Deep learning-based permafrost upper limit prediction method and device, and terminal

Abstract

The invention provides a permafrost upper limit prediction method and device based on deep learning, and a terminal, and relates to the technical field of permafrost deformation. A historical frozen soil upper limit is decomposed into a plurality of components, a long-term characteristic prediction model is adopted to pay attention to long-term change characteristics of the frozen soil upper limit, and a short-term characteristic prediction model is adopted to pay attention to short-term fluctuation characteristics of the frozen soil upper limit. The long-term change feature and the short-term fluctuation feature of the frozen soil upper limit are respectively predicted by using the advantage of extracting the long-term trend feature by the long-term feature prediction model and the advantage of extracting the short-term fluctuation feature by the short-term feature prediction model. According to the method, the data are decomposed into different feature items, and different models are adopted for processing according to the characteristics of each feature item, so that the nonlinear relationship in the data is more flexibly fitted. And the long-term trend prediction result and the short-term fluctuation prediction result in each characteristic item are combined, so that the real rule of the data can be better captured compared with a single model, and the accuracy of the permafrost upper limit prediction result is improved.

Description

Technical Field

[0001] The present invention relates to the field of frozen soil deformation technology, and in particular to a method, device and terminal for predicting the upper limit of permafrost based on deep learning. Background Art

[0002] The upper permafrost limit refers to the location of the top surface of the permafrost layer in permafrost areas. The upper permafrost limit is an important indicator for evaluating the stability of roadbeds in permafrost areas. Predicting the upper permafrost limit with high temperature instability is crucial to ensure the safe operation and maintenance of major projects.

[0003] At present, the prediction methods mainly include empirical models, physical models and machine learning. The structure of the empirical model is too simple and the accuracy is low. The establishment of a relatively complex physical model structure requires high data. Machine learning methods rely on data for training and optimization. They are not limited by expert knowledge and prior assumptions and can process data sets containing a large number of variables and features in a variety of application scenarios. Traditional machine learning methods usually directly train a single model with raw data for prediction.

[0004] The permafrost upper limit has the characteristics of regional differences, dynamics, and sensitivity. The depth and change trend of the permafrost upper limit vary in different geographical regions, and are affected by the combined effects of multiple factors such as local climate conditions, vegetation cover, and geological conditions. It is also very sensitive to changes in external temperature and changes dynamically with changes in external temperature. This change has certain complexity and uncertainty. The changes in the permafrost upper limit are multi-scale in time and space. From seasonal changes to long-term trends, from local areas to large areas, the interweaving of multiple factors increases the difficulty of predicting the permafrost upper limit. On the one hand, the changes in the permafrost upper limit are complex, and the original data cannot effectively reflect the changing characteristics of the permafrost upper limit. On the other hand, a single model can only effectively focus on some of the factors, and cannot fully and accurately describe the complex relationship between the permafrost upper limit and various influencing factors. The ability to extract the characteristics of the permafrost upper limit is limited, and the prediction accuracy is low. Summary of the invention

[0005] The embodiments of the present invention provide a method, device and terminal for predicting the upper limit of permafrost based on deep learning, so as to solve the problem of low prediction accuracy of the existing permafrost upper limit prediction method.

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting an upper limit of permafrost based on deep learning, comprising: Obtain historical permafrost environmental data and historical permafrost upper limit data; The historical permafrost upper limit data is subjected to variational mode decomposition to obtain a long-term trend term and a short-term seasonal fluctuation term and a random fluctuation term; The historical environmental data and long-term trend items are used as training sets to perform deep learning training to obtain a trained long-term feature prediction model; The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets to perform deep learning training to obtain a trained short-term feature prediction model; Based on the trained long-term feature prediction model and short-term feature prediction model, the upper limit of permafrost is predicted respectively to obtain the upper limit prediction result of permafrost.

[0007] In a possible implementation, the historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and the trained short-term feature prediction model includes: The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and the Bayesian optimization algorithm is used to tune the hyperparameters during the training process to obtain a trained short-term feature prediction model.

[0008] In a possible implementation, the upper limit of permafrost is predicted based on the trained long-term feature prediction model and the short-term feature prediction model, respectively, and the upper limit prediction result of permafrost is obtained, including: Obtain current environmental data on permafrost; Based on the trained long-term characteristic prediction model and current and historical environmental data, predict the long-term trend item of the upper limit of permafrost; Based on the trained short-term characteristic prediction model and current and historical environmental data, the short-term seasonal fluctuation term and random fluctuation term of the upper limit of permafrost are predicted; The predicted long-term trend term, seasonal fluctuation term and random fluctuation term are combined to obtain the upper limit prediction result of permafrost.

[0009] In a possible implementation, the environmental data includes: 2-meter air temperature, near-surface temperature, surface heat flux, daily average temperature, precipitation, wind speed or air humidity.

[0010] In a possible implementation, after obtaining the historical environmental data of permafrost and the historical upper limit data of permafrost, the method further includes: Based on the 3σ principle of normal distribution, outliers in historical environmental data and historical permafrost upper limit data are identified and removed.

[0011] In a possible implementation, obtaining historical environmental data and historical upper limit data of permafrost includes: Obtain historical permafrost upper limit data and multiple influencing factors data; Obtain the lag time for ambient temperature changes to affect the upper limit of permafrost; Based on the lag time, determining the correlation between each influencing factor data and the historical permafrost upper limit data; Based on the correlation of the data of various influencing factors, the data of influencing factors with correlation greater than the preset value are screened as the historical environmental data of permafrost.

[0012] In a possible implementation, obtaining the lag time of the impact of the ambient temperature change on the upper limit of permafrost includes: Obtain the time of extreme surface temperatures of permafrost; Obtain the time of extreme ground temperature at the upper limit of permafrost; The difference between the surface temperature extreme time and the ground temperature extreme time is taken as the lag time for the impact of ambient temperature changes on the upper limit of permafrost.

[0013] In a possible implementation, the long-term feature prediction model includes a long short-term memory network; the short-term feature prediction model includes a gradient boosting decision tree.

[0014] In a second aspect, an embodiment of the present invention provides a permafrost upper limit prediction device based on deep learning, the device comprising: An acquisition module is used to obtain historical environmental data of permafrost and historical permafrost upper limit data; A decomposition module is used to perform variational mode decomposition on the historical permafrost upper limit data to obtain long-term trend terms and short-term seasonal fluctuation terms and random fluctuation terms; A first training module is used to use the historical environmental data and long-term trend items as training sets to perform deep learning training to obtain a trained long-term feature prediction model; The second training module is used to use the historical environmental data, seasonal fluctuation items and random fluctuation items as training sets to perform deep learning training to obtain a trained short-term feature prediction model; The prediction module is used to predict the upper limit of permafrost based on the trained long-term feature prediction model and short-term feature prediction model, respectively, to obtain the upper limit prediction result of permafrost.

[0015] In a possible implementation, the historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and the trained short-term feature prediction model includes: The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and the Bayesian optimization algorithm is used to tune the hyperparameters during the training process to obtain a trained short-term feature prediction model.

[0016] In the third aspect, an embodiment of the present invention provides a permafrost upper limit prediction terminal based on deep learning, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the method in the first aspect or any possible implementation method of the first aspect when executing the computer program.

[0017] The embodiment of the present invention decomposes the historical permafrost upper limit into multiple components, trains different models with different components, uses a long-term feature prediction model to focus on the long-term change characteristics of the permafrost upper limit, and uses a short-term feature prediction model to focus on the short-term fluctuation characteristics of the permafrost upper limit. Then, the advantages of the long-term feature prediction model in extracting long-term trend characteristics and the advantages of the short-term feature prediction model in extracting short-term fluctuation characteristics can be used to predict the long-term change characteristics and short-term fluctuation characteristics of the permafrost upper limit respectively. The embodiment of the present invention decomposes the data into different feature items, and uses different models for processing according to the characteristics of each feature item, so as to more flexibly fit the nonlinear relationship in the data. Then, the long-term trend and short-term fluctuation prediction results of each feature item can be combined, so as to more accurately describe the nonlinear changes of the permafrost upper limit data, better capture the real laws of the data than using a single model, and improve the accuracy of the permafrost upper limit prediction results.

[0018] In addition, decomposing characteristic items also helps the model capture the changing characteristics of data in different time or space. Separating and modeling different components such as trend changes and seasonal changes in non-stationary permafrost upper limit data can better adapt to the temporal and spatial changes of data statistical characteristics, thereby improving the stability of permafrost upper limit prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is an application scenario diagram of the permafrost upper limit prediction method based on deep learning provided by an embodiment of the present invention; Figure 2 is a flowchart of an implementation method for predicting the upper limit of permafrost based on deep learning provided by an embodiment of the present invention; Figure 3 is a schematic diagram of raw temperature data provided by an embodiment of the present invention; Figure 4 is a schematic diagram of raw humidity data provided by an embodiment of the present invention; Figure 5 is a heat map of the upper limit correlation of frozen soil on the left shoulder provided by an embodiment of the present invention; Figure 6 is a heat map of the upper limit correlation of frozen soil in road holes provided by an embodiment of the present invention; Figure 7 is a heat map of the correlation of the upper limit of frozen soil in natural holes provided by an embodiment of the present invention; Figure 8is a schematic diagram of the upper limit modal decomposition of frozen soil on the left shoulder provided by an embodiment of the present invention; Fig. 9 It is a schematic diagram of the upper limit modal decomposition of frozen soil in a road hole provided by an embodiment of the present invention; Fig.10 It is a schematic diagram of the upper modal decomposition of frozen soil in a natural site provided by an embodiment of the present invention; Fig.11 It is a block diagram of an implementation method of a permafrost upper limit prediction method based on deep learning provided by an embodiment of the present invention; Fig.12 is a schematic diagram for comparing evaluation indicators of three monitoring points provided by an embodiment of the present invention; Fig.13 : is a comparison diagram of the upper limit prediction results of frozen soil achieved by three different models for the left shoulder provided by the embodiment of the present invention; wherein, Fig.13 Figure a in the middle shows the comparison between the predicted value and the measured value of the left shoulder of the road by the XGBoost model. Fig.13 Figure b in the middle shows the comparison between the predicted value and the measured value of the left shoulder LSTM-VMD-XGBoost model. Fig.13 Figure c in the middle shows the comparison between the predicted value and the measured value of the left shoulder by the LSTM-VMD-BO-XGBoost model; Fig.14 : is a comparison chart of the upper limit prediction results of frozen soil achieved by three different models of road holes provided in an embodiment of the present invention; wherein, Fig.14 Figure a in the middle shows the comparison between the predicted value and the measured value of the XGBoost model for the hole in the road. Fig.14 Figure b shows the comparison between the predicted value and the measured value of the LSTM-VMD-XGBoost model. Fig.14 Figure c in the middle shows the comparison between the predicted value and the measured value of the LSTM-VMD-BO-XGBoost model of the hole in the road; Fig.15 : is a comparison chart of the prediction results of the upper limit of frozen soil achieved by three different models of natural sites provided by the embodiment of the present invention; wherein, Fig.15 Figure a in the middle shows the comparison between the predicted value and the measured value of the XGBoost model for natural sites. Fig.15 Figure b shows the comparison between the predicted values ​​and measured values ​​of the LSTM-VMD-XGBoost model for natural sites. Fig.15 Figure c in the middle shows the comparison between the predicted values ​​and measured values ​​of the LSTM-VMD-BO-XGBoost model for natural sites; Fig.16 It is a structural schematic diagram of a permafrost upper limit prediction device based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Since the permafrost upper limit data has nonlinear and unstable characteristics, and its fluctuation is affected by many complex factors, it is difficult to achieve high accuracy and good generalization using machine learning models for direct prediction. Especially in the prediction of the artificial upper limit of permafrost in high-temperature unstable permafrost areas, the high-temperature unstable permafrost responds quickly to seasonal changes during the transition from transitional permafrost degradation to non-transitional permafrost, and the influencing factors are complex and changeable, making prediction difficult.

[0022] Figure 1 : is an application scenario diagram of the permafrost upper limit prediction method based on deep learning provided by an embodiment of the present invention. Figure 1 As shown in the figure, above the permafrost layer is the seasonal frozen soil layer and the roadbed. When carrying out engineering construction in permafrost areas, it is necessary to fully consider the changes in the upper limit of permafrost to ensure the stability of the roadbed and the safety of the project.

[0023] The embodiment of the present invention decomposes the permafrost upper limit signal by variational mode decomposition, which is conducive to solving the problem of modal aliasing in complex permafrost upper limit data, reducing the non-stationarity of the data, and highlighting its periodic trend characteristics and local fluctuation characteristics. The trend item data changes steadily and is predicted by the long-term feature prediction model. The periodic term and residual term are predicted using the short-term feature prediction model.

[0024] Figure 2 : is a flowchart of the implementation of the permafrost upper limit prediction method based on deep learning provided by an embodiment of the present invention. Figure 2 The embodiment of the present invention provides a method for predicting the upper limit of permafrost based on deep learning, comprising: Step 201, obtaining historical environmental data of permafrost and historical upper limit data of permafrost; In some embodiments, a permafrost fault in a certain area is taken as a target permafrost fault, and historical environmental data and historical permafrost upper limit data of the target permafrost fault are obtained.

[0025] For example, the target area has an open terrain. Due to the short frost-free period and low temperature, the growth of surface vegetation is low and permafrost is widely distributed. This area is a high-temperature permafrost area. The permafrost temperature within 20 meters underground is above -1°C, which belongs to high-temperature permafrost. The altitude is 4513.3m and the roadbed direction is 261.35°. Based on the roadbed direction of the monitoring section, the left side of the roadbed is the sunny side, and the right side corresponds to the shady side. This kind of frozen soil is very sensitive to environmental changes (such as climate warming, human activities, etc.).

[0026] For example, Figure 1This is the sensor layout diagram of the monitoring section, with a roadbed height of 0.9m. Temperature observation holes are set up at the left shoulder, the middle of the road and the natural surface of the section. The left shoulder observation hole has a temperature probe set every 0.5m in the depth direction from 0.5m below the roadbed surface, and the measurement depth is 20m; the middle observation hole has a temperature probe set every 0.5m in the depth direction from 0.5m below the roadbed surface, and a temperature probe set every 1m after exceeding 10m, and the measurement depth is 20m; the natural surface observation hole has a temperature probe set every 0.5m in the depth direction from 0.3m below the surface, and a temperature probe set every 1m after exceeding 10m, and the measurement depth is 16.3m. In addition, meteorological data of the target area and other factors that may affect the surface temperature and the upper limit of frozen soil changes were obtained. The data sampling interval is 2 times / month, and the monitoring data from January 2003 to December 2021 are used for analysis. The sorted data are used to build a permafrost upper limit prediction model, of which the first 90% of the data are used for model training and 10% of the data are used for model testing. It should be noted that the three positions of the left shoulder, the middle of the road and the natural surface can be used as processing objects of the embodiment method of the present invention respectively. For example, the environmental data and permafrost upper limit data collected at the left shoulder position can predict the permafrost upper limit at the left shoulder position.

[0027] Collecting appropriate data sets is the basis for successful prediction of machine learning. For example, 402 sets of field monitoring data were extracted from 2004 to 2021, with a total of more than 2,800 data points. For example, the data set may include: daily average temperature, rainfall, wind speed, air humidity, road surface temperature, road surface heat flow, road frozen soil upper limit, natural surface temperature, natural surface heat flow, natural surface frozen soil upper limit, left shoulder surface temperature, left shoulder surface heat flow, left shoulder frozen soil upper limit and other data.

[0028] In a possible implementation, the environmental data includes: 2-meter air temperature, near-surface temperature, surface heat flux, daily average temperature, precipitation, wind speed or air humidity.

[0029] Accurate training data is the key to successful model fitting. In the actual data collection process, problems such as data transmission errors, sensor failures, and changes in natural factors may introduce abnormal data. These noisy data may cause the model to over-focus on details and noise during training, resulting in poor performance on new data and lack of good generalization ability. Taking the original data of temperature and humidity at 2 meters above the ground of the target area weather station as an example, a detailed discussion and analysis is conducted.

[0030] In a possible implementation, after obtaining the historical environmental data and the historical permafrost upper limit data of permafrost, the method further includes: identifying and removing outliers in the historical environmental data and the historical permafrost upper limit data based on the 3σ principle of normal distribution.

[0031] Figure 3 is a schematic diagram of raw temperature data provided by an embodiment of the present invention; Figure 4 is a schematic diagram of the original humidity data provided by an embodiment of the present invention; Figure 3 , Figure 4 , the temperature changes periodically as a whole, and the humidity fluctuates irregularly, but is basically maintained within a certain range. In addition, the data occasionally has abnormal values ​​that are out of range. The principle is used to filter outliers and process them. Principle, data Should be at the mean Data that exceeds the mean plus or minus 3 standard deviations is considered an outlier. In order to ensure the integrity and periodicity of the data, this method helps to identify and handle these anomalies. For example, based on the linear relationship between adjacent data, a linear interpolation method is used to smooth out the outliers.

[0032] In one possible implementation, the acquisition of historical environmental data and historical permafrost upper limit data includes: acquiring historical permafrost upper limit data and multiple influencing factor data; acquiring the lag time for the impact of environmental temperature changes on the permafrost upper limit; based on the lag time, determining the correlation between each influencing factor data and the historical permafrost upper limit data; based on the correlation of each influencing factor data, screening the influencing factor data whose correlation is greater than a preset value as the historical environmental data of permafrost.

[0033] In some embodiments, obtaining the lag time for the impact of environmental temperature changes on the upper limit of permafrost includes: obtaining the surface temperature extreme time of permafrost; obtaining the ground temperature extreme time at the upper limit position of permafrost; and using the difference between the surface temperature extreme time and the ground temperature extreme time as the lag time for the impact of environmental temperature changes on the upper limit of permafrost.

[0034] It should be noted that for the correlation analysis between multiple influencing factors and target variables, the hysteresis effect of temperature change on the upper limit of frozen soil is considered here, and the hysteresis time is determined based on the time difference between the extreme surface temperature and the extreme ground temperature near the upper limit of frozen soil. For example, the left shoulder lags about 4.5 months, the middle of the road lags about 5 months, and the natural site lags about 2.5 months. Therefore, the correlation between the influencing factors and the upper limit of frozen soil can be calculated by considering the hysteresis effect within a freeze-thaw cycle.

[0035] Figures 5 to 7 It is a heat map of the correlation between various influencing factors and the upper limit of frozen ground at different locations. Figure 5 is a heat map of the upper limit correlation of frozen soil on the left shoulder provided by an embodiment of the present invention; Figure 6is a heat map of the upper limit correlation of frozen soil in road holes provided by an embodiment of the present invention; Figure 7 is a heat map of the correlation of the upper limit of natural hole frozen soil provided by the embodiment of the present invention; Figure 5 , Figure 6 , Figure 7 For the permafrost upper limit, air humidity, precipitation, near-surface temperature, and surface heat flux density are highly correlated. Air humidity shows a weak positive correlation with the upper limit change, while wind speed shows a weak negative correlation with the permafrost upper limit. Therefore, the correlation between each influencing factor and the permafrost upper limit may be different in different locations. Therefore, the embodiment of the present invention can further improve the accuracy of the original data and improve the accuracy of the permafrost upper limit prediction by performing correlation judgment on each influencing factor of the target object.

[0036] Step 202, performing variational mode decomposition on the historical permafrost upper limit data to obtain a long-term trend term and a short-term seasonal fluctuation term and a random fluctuation term; It should be noted that before performing variational modal decomposition (VMD) on the permafrost upper limit change data, it is necessary to determine the number of modal components k. First, starting from the preliminary estimated k value, VMD decomposition is performed, the frequency and energy distribution of each modal component is observed, and the reconstruction error is calculated. By comparing the decomposition effects under different k values, the k value is gradually adjusted according to the change of the reconstruction error and the characteristics of each modal component until the optimal decomposition result is obtained. After many experiments and adjustments, the k value was finally determined to be 3, at which time the decomposition result is optimal. This optimal k value not only minimizes the reconstruction error, but also the frequency and energy distribution of each modal component is relatively reasonable, successfully capturing the main characteristics and periodic change laws of permafrost upper limit changes. This provides a reliable basis for the subsequent prediction of permafrost upper limit changes. Modal decomposition of permafrost upper limit data at different monitoring points is shown in the figure. Figures 8 to 10 shown. Figure 8 is a schematic diagram of the upper limit modal decomposition of frozen soil on the left shoulder provided by an embodiment of the present invention; Fig. 9 It is a schematic diagram of the upper limit modal decomposition of frozen soil in a road hole provided by an embodiment of the present invention; Fig.10 It is a schematic diagram of the upper modal decomposition of frozen soil in a natural site provided by an embodiment of the present invention.

[0037] It can be seen from the permafrost upper limit modal decomposition diagram that IMF1 is a trend term with a relatively gentle change, representing the long-term change trend of the permafrost upper limit; while IMF2 is a periodic term, showing the seasonal fluctuation law of the permafrost upper limit with strongly correlated periodic influencing factors, and the value of IMF2 is usually relatively large, and its change has a greater impact on the permafrost upper limit. IMF1 and IMF2 directly reflect the overall change trend of the permafrost upper limit to a large extent. IMF3 is an error term, or residual term, with a relatively small amplitude change range and random fluctuations, but a large fluctuation frequency, which is mainly caused by a variety of other weakly correlated influencing factors. Its change has little overall impact on the permafrost upper limit. VMD decomposes the original signal of the permafrost upper limit change into global trend and local fluctuation characteristics, which helps to further predict the change of the permafrost upper limit.

[0038] The above steps obtain the training data. The following steps illustrate the model training set process.

[0039] In the embodiment of the present invention, the data of the upper limit change of frozen soil is decomposed into long-term trend items, seasonal fluctuation items and random fluctuation items. The long-term trend item is relatively stable, and the long-term feature prediction model is used for training and prediction, while the seasonal fluctuation item and the random fluctuation item have large volatility, and the short-term feature prediction model is used for training and prediction.

[0040] Step 203, using the historical environmental data and long-term trend items as training sets, performing deep learning training to obtain a trained long-term feature prediction model; Exemplarily, the long-term feature prediction model includes a long short-term memory network.

[0041] It should be noted that the long-term trend item here is the long-term trend item of the historical frozen soil upper limit data obtained in step 202, referred to as the frozen soil upper limit long-term trend item. For permafrost, changes in historical environmental data lead to changes in the frozen soil upper limit long-term trend item.

[0042] In some embodiments, historical environmental data and long-term trend items of the upper limit of permafrost are used as training sets, and deep learning training is performed on a pre-constructed long short-term memory network to obtain a trained long short-term memory network.

[0043] Step 204, using the historical environmental data, seasonal fluctuation items and random fluctuation items as training sets, performing deep learning training to obtain a trained short-term feature prediction model; Exemplarily, the short-term feature prediction model includes a gradient boosting decision tree.

[0044] In some embodiments, the historical environmental data, the seasonal fluctuation term of the upper limit of permafrost, and the random fluctuation term of the upper limit of permafrost are used as training sets, and deep learning training is performed on a pre-constructed gradient boosting decision tree to obtain a trained gradient boosting decision tree.

[0045] It should be noted that the seasonal fluctuation term reflects the regular changes in the upper limit of frozen ground with the change of seasons. For example, in warm summers, the upper limit of frozen ground may drop (increase in depth); in cold winters, the upper limit of frozen ground may rise (become shallower). This seasonal fluctuation term helps the gradient boosting decision tree model capture the changing pattern of the upper limit of frozen ground with the seasonal cycle.

[0046] In addition to the regular seasonal changes, the permafrost limit is also affected by some random factors, such as sudden extreme climate events and local geological activities, which lead to irregular fluctuations. The random fluctuation term contains the change information caused by these unpredictable factors, enabling the gradient boosting decision tree model to learn more complex permafrost limit change characteristics.

[0047] Gradient boosting decision tree is an ensemble learning method consisting of multiple decision trees. A decision tree is a model that makes decisions based on a tree structure. It builds a tree structure by partitioning features. Each internal node is a test on a feature, each branch is a test output, and each leaf node is a category or value. Gradient boosting is an iterative decision tree training method. Each round of training will train a new decision tree to fit the residual of the previous round of model. Through continuous iteration, the performance of the model is gradually improved.

[0048] Before training, the basic structure and parameters of the gradient boosting decision tree have been preliminarily set, such as the number of decision trees, tree depth, learning rate, etc., to determine the basic framework of the model so that the patterns and regularities of the data can be learned in the training set later.

[0049] In one possible implementation, the historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training to obtain a trained short-term feature prediction model, including: using the historical environmental data, seasonal fluctuation items and random fluctuation items as training sets for deep learning training, and using a Bayesian optimization algorithm to tune hyperparameters during the training process to obtain a trained short-term feature prediction model.

[0050] Exemplarily, the historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and during the training process, the Bayesian optimization algorithm is used to tune the hyperparameters to obtain a trained gradient boosting decision tree.

[0051] In order to improve the prediction accuracy, the embodiment of the present invention uses the Bayesian optimization algorithm to tune the hyperparameters during the prediction process. Exemplarily, the hyperparameters include learning rate, maximum depth, weight, number of trees, subsampling rate, and random sampling ratio. For example, the adjusted hyperparameter configuration is shown in Table 1 Hyperparameter Configuration Table: Table 1 Hyperparameter configuration table

[0052] Step 205, based on the trained long-term feature prediction model and the short-term feature prediction model, the upper limit of permafrost is predicted respectively to obtain a permafrost upper limit prediction result.

[0053] It should be noted that two different predictions about the upper limit of permafrost are obtained by using two different models. The two results differ in their emphasis on different influencing factors. For example, the long short-term memory network is better at capturing the long-term continuous change trend of the upper limit of permafrost, while the gradient boosting decision tree is more sensitive to short-term changes affected by multiple complex factors. Therefore, the two prediction results are comprehensively analyzed and compared to gain a more comprehensive understanding of the possible changes in the upper limit of permafrost.

[0054] In a possible implementation, the upper limit of permafrost is predicted based on the trained long-term feature prediction model and the short-term feature prediction model, respectively, and the upper limit prediction result of permafrost is obtained, including: Obtain current environmental data on permafrost; Based on the trained long-term characteristic prediction model and current and historical environmental data, predict the long-term trend item of the upper limit of permafrost; Based on the trained short-term characteristic prediction model and current and historical environmental data, the short-term seasonal fluctuation term and random fluctuation term of the upper limit of permafrost are predicted; The predicted long-term trend term, seasonal fluctuation term and random fluctuation term are combined to obtain the upper limit prediction result of permafrost.

[0055] Exemplarily, the predicted long-term trend term, seasonal fluctuation term and random fluctuation term are added and combined to obtain the upper limit prediction result of permafrost.

[0056] In another exemplary embodiment, the predicted long-term trend term, seasonal fluctuation term and random fluctuation term are weightedly added together to obtain the upper limit prediction result of permafrost.

[0057] The embodiment of the present invention decomposes the historical permafrost upper limit into multiple components, trains different models with different components, uses a long-term feature prediction model to focus on the long-term change characteristics of the permafrost upper limit, and uses a short-term feature prediction model to focus on the short-term fluctuation characteristics of the permafrost upper limit. Then, the advantages of the long-term feature prediction model in extracting long-term trend characteristics and the advantages of the short-term feature prediction model in extracting short-term fluctuation characteristics can be used to predict the long-term change characteristics and short-term fluctuation characteristics of the permafrost upper limit respectively. The embodiment of the present invention decomposes the data into different feature items, and uses different models for processing according to the characteristics of each feature item, so as to more flexibly fit the nonlinear relationship in the data. Then, the long-term trend and short-term fluctuation prediction results of each feature item can be combined, so as to more accurately describe the nonlinear changes of the permafrost upper limit data, better capture the real laws of the data than using a single model, and improve the accuracy of the permafrost upper limit prediction results.

[0058] In addition, decomposing characteristic items also helps the model capture the changing characteristics of data in different time or space. Separating and modeling different components such as trend changes and seasonal changes in non-stationary permafrost upper limit data can better adapt to the temporal and spatial changes of data statistical characteristics, thereby improving the stability of permafrost upper limit prediction.

[0059] The present invention is described below with a comprehensive embodiment. Fig.11 This is a block diagram of the implementation method of the permafrost upper limit prediction method based on deep learning provided by an embodiment of the present invention. The embodiment of the present invention uses variational mode decomposition (VMD) to perform multi-scale decomposition of permafrost upper limit data, effectively extracts intrinsic mode functions (IMFs) in the data, and decomposes permafrost upper limit data into trend terms, periodic terms, and residual terms. The trend item data changes smoothly and is predicted by a long short-term memory neural network (LSTM). The periodic term and residual term are predicted using the XGBoost model, and the XGBoost model hyperparameters are optimized by the Bayesian optimization algorithm (BO) to improve the accuracy of its prediction.

[0060] To verify the effectiveness of the combined model, XGBoost, LSTM-VMD-XGBoost, and LSTM-VMD-BO-XGBoost were used to predict the upper limit of frozen soil on the left shoulder, middle of the road, and natural site of the monitoring section and compared. The comparison results are shown in Table 2. XGBoost means that the XGBoost model is used alone for prediction. LSTM-VMD-XGBoost means that the LSTM-XGBoost combined model trained after variational mode decomposition is used for prediction. BO-XGBoost in LSTM-VMD-BO-XGBoost means that the XGBoost model has been tuned for hyperparameters using the Bayesian optimization algorithm.

[0061] Table 2 Performance comparison of different models

[0062] Table 2 shows the impact of each component of the proposed algorithm on the overall combined prediction model. For the same detection hole data, the LSTM-VMD-XGBoost prediction model performs better than the XGBoost model, and the LSTM-VMD-BO-XGBoost model performs better than the LSTM-VMD-XGBoost model, indicating that the use of variational mode decomposition and Bayesian hyperparameter optimization can improve the model prediction performance and computational efficiency to varying degrees.

[0063] For example, the evaluation indicators of three monitoring points are analyzed respectively, such as Fig.12 As shown. 1#, 4#, 7# are the prediction results of the XGBoost single model, 2#, 5#, 8# are the prediction results of the LSTM-VMD-BO-XGBoost model, and 3#, 6#, 9# are the prediction results of the LSTM-VMD-BO-XGBoost model. It can be seen that with the increase in the complexity of the combined model, the prediction performance of the model improves and the accuracy continues to improve. In the process of using different data for prediction, it can be seen that the distribution of the original data and the data quality also have a certain impact on the prediction results. Overall, the combined model significantly improves the accuracy relative to the single XGBoost prediction model. This shows that the combined model is of great significance in improving the accuracy and reliability of roadbed stability prediction in frozen soil areas, and helps to more effectively ensure the safety and durability of road infrastructure.

[0064] In order to more intuitively display the prediction results of the upper limit of permafrost, the single-value prediction errors of the single XGBoost model and those after adding VMD and BO algorithms respectively are compared. Fig.13 : is a comparison diagram of the upper limit prediction results of frozen soil achieved by three different models for the left shoulder provided by the embodiment of the present invention; wherein, Fig.13 Figure a in the middle shows the comparison between the predicted value and the measured value of the left shoulder of the road by the XGBoost model. Fig.13 Figure b in the middle shows the comparison between the predicted value and the measured value of the left shoulder LSTM-VMD-XGBoost model. Fig.13 Figure c in the middle shows the comparison between the predicted value and the measured value of the left shoulder LSTM-VMD-BO-XGBoost model. Fig.14 : is a comparison chart of the upper limit prediction results of frozen soil achieved by three different models of road holes provided in an embodiment of the present invention; wherein, Fig.14 Figure a in the middle shows the comparison between the predicted value and the measured value of the XGBoost model for the hole in the road. Fig.14 Figure b shows the comparison between the predicted value and the measured value of the LSTM-VMD-XGBoost model. Fig.14Figure c in the middle shows the comparison between the predicted value and the measured value of the LSTM-VMD-BO-XGBoost model for the hole in the road. Fig.15 : is a comparison chart of the prediction results of the upper limit of frozen soil achieved by three different models of natural sites provided by the embodiment of the present invention; wherein, Fig.15 Figure a in the middle shows the comparison between the predicted value and the measured value of the XGBoost model for natural sites. Fig.15 Figure b shows the comparison between the predicted values ​​and measured values ​​of the LSTM-VMD-XGBoost model for natural sites. Fig.15 Figure c shows the comparison between the predicted values ​​and measured values ​​of the LSTM-VMD-BO-XGBoost model for natural sites. Figure 13-Figure 15 , the prediction fluctuations of XGBoost, LSTM-VMD-XGBoost, and LSTM-VMD-BO-XGBoost models gradually decreased, the model stability increased, and the prediction error of LSTM-VMD-BO-XGBoost showed a significant downward trend compared with XGBoost. The prediction errors of the LSTM-VMD-BO-XGBoost model were smaller than those of the XGBoost and LSTM-VMD-XGBoost prediction models, showing good prediction performance. The combined model was verified at three different monitoring points and had good generalization.

[0065] In summary, the embodiment of the present invention combines different machine learning models to predict the upper limit of permafrost, and mainly applies the XGBoost and LSTM prediction models to perform division of labor prediction. First, the permafrost upper limit data is decomposed into trend terms, periodic terms, and residual terms using variational mode decomposition (VMD). The trend term is predicted by the LSTM model, and the periodic term and residual term are predicted by the XGBoost model. The prediction results are then reconstructed to obtain the final prediction results, and good performance has been achieved in the prediction of the upper limit of permafrost on the left shoulder, the middle of the road, and natural sites. By decomposing the complex permafrost upper limit data into different feature terms, the model effectively solves the nonlinear and non-stationary problems of the data.

[0066] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0067] The following is an embodiment of the device of the present invention. For details not described in detail therein, reference may be made to the corresponding method embodiment described above.

[0068] Fig.16 The schematic diagram of the structure of the permafrost upper limit prediction device based on deep learning provided by an embodiment of the present invention is shown. For the convenience of explanation, only the part related to the embodiment of the present invention is shown, which is described in detail as follows: like Fig.16 As shown, the permafrost upper limit prediction device 16 based on deep learning includes: An acquisition module 161 is used to acquire historical environmental data of permafrost and historical upper limit data of permafrost; A decomposition module 162 is used to perform variational mode decomposition on the historical frozen soil upper limit data to obtain long-term trend terms and short-term seasonal fluctuation terms and random fluctuation terms; A first training module 163 is used to use the historical environmental data and the long-term trend items as a training set to perform deep learning training to obtain a trained long-term feature prediction model; A second training module 164 is used to use the historical environmental data, seasonal fluctuation items and random fluctuation items as training sets to perform deep learning training to obtain a trained short-term feature prediction model; The prediction module 165 is used to predict the upper limit of permafrost based on the trained long-term feature prediction model and the short-term feature prediction model, respectively, to obtain a permafrost upper limit prediction result.

[0069] An embodiment of the present invention also provides a permafrost upper limit prediction terminal based on deep learning, including a memory and a processor, the memory stores a computer program, and the processor implements the method in the above method embodiment when executing the computer program.

[0070] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. If there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced to each other. The technical features in different embodiments can be combined to form a new embodiment according to their internal logical relationship.

[0071] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A permafrost upper limit prediction method based on deep learning, characterized in that: include: Obtain historical permafrost environmental data and historical permafrost upper limit data; The historical permafrost upper limit data is subjected to variational mode decomposition to obtain a long-term trend term and a short-term seasonal fluctuation term and a random fluctuation term; The historical environmental data and long-term trend items are used as training sets to perform deep learning training to obtain a trained long-term feature prediction model; The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets to perform deep learning training to obtain a trained short-term feature prediction model; Based on the trained long-term feature prediction model and short-term feature prediction model, the upper limit of permafrost is predicted respectively to obtain the upper limit prediction result of permafrost.

2. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets to perform deep learning training, and the trained short-term feature prediction model includes: The historical environmental data, seasonal fluctuation items and random fluctuation items are used as training sets for deep learning training, and the Bayesian optimization algorithm is used to tune the hyperparameters during the training process to obtain a trained short-term feature prediction model.

3. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: The permafrost upper limit is predicted based on the trained long-term feature prediction model and the short-term feature prediction model, respectively, and the permafrost upper limit prediction results obtained include: Obtain current environmental data on permafrost; Based on the trained long-term characteristic prediction model and current and historical environmental data, predict the long-term trend item of the upper limit of permafrost; Based on the trained short-term characteristic prediction model and current and historical environmental data, the short-term seasonal fluctuation term and random fluctuation term of the upper limit of permafrost are predicted; The predicted long-term trend term, seasonal fluctuation term and random fluctuation term are combined to obtain the upper limit prediction result of permafrost.

4. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: The environmental data include: 2-meter air temperature, near-surface temperature, surface heat flux, daily average temperature, precipitation, wind speed or air humidity.

5. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: After obtaining the historical environmental data of permafrost and the historical upper limit data of permafrost, the method further includes: Based on the 3σ principle of normal distribution, outliers in historical environmental data and historical permafrost upper limit data are identified and removed.

6. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: The acquisition of historical environmental data of permafrost and historical upper limit data of permafrost includes: Obtain historical permafrost upper limit data and multiple influencing factors data; Obtain the lag time for ambient temperature changes to affect the upper limit of permafrost; Based on the lag time, determining the correlation between each influencing factor data and the historical permafrost upper limit data; Based on the correlation of the data of various influencing factors, the data of influencing factors with correlation greater than the preset value are screened as the historical environmental data of permafrost.

7. The method for predicting the upper limit of permafrost based on deep learning according to claim 6, characterized in that: The lag time for obtaining the impact of ambient temperature changes on the upper limit of permafrost includes: Obtain the time of extreme surface temperatures of permafrost; Obtain the time of extreme ground temperature at the upper limit of permafrost; The difference between the surface temperature extreme time and the ground temperature extreme time is taken as the lag time for the impact of ambient temperature changes on the upper limit of permafrost.

8. The method for predicting the upper limit of permafrost based on deep learning according to claim 1, characterized in that: The long-term feature prediction model includes a long short-term memory network; the short-term feature prediction model includes a gradient boosting decision tree.

9. A permafrost upper limit prediction device based on deep learning, characterized in that: The device comprises: An acquisition module is used to obtain historical environmental data of permafrost and historical permafrost upper limit data; A decomposition module is used to perform variational mode decomposition on the historical permafrost upper limit data to obtain long-term trend terms and short-term seasonal fluctuation terms and random fluctuation terms; A first training module is used to use the historical environmental data and long-term trend items as training sets to perform deep learning training to obtain a trained long-term feature prediction model; The second training module is used to use the historical environmental data, seasonal fluctuation items and random fluctuation items as training sets to perform deep learning training to obtain a trained short-term feature prediction model; The prediction module is used to predict the upper limit of permafrost based on the trained long-term feature prediction model and short-term feature prediction model, respectively, to obtain the upper limit prediction result of permafrost.

10. A permafrost upper limit prediction terminal based on deep learning, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method for predicting the upper limit of permafrost based on deep learning as described in any one of claims 1 to 8 is implemented.

Images