA-meteorological coupling model. Background Technology

[0002] With the rapid development of cross-sea transportation networks, long-span bridges face severe structural safety challenges in complex marine environments (salt spray corrosion, typhoons, ocean current erosion, drastic temperature and humidity changes, etc.). These dynamic meteorological loads cause multi-scale deformation of bridges, including long-term foundation settlement, seasonal expansion and contraction caused by temperature, and short-term vibrations caused by typhoons. Traditional monitoring methods are insufficient to achieve high-precision deformation prediction of cross-sea bridges.

[0003] Current methods for predicting the deformation of cross-sea bridges are mainly divided into three categories:

[0004] (1) The physical mechanism model is based on the finite element analysis to construct the structure-environment coupling equation and simulate the deformation response by inputting meteorological parameters. However, it depends on precise boundary conditions and material parameters. Time-varying damage (such as local corrosion and foundation scour) will cause the model to be inaccurate. Moreover, the computation cost of solving a large number of differential equations is high, which makes it difficult to meet the real-time prediction requirements. It also lacks the ability to model the nonlinear effects of extreme events such as strong typhoons.

[0005] (2) In classic time series models, a single ARIMA model is only applicable to stationary series. However, the deformation data of cross-sea bridges exhibits strong trends, seasonality, and random abrupt changes due to meteorological interference, violating the stationarity assumption. The STL-ARIMA combined model, which is closer to the existing technology, uses seasonal decomposition (STL) to split the original deformation data into trend, seasonal, and residual terms, and then establishes ARIMA sub-models for each component. The core defects of this scheme are: the lack of real-time meteorological parameters (wind speed / temperature and humidity / atmospheric pressure) leads to the failure of prediction for sudden meteorological events (such as typhoons); random ocean interference (splash impact) makes the residuals highly heteroscedastic, and the prediction accuracy of traditional ARIMA for abrupt residual changes drops sharply; the trend term still implies meteorological influences and lacks an explicit meteorological-deformation correlation mechanism, resulting in significant cumulative errors in long-term predictions.

[0006] (3) Machine learning models (such as LSTM and SVR) fit deformation patterns through nonlinear mapping, but they rely on large-scale labeled data while the actual measured datasets of cross-sea bridges are scarce. The black box nature of the model leads to poor interpretability and weak generalization ability for untrained extreme weather scenarios.

[0007] Therefore, there is an urgent need to construct a meteorological-structural response coupled prediction framework that, while inheriting the advantages of STL-ARIMA time series modeling, deeply integrates the multi-scale influence mechanisms of meteorological factors. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method for predicting the deformation of a cross-sea bridge based on the STL-ARIMA-meteorological coupling model, comprising: S1, data acquisition and time alignment: acquiring deformation data and meteorological data of the cross-sea bridge, associating the deformation data and meteorological data of the cross-sea bridge to the same time dimension, setting the timestamp as the index, and resampling the deformation data and meteorological data by setting a fixed time interval to make the deformation data and meteorological data fully aligned in time scale. The deformation data is the cumulative settlement, and the meteorological data includes temperature, precipitation and wind speed.

[0009] S2. Construction of dynamic feature engineering: Deformation features and meteorological features are constructed based on the deformation data and meteorological data of the cross-sea bridge, respectively.

[0010] S3. Adaptive STL parameter optimization: Based on deformation data, optimize the seasonal cycle and trend window of the STL algorithm, and use the parameter-optimized STL algorithm to decompose the cumulative settlement time series into trend term, seasonal term and residual term.

[0011] S4. Meteorological-Residual Coupling Modeling: A random forest regression model is trained using residual terms obtained from STL algorithm decomposition, meteorological features, and deformation features. The trained random forest regression model is used to predict the influence of meteorological factors on the residuals. The meteorological correction residuals are calculated based on the influence of meteorological factors on the residuals and residual terms obtained from STL algorithm decomposition.

[0012] S5. Trend-Residual Reconstruction: The trend term obtained by decomposing the STL algorithm is added to the meteorological correction residual to obtain the trend-correction residual.

[0013] S6. Weather-driven ARIMAX modeling: The auto_arima algorithm is used to select the optimal parameter combination of the ARIMAX model. The trend-corrected residuals and meteorological features are used to fit the ARIMAX model. The fitted ARIMAX model is then used to predict the trend-corrected residuals.

[0014] S7. Multi-scale forecast reconstruction: Calculate the predicted values ​​of future meteorological impacts and extended seasonal terms, and use the trend-corrected residual predicted values, the predicted values ​​of future meteorological impacts and extended seasonal terms to calculate the predicted values ​​of settlement.

[0015] S8. Model Evaluation and Visualization: The accuracy of settlement prediction is evaluated using the coefficient of determination R2, root mean square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE, normalized square error NSE, and residual standard deviation RSTD. A comparison chart of historical observations, model fitted values, and future predicted values ​​is plotted, and the current time point is marked as the boundary between the past and the future.

[0016] Optionally, step S2 includes: calculating the difference between the highest temperature and the lowest temperature to obtain the temperature change rate, calculating the difference between the current cumulative precipitation and the previous cumulative precipitation to obtain the precipitation change rate, and constructing a meteorological feature consisting of average temperature, highest temperature, lowest temperature, temperature change rate, average wind speed, maximum wind speed, cumulative precipitation and precipitation change rate.

[0017] The rolling mean of the cumulative settlement over the previous three periods and the rolling standard deviation over the previous six periods were calculated. Anomaly detection using a rolling window was used to define the cumulative settlement exceeding the mean ± 3 times the standard deviation as outliers. The PCHIP interpolation method was used for correction. Deformation characteristics were constructed by the cumulative settlement over the previous three periods, the rolling mean of the previous three periods, and the rolling standard deviation over the previous six periods.

[0018] Optionally, step S3 includes: calculating the correlation of cumulative settlement under different lag orders through the autocorrelation function, identifying significant peaks by combining the peak detection algorithm, adding 1 to the lag order corresponding to the first significant peak as the optimal seasonal period, adding 1 to the optimal seasonal period as the optimal trend window and adjusting it to an odd number, and using the parameter-optimized STL algorithm to decompose the time series of cumulative settlement into trend terms, seasonal terms and residual terms.

[0019] Optionally, step S4 includes: using the residual terms obtained by the STL algorithm decomposition as the target variable, meteorological features and deformation features as input features, and using a random forest regression model to establish a nonlinear mapping relationship between meteorological features and residual terms.

[0020] The meteorological influence component is obtained by using a trained random forest regression model to predict the influence of meteorological factors on the residuals.

[0021] The meteorological correction residual is obtained by calculating the difference between the residual terms decomposed by the STL algorithm and the meteorological impact components predicted by the random forest model.

[0022] Optionally, step S6 includes: using the auto_arima algorithm to set the maximum values ​​of the parameters p, d, and q of the ARIMAX model to expand the parameter search range and extend the optional parameter space; using the Akaike information criterion as the optimization objective, selecting the parameters p, d, and q corresponding to the lowest Akaike information criterion as the optimal parameter combination for the ARIMAX model.

[0023] When constructing an ARIMAX model based on the optimal parameter combination, the core relationship can be expressed by the formula:

[0024] .

[0025] in, φ represents the trend-corrected residual prediction value, φ represents the optimal parameter combination selected by the auto_arima algorithm, Y represents the trend-corrected residual, and X represents the meteorological characteristics.

[0026] Optionally, step S7 includes: performing linear extrapolation of meteorological characteristics based on the most recent 12 meteorological data periods to obtain the predicted value of future meteorological impacts.

[0027] The extended seasonal term is obtained by copying the last complete cycle of the historical seasonal term and extending it to the forecast period.

[0028] The settlement prediction is obtained by summing the trend-corrected residual prediction, the extended seasonal term, and the future meteorological impact prediction.

[0029] By adopting the above technical solution, the present invention has at least the following beneficial effects:

[0030] (1) This invention introduces the temperature / precipitation change rate into meteorological features and the rolling mean of the first 3 periods and the rolling standard deviation of the first 6 periods into deformation features. The cumulative settlement exceeding the mean ± 3 times the standard deviation is defined as an outlier. Combined with PCHIP interpolation correction, dynamic enhancement and noise robustness processing of meteorological data and deformation data are achieved.

[0031] (2) The present invention designs a meteorological-residual coupling modeling mechanism, which uses random forest regression to model the residual terms obtained by STL decomposition and meteorological factors to generate meteorological influence components, and then calculates meteorological correction residuals to explicitly separate the interference of meteorological noise on deformation.

[0032] (3) This invention uses meteorological-driven ARIMAX modeling, superimposes the trend term obtained by STL decomposition with the meteorological correction residual to generate the trend-correction residual, and uses it as the ARIMAX modeling target. The search space is expanded by using the auto_arima algorithm, and the ARIMAX modeling is driven by meteorological features, which effectively improves the deformation prediction accuracy of bridges.

[0033] (4) This invention integrates ARIMAX model output, historical seasonal term replication and meteorological feature prediction results to achieve multi-scale deformation reconstruction. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0035] Figure 1 This is a flowchart illustrating a method for predicting the deformation of a cross-sea bridge based on an STL-ARIMA-meteorological coupling model, provided in an embodiment of this disclosure.

[0036] Figure 2 A comparison chart of historical observations, model fit values, and future predicted values ​​provided for embodiments of this disclosure. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] The existing STL-ARIMA combinatorial model has three core flaws:

[0039] First, the coupling mechanism of meteorological factors is lacking. The model relies solely on historical deformation data and fails to incorporate real-time meteorological parameters (wind speed, temperature, humidity, atmospheric pressure, etc.), leading to the failure of deformation prediction under sudden meteorological events (such as typhoons and rainstorms). The fundamental reason is that STL decomposition only extracts the trend / seasonal / residual components of the data itself, without establishing an explicit mapping relationship between meteorological variables and each component. The meteorological influence is implicit in the residual term, making it impossible to distinguish the contribution of specific factors. The difficulty in solving this problem lies in the nonlinear time-varying coupling between meteorological parameters and deformation components (such as the critical threshold effect of wind speed on bridge vibration), making it difficult for traditional linear models to construct a dynamic weighting mechanism.

[0040] Second, the handling of strong heteroscedasticity in the residual terms is insufficient. Random disturbances in the marine environment (splash impact, ship disturbances) cause abrupt changes in the variance of the residual terms over time. Traditional ARIMA models assume that the residuals are stationary, leading to significant deviations in predicted values ​​under disturbance events. This is because the residuals after STL decomposition still contain unmodeled meteorological abrupt noise, and standard ARIMA has poor adaptability to heteroscedastic sequences. The difficulty in solving this problem lies in the fact that heteroscedasticity is triggered by external random events, requiring the establishment of a correlation model between the residuals and real-time environmental parameters. However, the causal relationship between meteorological disturbances and deformation responses is difficult to quantify.

[0041] Third, there is the cumulative error in long-term forecasts. The trend term implies unexposed meteorological influences (such as material expansion and contraction caused by temperature and humidity cycles), and the continuous accumulation of meteorological effects in long-term forecasts causes forecast drift. This is because STL decomposition cannot completely separate meteorological-driven spurious trends (such as slow deformation caused by rising annual average temperature) and lacks a feedback correction mechanism between the meteorological and trend terms. The technical obstacle lies in the lag effect of meteorology on long-term deformation (such as salt spray corrosion taking several months to manifest), which makes data labeling difficult, and traditional time series models cannot capture cross-period dependencies.

[0042] During the improvement process, three technical bottlenecks were encountered: First, the influence weight of meteorological factors changes dynamically with the type of event (e.g., wind speed dominates vibration during typhoons, while its influence is negligible on ordinary days), which the fixed-weight model cannot adapt to; second, the difference in sampling frequency between meteorological data (minute-level) and deformation data (hour-level) makes feature alignment difficult, and forced interpolation will introduce noise; third, simply adding meteorological input to STL-ARIMA will destroy the mathematical assumption of STL decomposition (the asynchronous periodicity of meteorological variables and the original sequence), and the decomposition logic needs to be reconstructed.

[0043] like Figure 1 As shown, for a certain cross-sea bridge, this disclosure provides a method for predicting the deformation of a cross-sea bridge based on an STL-ARIMA-meteorological coupled model, including:

[0044] S1, Data Acquisition and Time Alignment.

[0045] Deformation data (target variable, including cumulative settlement) and meteorological data (independent variables, including temperature, precipitation and wind speed) of the cross-sea bridge are collected. The deformation data and meteorological data of the cross-sea bridge are associated with the same time dimension, and the timestamp is set as the index. The deformation data and meteorological data are resampled by setting a fixed time interval (12 days) to make the deformation data and meteorological data fully aligned with the time scale.

[0046] S2, Dynamic Feature Engineering Construction.

[0047] The temperature change rate (temp_change, reflecting the intensity of temperature fluctuations) is obtained by calculating the difference between the highest and lowest temperatures. The precipitation change rate (precip_change, reflecting the increase in precipitation) is obtained by calculating the difference between the current period's cumulative precipitation and the previous period's cumulative precipitation. A meteorological feature is constructed consisting of average temperature (average temp), maximum temperature (max temp), minimum temperature (min temp), temperature change rate (temp_change), average wind speed (average wind), maximum wind speed (max wind), cumulative precipitation, and precipitation change rate (precip_change).

[0048] The cumulative settlement over the previous three periods is introduced as a lag feature (lag1_subsidence, lag2_subsidence, lag3_subsidence, used to capture the time dependence of deformation). The rolling mean of the cumulative settlement over the previous three periods (rolling_3_mean, reflecting the short-term deformation trend) and the rolling standard deviation of the cumulative settlement over the previous six periods (rolling_6_std, reflecting the degree of deformation fluctuation) are calculated. The cumulative settlement exceeding the mean ± 3 times the standard deviation is defined as an outlier and corrected using the PCHIP interpolation method. The deformation feature is constructed by the cumulative settlement over the previous three periods (lag1_subsidence, lag2_subsidence, lag3_subsidence), the rolling mean of the previous three periods (rolling_3_mean), and the rolling standard deviation of the previous six periods (rolling_6_std).

[0049] S3, Adaptive STL parameter optimization.

[0050] The correlation of cumulative settlement at different lag orders is calculated using the autocorrelation function (ACF). Significant peaks are identified using a peak detection algorithm. The lag order corresponding to the first significant peak (the lag number of a few past periods used in the model, where "order" represents the specific use of "previous periods" of data; if no peak is detected, the default period is 30) is incremented by 1 to obtain the optimal seasonal period. The optimal seasonal period is incremented by 1 to obtain the optimal trend window and adjusted to an odd number (while limiting it to no more than 90% of the data length). The parameter-optimized STL algorithm is used to decompose the cumulative settlement time series into a trend term (stl_result.trend, reflecting the long-term deformation trend), a seasonal term (stl_result.seasonal, reflecting periodic fluctuations), and a residual term (stl_result.resid, reflecting random fluctuations).

[0051] S4, Meteorological-Residual Coupled Modeling.

[0052] The residual term (resid_component=stl_result.resid) obtained by STL algorithm decomposition is used as the target variable. Meteorological features (average temperature, maximum temperature, minimum temperature, temperature change rate, average wind speed, maximum wind speed, cumulative precipitation, and precipitation change rate) and deformation features (cumulative subsidence, lag1_subsidence, lag2_subsidence, lag3_subsidence, rolling 3_mean, and rolling 6_std) of the first 3 periods are used as input features. A random forest regression model (with 500 trees and a maximum depth of 12) is used to establish a nonlinear mapping relationship between meteorological features and residual term.

[0053] Meteorological characteristics are standardized (by using met_scaler to convert the mean to 0 and the standard deviation to 1, thus eliminating dimensional differences).

[0054] The trained random forest regression model is used to predict the influence of meteorological factors on the residuals (resid_met_pred, meteorological influence component).

[0055] The difference between the residual term (resid_component) from the STL algorithm decomposition and the meteorological impact component (resid_met_pred, meteorological impact component) predicted by the random forest model is used to obtain the meteorological correction residual (resid_corrected). The influence of meteorological factors is then removed from the residual to obtain a residual sequence containing only non-meteorological factors. The specific formula is as follows:

[0056] resid_corrected=resid_component-resid_met_pred.

[0057] S5, Trend-Residual Reconstruction.

[0058] The trend term (stl_result.trend) obtained from the STL algorithm decomposition is added to the meteorological correction residual (resid_corrected) to obtain the trend-correction residual (trend_resid_corrected), as shown in the following formula:

[0059] trend_resid_corrected=stl_result.trend+resid_corrected.

[0060] As the core target variable for ARIMAX modeling, it retains long-term trend characteristics while eliminating random fluctuations caused by weather, providing a stable input for subsequent time series modeling.

[0061] S6, Weather-Driven ARIMAX Modeling.

[0062] The auto_arima algorithm is adopted. By setting the parameter search range of max_p=7, max_d=3, and max_q=7, the space of optional parameters is expanded. With the Akaike information criterion as the optimization objective, the parameters p, d, and q corresponding to the lowest Akaike information criterion are selected as the optimal parameter combination of the ARIMAX model.

[0063] When constructing an ARIMAX model based on the optimal parameter combination, the core relationship can be expressed by the formula:

[0064] .

[0065] in, Here, φ represents the predicted value of the trend-corrected residual (forecast_trend_resid.values), φ represents the optimal parameter combination selected by the auto_arima algorithm, Y represents the trend-corrected residual (trend_resid_corrected, which is the target sequence for modeling), and X represents meteorological features. Meteorological features include both raw meteorological data such as temperature, precipitation, and wind speed, as well as derived features such as the rate of temperature change and the rate of precipitation change obtained through dynamic feature engineering. This allows the model to simultaneously capture the autocorrelation of the sequence itself (through Y) to characterize the temporal dynamics of subsidence through the above formula, and to quantify the external driving effect of meteorological factors through X, thereby achieving accurate modeling and prediction of the target sequence.

[0066] S7, Multi-scale Prediction and Reconstruction.

[0067] The future meteorological impact prediction (future_resid_met_pred) is obtained by linear extrapolating the meteorological characteristics based on the most recent 12 meteorological data.

[0068] The extended seasonal term (repeated_seasons) is obtained by copying the last complete cycle of the historical seasonal term and extending it to the forecast period.

[0069] The subsidence forecast value (forecast_full) is obtained by summing the trend-corrected residual forecast value (forecast_trend_resid.values), the extended seasonal term (repeated_seasons), and the future meteorological impact forecast value (future_resid_met_pred). The specific formula is as follows:

[0070] forecast_full=forecast_trend_resid.values+repeated_seasons+future_resid_met_pred.

[0071] By taking into account long-term trends, cyclical fluctuations, and the influence of meteorological factors, a complete deformation prediction result is obtained.

[0072] S8. Model Evaluation and Visualization.

[0073] Four bridge hotspots (points A and D are located on the bridge deck, and points B and C are located on the artificial island, representing two deformation modes respectively) were used to evaluate the deformation prediction accuracy of the traditional STL-ARIMA model and the STL-ARIMA-meteorological coupled model. The coefficient of determination R2, root mean square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE, normalized square error NSE, and residual standard deviation RSTD were calculated, as shown in Table 1.

[0074] Table 1. Comparison of indices between the traditional STL-ARIMA model and the STL-ARIMA-meteorological coupled model.

[0075]

[0076] Comparative verification shows that the STL-ARIMA-meteorological coupling model of this invention can effectively improve the deformation prediction accuracy of bridges compared with the previous STL-ARIMA model.

[0077] Plot a comparison graph of historical observations, model-fitted values, and future predictions, and mark the current time point as the boundary between history and the future. Figure 2 As shown, the deformation prediction results for the four hotspot areas of AD demonstrate that the STL-ARIMA-meteorological coupling model of this invention can better capture the characteristics of the original time series, and the fitted curve is basically consistent with the original data curve. Furthermore, based on this model, the time series deformation of bridges for the next year (March 2025 - March 2026) was predicted.

[0078] The present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

[0079] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting deformation of a cross-sea bridge based on an STL-ARIMA-meteorological coupling model, characterized in that, The method comprises the following steps: S1, data acquisition and time alignment: collect the deformation data and meteorological data of the cross-sea bridge, correlate the deformation data and meteorological data of the cross-sea bridge to the same time dimension, set the time stamp as the index, and resample the deformation data and meteorological data by setting a fixed time interval to completely align the deformation data and meteorological data in the time scale, wherein the deformation data is the cumulative settlement, and the meteorological data includes temperature, precipitation and wind speed; S2, dynamic feature engineering construction: deformation features and meteorological features are constructed based on the deformation data and meteorological data of the cross-sea bridge; S3, adaptive STL parameter optimization: the seasonal cycle and trend window of the STL algorithm are optimized based on the deformation data, and the cumulative settlement time series is decomposed into a trend item, a seasonal item and a residual item by using the STL algorithm after parameter optimization; S4, meteorological-residual coupling modeling: a random forest regression model is trained by using the residual item obtained by the STL algorithm, meteorological features and deformation features, the influence of meteorological factors on the residual is predicted by using the trained random forest regression model, and the meteorological correction residual is calculated according to the influence of meteorological factors on the residual and the residual item obtained by the STL algorithm; S5, trend-residual reconstruction: the trend item obtained by the STL algorithm is added to the meteorological correction residual to obtain a trend-correction residual; S6, ARIMAX modeling driven by meteorology: the optimal parameter combination of the ARIMAX model is screened by using the auto_arima algorithm, the trend-correction residual and meteorological features are used to fit the ARIMAX model, and the trend-correction residual is predicted by using the fitted ARIMAX model; S7, multi-scale prediction reconstruction: the future meteorological influence prediction value and the extended seasonal item are calculated, the settlement prediction value is calculated by using the trend-correction residual prediction value, the future meteorological influence prediction value and the extended seasonal item; S8, model evaluation and visualization: the settlement prediction accuracy is evaluated by using the coefficient of determination R2, the root mean square error RMSE, the mean absolute error MAE, the mean absolute percentage error MAPE, the normalized square error NSE and the residual standard deviation RSTD, and the comparison chart of historical observation value-model fitting value-future prediction value is drawn and the current time point is marked as the boundary between history and future.

2. The method of claim 1, wherein the method is based on an STL-ARIMA-weather coupled model. The step S2 comprises: The difference between the maximum temperature and the minimum temperature is calculated to obtain the temperature change rate, the difference between the current cumulative precipitation and the previous cumulative precipitation is calculated to obtain the precipitation change rate, and the meteorological features composed of the average temperature, the maximum temperature, the minimum temperature, the temperature change rate, the average wind speed, the maximum wind speed, the cumulative precipitation and the precipitation change rate are constructed; The rolling mean of the cumulative settlement of the previous 3 periods and the rolling standard deviation of the previous 6 periods are calculated, the cumulative settlement exceeding the mean value ± 3 times the standard deviation is defined as an abnormal value, the PCHIP interpolation method is used for correction, and the deformation features composed of the cumulative settlement of the previous 3 periods, the rolling mean of the previous 3 periods and the rolling standard deviation of the previous 6 periods are constructed.

3. The method of claim 2, wherein the method is based on an STL-ARIMA-weather coupled model. The step S3 comprises: The correlation of the cumulative settlement amount at different lag orders is calculated by the autocorrelation function, a significant peak is identified by combining a peak detection algorithm, the lag order corresponding to the first significant peak is added by 1 to obtain an optimal seasonal period, the optimal seasonal period is added by 1 to obtain an optimal trend window and is adjusted to be an odd number, and the cumulative settlement amount time series is decomposed into a trend item, a seasonal item and a residual item by using the STL algorithm with optimized parameters.

4. The method of claim 3, wherein the method is based on an STL-ARIMA-weather coupled model. The step S4 comprises: The residual item obtained by decomposition by using the STL algorithm is taken as a target variable, meteorological features and deformation features are taken as input features, and a random forest regression model is used to establish a nonlinear mapping relationship between the meteorological features and the residual item; An influence amount of the meteorological factors on the residual is predicted by using the trained random forest regression model to obtain a meteorological influence component; A difference between the residual item decomposed by using the STL algorithm and the meteorological influence component predicted by using the random forest model is calculated to obtain a meteorological correction residual.

5. The method of claim 4, wherein the method is based on an STL-ARIMA-weather coupled model. The step S6 comprises: An auto_arima algorithm is used, a parameter p, d, q of an ARIMAX model is set, a maximum value is expanded to expand a parameter search range to expand a selectable parameter space, and an AIC (Akaike Information Criterion) is taken as an optimization target, so that a parameter p, d, q corresponding to the lowest AIC is selected as an optimal parameter combination of the ARIMAX model; When the ARIMAX model is constructed based on the optimal parameter combination, a core relationship can be expressed by a formula as follows: ； where, is the trend-corrected residual prediction, φ is the optimal parameter combination selected by the auto arima algorithm, Y is the trend-corrected residual, and X is the meteorological feature.

6. The method of claim 5, wherein the method is based on an STL-ARIMA-weather coupled model. The step S7 comprises: Future meteorological influence prediction values are obtained by linear extrapolation of the meteorological features based on the last 12 periods of meteorological data; An extended seasonal item is obtained by copying and extending the last complete period of the historical seasonal item to the prediction period; A sum of the trend-correction residual prediction value, the extended seasonal item and the future meteorological influence prediction value is calculated to obtain a settlement prediction value.

Images