A Real-time Prediction Method for Dynamic Landslide Displacement Based on PFA-GRU-OL-MRT

Abstract

The present invention provides a method for real-time prediction of dynamic landslide displacement based on PFA-GRU-OL-MRT, belonging to the field of landslide prediction. The content includes: obtaining the online monitored landslide displacement data and preprocessing the data; optimizing the hyperparameters of the gated recurrent unit (GRU) based on the particle fireworks algorithm (PFA) to establish an initial PFA-GRU submodel; secondly, every time the latest monitored data is obtained, based on the online learning (OL) strategy, the input submodel is trained in real time and the next-round displacement is predicted; finally, after obtaining the next monitoring data, the prediction accuracy is evaluated by combining the online learning prediction results, and it is judged whether to use the online learning submodel or the PFA-GRU dynamic reconstruction (MRT) submodel as the submodel for the next-round prediction, so as to realize the dynamic real-time prediction of landslide displacement. The present invention can provide a basis for accurately carrying out the real-time prediction of dynamic landslide displacement and landslide prevention and control.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide prediction, and in particular, to a real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT. Background Art

[0002] Landslide geological disasters will cause significant casualties and property losses, and landslide displacement is the most important basis for preventing and controlling landslide geological disasters. Therefore, predicting landslide displacement is of great practical significance for reducing casualties and economic losses.

[0003] Existing landslide displacement predictions mainly focus on the prediction of long-term landslide displacement trends, and there are relatively few predictions for short-term, especially real-time landslide displacement trends. There are the following difficulties in real-time landslide displacement trend prediction: on the one hand, the real-time dynamic characteristics of landslides determine that static models have insufficient accuracy in predicting landslide displacement, especially when the landslide trend changes, they cannot effectively predict. Therefore, a dynamic model needs to be established to improve the accuracy of real-time prediction; on the other hand, dynamic models greatly increase the training cost of the model. Especially when using optimization algorithms to optimize model hyperparameters, a large number of evaluations are required to find the optimal solution, which further increases the training cost of the model and reduces the timeliness of prediction. Therefore, how to establish a dynamic real-time prediction model suitable for landslide trend changes and maximize the reduction of calculation costs while improving accuracy is crucial for high-precision and high-efficiency real-time prediction of landslide displacement. Summary of the Invention

[0004] The purpose of the present invention is to provide a real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT, which can solve the problem that traditional static prediction models are difficult to accurately predict dynamic landslide trends, while dynamic prediction models often increase the calculation cost, bringing difficulties to real-time prediction, and provide reliable technical support for improving the accuracy and efficiency of landslide displacement prediction.

[0005] The present invention is implemented as follows:

[0006] The technical solution to achieve the purpose of the present invention is: a real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT, characterized by including the following steps:

[0007] Step 1: Obtain historical landslide displacement data obtained from online monitoring and preprocess the data;

[0008] Step 2: Set the hyperparameters required for the PFA-GRU-OL-MRT model;

[0009] The PFA-GRU-OL-MRT model includes three sub-models, namely the PFA-GRU initial sub-model, the online learning sub-model, and the PFA-GRU dynamic reconstruction sub-model;

[0010] Step 3: Based on the particle firework algorithm and historical landslide displacement data, adopt a sliding window transfer training method to optimize the hyperparameters of the gated recurrent unit;

[0011] Step 4: Based on the optimized hyperparameters, establish a PFA-GRU initial sub-model in combination with historical landslide displacement data;

[0012] Step 5: Each time the latest monitoring data D t is obtained, the input sub-model is trained in real time based on the online learning strategy to obtain an online learning sub-model and predict the landslide displacement DP at time t + 1 t+1 , and the prediction result DP is output t+1 ;

[0013] Step 6: After obtaining the next monitoring data D t+1 , combine it with DP t+1 to evaluate the accuracy of the prediction result, and then assign D t+1 to D t , that is, D t = D t+1 . If the prediction deviation is less than the allowable value, the online learning sub-model is used as the input sub-model and returned to Step 5 for the next round of prediction; if the prediction deviation is greater than the allowable value, the sliding window transfer training method is adopted, and the PFA-GRU dynamic model is reconstructed. The PFA-GRU dynamic reconstruction sub-model is used as the input sub-model and returned to Step 5 for the next round of prediction.

[0014] Furthermore, in Step 2, the hyperparameters required for the PFA-GRU-OL-MRT model include the window size W1 of the PFA-GRU initial sub-model, the window size W2 of the online learning sub-model, the window reduction scale W minus of the PFA-GRU dynamic reconstruction sub-model, the minimum window size W min of the PFA-GRU dynamic reconstruction sub-model, the training iteration times initial_epochs of the PFA-GRU initial sub-model, the training iteration times OL_epochs of the online learning sub-model, the training iteration times MRT_epochs of the PFA-GRU dynamic reconstruction sub-model, the sliding window transfer training times initial_num of the PFA-GRU initial sub-model, the sliding window transfer training times MRT_num of the PFA-GRU dynamic reconstruction sub-model, the total number of PFA particles swarmsize, the PFA iteration times maxiter, the allowable value Re of the prediction deviation, the upper limit ub and the lower limit lb of the hyperparameter optimization range.

[0015] Further, in step three, the particle firework algorithm is an improved algorithm of the particle swarm optimization algorithm, and its steps are as follows:

[0016] (I) Initialization

[0017] In the particle firework algorithm, the initial particles are generated by random initialization:

[0018] X i 0 = lb + (ub - lb) × rand(0, 1)

[0019] where i is the particle number, X i represents the position of the i-th particle, the superscript represents the iteration number, where 0 represents the initialization stage, lb is the lower limit of the optimization range, ub is the upper limit of the optimization range, rand(0, 1) represents an M-dimensional vector composed of random numbers between 0 and 1, and M represents the dimension of the problem to be solved; in the initialization stage, the particle firework algorithm only randomly generates 80% of the total number of particles, and the remaining 20% of the particles are generated in the local intensification search stage;

[0020] Evaluate the fitness of all initialized particles, and define the particle with the best fitness as the firework particle;

[0021] (II) Local intensification search

[0022] Centered on the firework particle, randomly generate the remaining 20% of the particles within a certain range around it, and call these individuals spark particles; the total number of spark particles N k and the position G bj of each spark particle are expressed as:

[0023] N k = 0.2 × N

[0024] G bj = G b + 0.2 × (ub - lb) × e -2t / T × (2 × rand(0, 1) - 1)

[0025] where N is the total number of particles, j is the number of the spark particle, G b is the position of the firework particle, t represents the current iteration number, and T represents the maximum iteration number;

[0026] Subsequently, calculate the fitness values of all spark particles;

[0027] (III) Particle update

[0028] In the particle update stage, different treatments are carried out in three cases:

[0029] (1) For particles other than fireworks particles and spark particles, their update process is as follows:

[0030] V i t+1 = (1 - t / T)V i t / 2 + c1r1(P i t - X i t ) + c2r2(G b t - X i t )

[0031] X i t+1 = X i t + V i t+1

[0032] where V represents velocity, c1 and c2 are learning factors; r1 and r2 are random numbers between 0 and 1, and P i represents the optimal position in the search history of particle i.

[0033] (2) For fireworks particles, in each iteration, they are directly updated to the position of the particle with the best fitness among all fireworks particles and spark particles, and the velocity is defined as 0;

[0034] (3) Delete all spark particles;

[0035] (4) Elimination mechanism

[0036] Recalculate the fitness values of all particles, redefine the particle with the best fitness among all particles as the fireworks particle, update the historical optimal position P i of particle i, and at the same time delete the particle with the worst fitness and randomly generate a new particle within the solution domain;

[0037] For each iteration, steps (2) to (4) are repeated until the convergence condition is met or the maximum number of iterations is reached, and then the position of the optimal solution is output.

[0038] In this way, not only is the global search ability and the ability to jump out of local optima of the PFA algorithm guaranteed, but also the search attention is more focused on the vicinity of the current globally optimal with the best fitness, thus increasing the local fast convergence ability of the algorithm.

[0039] Further, in step three, the sliding window transfer training method refers to using a sliding window for model training. The window means that the scale of the training data selected from the time series data is fixed, and all the data within the window is used as the input, and the next data outside the window is used as the output for training. Using a sliding window for model training means that after each window is trained, the window slides along the time series towards the direction of the latest monitored data. The sliding will delete the earliest monitored data in the window and add the next data after the window for retraining.

[0040] Under the sliding window transfer training method, the MSE ave is used as the fitness function when the particle firework algorithm optimizes the hyperparameters of the gated recurrent unit. The calculation method of MSE ave is as follows: ① Calculate the mean square error MSE between the predicted value and the corresponding measured value obtained at each sliding window position. ② Obtain the average value of MSE at all sliding window positions, which is the MSE. ave .

[0041] Further, in step three, the hyperparameters include the number of units in the gated recurrent unit model, the learning rate, the Dropout rate, and the L2 regularization strength.

[0042] Further, in step five, the online learning strategy means that every time a latest monitored data D t is obtained, training data is acquired with the online learning window size W2, that is, using the W2 data before D t as the training input and D t as the training output to further train the PFA-GRU model for parameter fine-tuning, where W2 < the initial sub-model window size W1 of the PFA-GRU.

[0043] Further, in step six, for the dynamic model reconstruction, the model reconstruction window size W3 is dynamic. In the initial stage of model reconstruction, W3 = the online learning window size W2. If the prediction deviation after model reconstruction is still greater than the allowable value, then each time W minus earlier monitored data is deleted to reduce the window size, and then the model is reconstructed continuously until the minimum window size W min is reached, at which point the model reconstruction stops, and the sub-model with the highest prediction accuracy is output.

[0044] The beneficial effects of the present invention are as follows: The present invention provides a real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT. Among them, the Particle Fireworks Algorithm (PFA) can enhance the local search ability and accelerate the convergence speed; establishing a PFA-GRU landslide prediction model through historical landslide displacement data can fully exploit the historical landslide displacement information; through online learning (OL), the PFA-GRU model can be fine-tuned with a small computational cost and the latest landslide displacement trend can be obtained; through model reconstruction (MRT), the latest displacement trend can be accurately exploited when the displacement trend changes significantly, so that the model has better prediction accuracy and shorter prediction time, providing reliable technical support for real-time prediction of landslide displacement. Description of the Drawings

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0046] Figure 1 It is a flowchart of a real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT provided by an embodiment of the present invention;

[0047] Figure 2 It is a flowchart of optimizing GRU hyperparameters based on PFA provided by an embodiment of the present invention;

[0048] Figure 3 It is a comparison chart of the prediction results of the PFA-GRU and GRU models provided by an embodiment of the present invention;

[0049] Figure 4 It is a comparison chart of the prediction results of the PFA-GRU-OL and GRU-OL models provided by an embodiment of the present invention;

[0050] Figure 5 It is a comparison chart of the prediction results of the PFA-GRU-OL-MRT and GRU-OL-MRT models provided by an embodiment of the present invention. Detailed Embodiments

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the implementation cases and drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0052] The following takes a specific case of real-time prediction of landslide displacement as an example to illustrate the method of the present invention.

[0053] Such as Figure 1 , a method for real-time prediction of dynamic landslide displacement based on PFA-GRU-OL-MRT, characterized by comprising the following steps:

[0054] Step 1: Obtain historical landslide displacement data obtained from online monitoring and preprocess the data;

[0055] Step 2: Set the hyperparameters required for the PFA-GRU-OL-MRT model;

[0056] The PFA-GRU-OL-MRT model includes three sub-models, namely the PFA-GRU initial sub-model, the online learning sub-model, and the PFA-GRU dynamic reconstruction sub-model;

[0057] Step 3: Based on the particle fireworks algorithm and historical landslide displacement data, adopt a sliding window transfer training method to optimize the hyperparameters of the gated recurrent unit;

[0058] Step 4: Based on the optimized hyperparameters, establish a PFA-GRU initial sub-model in combination with historical landslide displacement data;

[0059] Step 5: Each time the latest monitoring data D t is obtained, based on the online learning strategy, the input sub-model is trained in real time to obtain an online learning sub-model and predict the landslide displacement DP at the t+1 moment t+1 , and output the prediction result DP t+1 ;

[0060] Step 6: After obtaining the next monitoring data D t+1 , combine it with DP t+1 to evaluate the accuracy of the prediction result, and then use D t+1Assign to D t , that is, D t = D t+1 , if the prediction deviation is less than the allowable value, the online learning sub-model is used as the input sub-model and returned to step five for the next round of prediction; if the prediction deviation is greater than the allowable value, the sliding window transfer training method is adopted, then the PFA-GRU dynamic model reconstruction is carried out, and the PFA-GRU dynamic reconstruction sub-model is used as the input sub-model and returned to step five, and the next round of prediction is carried out.

[0061] Furthermore, in step two, the hyperparameters required for the PFA-GRU-OL-MRT model include the window size W1 of the PFA-GRU initial sub-model, the window size W2 of the online learning sub-model, the window reduction scale W of the PFA-GRU dynamic reconstruction sub-model minus , the minimum window size W of the PFA-GRU dynamic reconstruction sub-model min , the training iteration times initial_epochs of the PFA-GRU initial sub-model, the training iteration times OL_epochs of the online learning sub-model, the training iteration times MRT_epochs of the PFA-GRU dynamic reconstruction sub-model, the sliding window transfer training times initial_num of the PFA-GRU initial sub-model, the sliding window transfer training times MRT_num of the PFA-GRU dynamic reconstruction sub-model, the total number of PFA particles swarmsize, the PFA iteration times maxiter, the allowable value Re of the prediction deviation, the upper limit ub and the lower limit lb of the hyperparameter optimization range.

[0062] Furthermore, in step three, the particle firework algorithm is an improved algorithm of the particle swarm optimization algorithm, and its steps are as follows:

[0063] (1) Initialization

[0064] In the particle firework algorithm, the initial particles are generated by random initialization:

[0065] X i 0 = lb+(ub-lb)×rand(0,1)

[0066] where i is the particle number, X i represents the position of the i-th particle, the superscript represents the iteration times, where 0 represents the initialization stage, lb is the lower limit of the optimization range, ub is the upper limit of the optimization range, rand(0,1) represents an M-dimensional vector composed of random numbers between 0 and 1, where M represents the dimension of the problem to be solved; in the initialization stage, the particle firework algorithm only randomly generates 80% of the total number of particles, and the remaining 20% of the particles are generated in the local reinforcement search stage;

[0067] Evaluate the fitness of all initialized particles, and define the particle with the best fitness as the firework particle;

[0068] (II) Local intensification search

[0069] Randomly generate the remaining 20% of the particles within a certain range around the firework particle, and call these individuals spark particles; The total number N of spark particles k and the position G of each spark particle bj are expressed as:

[0070] N k = 0.2 × N

[0071] G bj = G b + 0.2 × (ub - lb) × e -2t / T × (2 × rand(0,1) - 1)

[0072] where N is the total number of particles, j is the number of the spark particle, G b is the position of the firework particle, t represents the current iteration number, and T represents the maximum iteration number;

[0073] Subsequently, calculate the fitness values of all spark particles;

[0074] (III) Particle update

[0075] In the particle update stage, different treatments are carried out in three cases:

[0076] (1) For other particles except the firework particle and the spark particles, its update process is:

[0077] V i t+1 = (1 - t / T)V i t / 2 + c1r1(P i t - X i t ) + c2r2(G b t - X i t )

[0078] X i t+1 = X i t + V i t+1

[0079] where V represents the velocity, c1 and c2 are learning factors; r1 and r2 are random numbers between 0 and 1, P iRepresents the optimal position in the search history of particle i.

[0080] (2) For the fireworks particles, directly update their positions to the position of the particle with the best fitness among all the fireworks particles and spark particles in each iteration, and define the velocity as 0;

[0081] (3) Delete all the spark particles;

[0082] (4) Elimination mechanism

[0083] Recalculate the fitness values of all particles, redefine the particle with the best fitness among all particles as the fireworks particle, update the historical optimal position P of particle i i , and at the same time delete the particle with the worst fitness, and randomly generate a new particle within the solution domain;

[0084] For each iteration, repeat steps (2) to (4) until the convergence condition is met or the maximum number of iterations is reached, and then output the position of the optimal solution.

[0085] In the above way, it not only ensures that the PFA algorithm has the ability of global search and jumping out of local optima, but also focuses more on the vicinity of the current globally optimal with the best fitness, thus increasing the local fast convergence ability of the algorithm.

[0086] Furthermore, in step (3), the sliding window transfer training method means that the model is trained using a sliding window. The window refers to a fixed scale of training data selected from the time series data, and all the data within the window is used as the input, and the next data outside the window is used as the output for training; and using a sliding window for model training means that after each window is trained, the window will slide along the time series towards the direction of the latest monitored data. The sliding will delete the earliest monitored data in the window and add the next monitored data after the window for retraining;

[0087] Under the sliding window transfer training method, use MSE ave as the fitness function when the particle fireworks algorithm optimizes the hyperparameters of the gated recurrent unit. The calculation method of MSE ave is as follows: ① Calculate the mean square error MSE between the predicted value and the corresponding measured value obtained at each sliding window position, ② Obtain the average value of MSE at all sliding window positions, which is MSE ave .

[0088] In the above way, it is possible to fully exploit the historical information of the landslide.

[0089] Furthermore, in step (3), the hyperparameters refer to the structural parameters of the GRU model, including the number of units, learning rate, Dropout rate, and L2 regularization strength of the GRU model.

[0090] Further, in step five, the online learning strategy means that for each newly obtained monitoring data D t , training data is acquired with an online learning window size W2, that is, using the W2 data before D t as the training input and D t as the training output to further train the PFA-GRU model for parameter fine-tuning, where W2 < the initial sub-model window size W1 of PFA-GRU.

[0091] Through online learning, the network parameters of the PFA-GRU prediction model will be continuously accumulated and updated as the monitoring data increases, improving the real-time prediction accuracy; on the other hand, since there is no need to reconstruct the model, the computational cost is minimized to the greatest extent, ensuring the prediction timeliness.

[0092] Further, in step six, for the dynamic model reconstruction, its model reconstruction window size W3 is dynamic. In the initial stage of model reconstruction, W3 = the online learning window size W2. If the prediction deviation after model reconstruction is still greater than the allowable value, then each time W minus earlier monitoring data is deleted to reduce the window size, and then the model is reconstructed continuously until the minimum window size W min is reached, at which point the model reconstruction stops and the sub-model with the highest prediction accuracy is output.

[0093] Through model reconstruction, the model can be automatically adjusted and the recent landslide trend can be identified when the landslide trend changes significantly, thus greatly improving the prediction accuracy when the landslide trend changes.

[0094] The test data comes from the real-time monitoring data of a landslide in Sichuan Province. There was an obvious change in the landslide trend of this landslide from 19:00 on August 26, 2020 to 16:00 on September 11, 2020. A total of 400 groups of measured data were obtained during this period.

[0095] The specific implementation process is as follows:

[0096] As Figure 1 , a dynamic real-time landslide displacement prediction method based on PFA-GRU-OL-MRT includes the following steps:

[0097] Step one: Obtain the historical landslide displacement data obtained from online monitoring and preprocess the data;

[0098] Step two: Set the hyperparameters required for the PFA-GRU-OL-MRT model;

[0099] The PFA-GRU-OL-MRT model includes three sub-models, namely the PFA-GRU initial sub-model, the online learning sub-model, and the PFA-GRU dynamic reconstruction sub-model;

[0100] Among them, the initial model window size W1 = 250, the online learning window size W2 = 50, the model reconstruction window reduction scale W minus = 5, the minimum window size W min = 20, the initial model training iteration times initial_epochs = 50, the online learning training iteration times OL_epochs = 5, the model reconstruction training iteration times MRT_epochs = 30, the initial model sliding window transfer training times initial_num = 5, the model reconstruction sliding window transfer training times MRT_num = 5, the total number of PFA particles swarmsize = 10, the PFA iteration times maxiter = 5, the allowable prediction deviation value Re = 5, and the upper limits ub = [100, 0.1, 0.5, 0.01] and lower limits lb = [10, 0.0001, 0, 01] of the hyperparameter optimization ranges of the gated recurrent unit (GRU), representing the number of units, learning rate, Dropout rate, and L2 regularization strength respectively.

[0101] Step 3: Based on the particle firework algorithm and historical landslide displacement data, adopt the sliding window transfer training method to optimize the hyperparameters of the gated recurrent unit, as Figure 2 shown;

[0102] Step 4: Based on the optimized hyperparameters, combine the historical landslide displacement data to establish the PFA-GRU initial sub-model;

[0103] Step 5: Each time the latest monitoring data D t is obtained, based on the online learning strategy, the input sub-model is trained in real time to obtain the online learning sub-model and predict the landslide displacement DP t+1 at the t+1 moment, and the prediction result DP t+1 is output;

[0104] Step 6: After obtaining the next monitoring data D t+1 , combine DP t+1 to evaluate the accuracy of the prediction result, and then assign D t+1 to D t , that is, D t = D t+1 . If the prediction deviation is less than the allowable value, the online learning sub-model is used as the input sub-model to return to Step 5 for the next round of prediction; if the prediction deviation is greater than the allowable value, the sliding window transfer training method is adopted, then the PFA-GRU dynamic model reconstruction is performed, and the PFA-GRU dynamic reconstruction sub-model is used as the input sub-model to return to Step 5 and perform the next round of prediction.

[0105] To verify the advantages of the method of the present invention and the influence of PFA, online learning, and model dynamic reconstruction on the prediction effect, the prediction results of a GRU static model without PFA optimization training, a static + online learning fine-tuning model (GRU-OL), a static + online learning fine-tuning + dynamic reconstruction model (GRU-OL-MRT), as well as an initial static model of GRU optimized by PFA (PFA-GRU) and a static + online learning fine-tuning model (PFA-GRU-OL) were compared. Among them, when there is no PFA optimization, the number of units, learning rate, Dropout rate, and L2 regularization strength of the GRU model are 50, 0.01, 0.25, and 0.005 respectively, and the other model parameters involved are the same as those set in the model proposed in this paper.

[0106] Figures 3 to 5 The comparison of the prediction effects of three models under PFA optimization and three models without PFA optimization for 150 prediction cycles is respectively shown. From Figure 3 it can be seen that the PFA-GRU model has obvious advantages over the single GRU model because PFA finds better GRU structure parameters and the trained GRU model is more excellent. However, due to the obvious change in the landslide trend in the later stage, even the better PFA-GRU model cannot adapt to the rapidly changing landslide trend.

[0107] From Figure 4 and Figure 5 it can be seen that after adding online learning and model reconstruction, the accuracy of landslide displacement prediction is significantly improved. Among them, although the prediction trends of the PFA-GRU-OL model and the GRU-OL model are consistent with the measured values, due to the rapid change of the landslide trend, it causes large fluctuations in the predicted values and affects the accuracy of the prediction. The prediction accuracy of the PFA-GRU-OL-MRT model and the GRU-OL-MRT model has been greatly improved due to the introduction of the reconstruction mechanism. However, when not optimized by PFA, the volatility of the GRO model is significantly stronger, indicating that unreasonable hyperparameters will lead to a decline in prediction performance.

[0108] To further explore the performance of each model, four indicators, namely mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and goodness of fit R 2 are used to quantitatively evaluate each model. Among them, the closer MAE, MAPE, and RMSE are to 0, the better the effect, and the closer R 2 is to 1, the more accurate the prediction. The evaluation results of each model are shown in Table 1:

[0109] Table 1 Comparison of each evaluation index of each model

[0110] model MAE MAPE% RMSE <![CDATA[R 2 > GRU 17.2 60.4 24.9 0.188 GRU-OL 2.42 7.84 4.91 0.840 GRU-OL-MRT 1.73 7.82 2.54 0.917 PFA-GRU 4.56 16.2 6.63 0.783 PFA-GRU-OL 1.99 6.50 3.77 0.877 PFA-GRU-OL-MRT 0.672 3.08 1.20 0.961

[0111] As shown in Table 1, the PFA-GRU-OL-MRT model achieved 0.672, 3.08, 1.20, and 0.961 in the four indicators of MAE, MAPE, RMSE, and goodness of fit R2, respectively, all of which were the optimal values in the evaluation of each model, fully demonstrating the advantages of this model in predicting trend-changing landslides. The results of the present invention have very important practical significance for improving the accuracy and response speed of landslide geological disaster prediction and increasing the emergency response time after early warning.

[0112] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT, characterized in that, It includes the following steps: Step 1: Obtain the historical landslide displacement data obtained from on-line monitoring and preprocess the data; Step 2: Set the hyperparameters required for the PFA-GRU-OL-MRT model; The PFA-GRU-OL-MRT model includes three sub-models, namely the PFA-GRU initial sub-model, the online learning sub-model and the PFA-GRU dynamic reconstruction sub-model; The hyperparameters required for the PFA-GRU-OL-MRT model include the window size W1 of the PFA-GRU initial submodel, the window size W2 of the online learning submodel, and the window reduction scale W of the PFA-GRU dynamic reconstruction submodel minus , the minimum window size W of the PFA-GRU dynamic reconstruction submodel min , the number of training iterations initial_epochs of the PFA-GRU initial submodel, the number of training iterations OL_epochs of the online learning submodel, the number of training iterations MRT_epochs of the PFA-GRU dynamic reconstruction submodel, the number of sliding window transfer training times initial_num of the PFA-GRU initial submodel, the number of sliding window transfer training times MRT_num of the PFA-GRU dynamic reconstruction submodel, the total number of PFA particles swarmsize, the number of PFA iterations maxiter, the allowable value of the prediction deviation Re, the upper limit ub and the lower limit lb of the hyperparameter optimization range; Step 3: Based on the particle fireworks algorithm and the historical landslide displacement data, adopt the sliding window transfer training method to optimize the hyperparameters of the gated recurrent unit; The particle fireworks algorithm is an improved algorithm of the particle swarm optimization algorithm, and its steps are as follows: (1) Initialization In the particle fireworks algorithm, the initial particles are generated by random initialization: X i 0 = lb+(ub - lb)×rand(0,1) where i is the particle number, and X i represents the position of the i-th particle, and the superscript represents the iteration number, where 0 represents the initialization stage, lb is the lower limit of the optimization range, ub is the upper limit of the optimization range, rand(0,1) represents an M-dimensional vector composed of random numbers between 0 and 1, where M represents the dimension of the problem to be solved; in the initialization stage, the particle firework algorithm only randomly generates 80% of the total number of particles, and the remaining 20% of the particles are generated in the local intensification search stage; Evaluate the fitness of all the initialized particles, and define the particle with the optimal fitness as the firework particle; (2) Local intensification search Centered on the firework particles, the remaining 20% of the particles are randomly generated within a certain range around them, and these individuals are called spark particles; the total number N of the spark particles k and the position G of each spark particle bj are expressed as: N k = 0.2 × N G bj = G b + 0.2×(ub - lb)×e -2t / T ×(2×rand(0,1) - 1) where N is the total number of particles, j is the number of spark particles, G b is the position of the fireworks particle, t represents the current iteration number, and T represents the maximum iteration number; Subsequently, calculate the fitness values of all the spark particles; (3) Particle update In the particle update stage, different treatments are carried out in three cases: (1) For the particles other than the firework particles and the spark particles, their update process is: V i t+1 = (1 - t / T)V i t / 2 + c1r1(P i t - X i t ) + c2r2(G b t - X i t ) X i t+1 = X i t + V i t+1 Among them, V represents velocity, c1 and c2 are learning factors; r1 and r2 are random numbers between 0 and 1, and P i represents the optimal position in the search history of particle i; (2) For the firework particles, directly update to the position of the particle with the optimal fitness among all the firework particles and the spark particles in each iteration, and the speed is defined as 0; (3) Delete all the spark particles; (4) Elimination mechanism Recalculate the fitness values of all particles, redefine the particle with the best fitness among all particles as the firework particle, and update the historical best position \(P\) of particle \(i\). i Meanwhile, delete the particle with the worst fitness and randomly generate a new particle within the solution domain. For each iteration, repeat steps (2) to (4) until the convergence condition is met or the maximum number of iterations is reached, and then output the position of the optimal solution; Step 4: Based on the optimized hyperparameters, establish the PFA-GRU initial sub-model in combination with the historical landslide displacement data; Step Five: Obtain the latest monitoring data D each time t , and perform real-time training on the input sub-model based on the online learning strategy to obtain an online learning sub-model and predict the landslide displacement DP at time t+1 t+1 , and output the prediction result DP t+1 ; Step 6: Obtain the next monitoring data D t+1 After that, combine with DP t+1 Evaluate the accuracy of the prediction result, and then assign D t+1 to D t , that is, D t = D t+1 . If the prediction deviation is less than the allowable value, return the online learning sub-model as the input sub-model to Step 5 for the next round of prediction; if the prediction deviation is greater than the allowable value, adopt the sliding window transfer training method to perform PFA-GRU dynamic model reconstruction, return the PFA-GRU dynamic reconstruction sub-model as the input sub-model to Step 5, and conduct the next round of prediction.

2. The real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT according to claim 1, characterized in that In step 3, the sliding window transfer training method refers to using a sliding window for model training. The window refers to that the scale of the training data selected from the time series data is fixed, and all the data within the window are used as the input, and the next data outside the window is used as the output for training; and using a sliding window for model training means that after each window is trained, the window will slide along the time series towards the direction of the latest monitoring data. The sliding will delete the earliest monitoring data in the window and add the next monitoring data after the window to re-train; Under the sliding window rotation training method, the MSE ave is used as the fitness function when the particle fireworks algorithm optimizes the hyperparameters of the gated recurrent unit. The calculation method of the MSE ave is as follows: ① Calculate the mean square error MSE between the predicted value and the corresponding measured value obtained at each sliding window position, and ② Obtain the average value of the MSE at all sliding window positions, which is the MSE ave .

3. A real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT according to claim 1, characterized in that, In step 3, the hyperparameters include the number of units of the gated recurrent unit model, the learning rate, the Dropout rate, and the L2 regularization strength.

4. A real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT according to claim 1, characterized in that, In step 5, the online learning strategy means that every time a latest monitoring data D is obtained t , training data is acquired with an online learning window size W2, that is, using the W2 data before D t as the training input and D t as the training output to further train the PFA-GRU model for parameter fine-tuning, where W2 < the initial sub-model window size W1 of the PFA-GRU.

5. A real-time prediction method for dynamic landslide displacement based on PFA-GRU-OL-MRT according to claim 1, characterized in that In Step 6, for the dynamic model reconstruction, the window size W3 of the model reconstruction is dynamic. In the initial stage of the model reconstruction, W3 = the online learning window size W2. If the prediction deviation after the model reconstruction is still greater than the allowable value, then each time minus monitoring data with earlier time is deleted to reduce the window size, and then the model is reconstructed continuously until the minimum window size W min is reached, the model reconstruction is stopped, and the sub-model with the highest prediction accuracy is output.

Images