A Landslide Displacement Prediction Method Based on Singular Spectrum Analysis and ARIMA-LSTM Combined Model

Abstract

This invention relates to the field of landslide displacement prediction technology, specifically to a landslide displacement prediction method based on a combined SSA-SE-ARIMA-LSTM model. The prediction process is controlled by a Python program, automating landslide prediction. This method uses historical landslide displacement monitoring data as the research object. First, a landslide data preprocessing method based on singular spectrum analysis (SSA) employs a "decomposition-denoising-reconstruction" approach to reduce noise in the landslide displacement data. Second, an LSTM model and an ARIMA model are constructed to predict the reconstructed non-stationary and stationary sequences, respectively. Finally, the prediction results are weighted and summed to obtain a fused cumulative displacement prediction sequence. Compared with existing technologies, the SSA-SE-ARIMA-LSTM model effectively removes noise from the original displacement sequence, improving the accuracy of landslide prediction. Furthermore, the accuracy and reliability of the SSA-SE-ARIMA-LSTM model in univariate prediction are superior to those of a single model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of landslide displacement prediction, and particularly relates to a landslide displacement prediction method based on singular spectrum analysis and an ARIMA-LSTM combined model. BACKGROUND

[0002] Landslide is a common and harmful geological disaster. The occurrence of landslide not only threatens people's life and property safety, but also causes damage to the ecological environment. Reliable and accurate prediction of the motion state of landslide is a key means to reduce the damage and loss caused by landslide disasters.

[0003] The evolution of landslide is a complex nonlinear process superimposed by multiple influencing factors. Direct use of the original displacement sequence for prediction cannot guarantee accuracy, and the original data has a lot of noise interference. Therefore, the displacement sequence is decomposed into a periodic term affected by external factors and a trend term controlled by the mechanical factors of the landslide body for separate prediction. Current methods for decomposing landslide displacement sequences include: classical decomposition method and its improved method X-11 decomposition method; STL decomposition method which decomposes time series into local trend, seasonality and residual; empirical mode decomposition (EMD) and its improved methods EEMD, CEEMD, etc. which decompose time series into multiple intrinsic mode functions (IMF) and a residual term; and variational mode decomposition (VMD) which also decomposes time series into multiple IMFs; singular spectrum analysis (SSA) which can decompose time series into trend, periodicity, noise, etc. and remove noise in the original sequence.

[0004] Some nonlinear time series prediction methods are widely used in component sequence prediction, such as multivariate nonlinear regression analysis and artificial neural network. Luohuanyuan et al. used VMD to decompose the original displacement and applied GRU recurrent neural network to the prediction of component displacement. Anbei et al. used VMD to decompose the original displacement, used grey correlation analysis to determine the influencing factors of the periodic term, and used BP neural network for periodic term prediction and quadratic exponential smoothing for trend term prediction. Shangmin et al. used exponential smoothing to decompose the original displacement of landslide, used the influencing factors as input, used the SVR model optimized by simulated annealing to output the prediction value of the periodic term, and used a cubic polynomial to predict the trend term.

[0005] The common point of the above models is that the influencing factors such as rainfall and reservoir water level are added to the prediction model, which improves the prediction accuracy of the model to a certain extent. However, in most cases, the data of landslide influencing factors such as rainfall is difficult to obtain, and the accuracy of prediction using only landslide displacement data cannot meet the expectation. Therefore, it is crucial to establish a landslide displacement univariate prediction model with practical physical significance. SUMMARY

[0006] The present application aims to provide a landslide displacement prediction method based on singular spectrum analysis and ARIMA-LSTM combined model, aiming at automatically predicting the landslide with single displacement monitoring data, while improving the prediction accuracy and efficiency.

[0007] To achieve the above-mentioned purpose, the present application provides a landslide displacement prediction method based on singular spectrum analysis and ARIMA-LSTM combined model, comprising the following steps:

[0008] Step 1: obtaining landslide cumulative displacement history monitoring data from the database;

[0009] Step 2: decomposing, denoising and reconstructing the landslide cumulative displacement history monitoring data based on singular spectrum analysis and sample entropy to obtain displacement component sequence;

[0010] Step 3: performing stationarity test on the obtained displacement component sequence, and dividing the data set and normalizing respectively;

[0011] Step 4: establishing and training ARIMA-LSTM model according to the stationarity test result;

[0012] Step 5: saving the optimal model and optimal parameters to a file respectively;

[0013] Step 6: calling the saved optimal model file and optimal parameter file to predict the component sequence;

[0014] Step 7: reversing the normalization of each component sequence prediction result, and adding each component with equal weight to obtain the landslide cumulative displacement prediction sequence.

[0015] Optionally, the process of "decomposition, denoising and reconstruction" comprises the following steps:

[0016] Using SSA to decompose the original landslide cumulative displacement into periodic sequence, trend sequence and noise sequence, calculating the contribution rate of each component sequence and removing the noise component to retain the effective displacement sequence;

[0017] Reconstructing the sample entropy of each effective displacement sequence, and reconstructing the sequences with similar entropy values according to the size of sample entropy value to form high-frequency sequence, medium-frequency sequence and low-frequency sequence;

[0018] Performing stationarity test on the component sequence, distinguishing between stationary sequence and non-stationary sequence, dividing the training set and test set according to the ratio of 7:3, and normalizing the data.

[0019] Optionally, in the process of sample entropy reconstruction, the pattern dimension m is 2 and the similar tolerance threshold r is 0.2.

[0020] Optionally, in the process of establishing and training the ARIMA-LSTM model, the stationary sequence is trained by constructing an ARIMA model, and the non-stationary sequence is trained by constructing an LSTM model; the ARIMA model parameters are optimized by a grid search algorithm, and the optimal parameter combination and optimal model are saved to a file; the LSTM model parameters are optimized using a genetic algorithm to obtain the optimal model and optimal parameters.

[0021] Optionally, in the process of optimizing the ARIMA model parameters by the grid search algorithm, the parameter interval of the ARIMA model is set as follows: the autoregressive term number p is in the value interval [0, 6], the difference order d is in the value interval [0, 5], the moving average term number q is in the value interval [0, 5], the step size is set to 1, and the coefficient of determination R 2 is the fitness function.

[0022] Optionally, in the process of optimizing the LSTM model parameters using the genetic algorithm, the evolution number is 30, the population number is 20, the crossover rate is 0.6, the mutation rate is 0.1, and the coefficient of determination R 2 is the fitness function.

[0023] The present application provides a landslide displacement prediction method based on singular spectrum analysis and ARIMA-LSTM combined model, which realizes the automation of landslide prediction by controlling the prediction process through Python program. The method takes landslide displacement historical monitoring data as the research object, first uses the "decomposition-denoising-reconstruction" landslide data preprocessing method based on singular spectrum analysis to achieve the purpose of denoising the landslide displacement data. Secondly, an LSTM model and an ARIMA model are constructed to predict the reconstructed non-stationary sequence and stationary sequence respectively, and finally the prediction results are added equally to obtain the fused cumulative displacement prediction sequence. Compared with the prior art, the "decomposition-denoising-reconstruction" method based on singular spectrum analysis effectively eliminates the noise in the original displacement sequence, improves the accuracy of landslide prediction, and the accuracy and reliability of the SSA-SE-ARIMA-LSTM model in single variable prediction are better than those of a single model. The Python program controls the running process of the model, automatically processes the intermediate data, and makes the various components of the model run efficiently and orderly. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below for the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0025] Figure 1is a displacement prediction flowchart of a landslide displacement prediction method based on singular spectrum analysis and ARIMA-LSTM combined model.

[0026] Figure 2 is a singular spectrum analysis result schematic diagram of the present application.

[0027] Figure 3 is a prediction model structure and flowchart schematic diagram of the present application.

[0028] Figure 4 is an ARIMA model prediction stationary sequence result graph of the present application.

[0029] Figure 5 is an LSTM model prediction non-stationary sequence result graph of the present application.

[0030] Figure 6 is a cumulative displacement prediction result schematic diagram of the present application. DETAILED DESCRIPTION

[0031] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0032] Some English abbreviations used in the present application are interpreted as follows:

[0033] ARIMA: Auto Regressive Integrated Moving Average model, difference integrated moving average autoregressive model;

[0034] LSTM: Long Short Term Memory, long short term memory recurrent neural network;

[0035] SSA: Singular Spectrum Analysis, singular spectrum analysis;

[0036] GA: Genetic Algorithm, genetic algorithm;

[0037] R 2 : R-Square, determination coefficient;

[0038] RMSE: Root Mean Square Error, root mean square error;

[0039] MAPE: Mean Absolute Percentage Error, mean absolute percentage error;

[0040] The application provides a landslide displacement prediction method based on a singular spectrum analysis and ARIMA-LSTM combined model, and comprises the following steps:

[0041] S1: obtaining landslide cumulative displacement historical monitoring data from a database;

[0042] S2: performing "decomposition-denoising-reconstruction" on the landslide cumulative displacement historical monitoring data based on singular spectrum analysis and sample entropy to obtain a displacement component sequence;

[0043] S3: performing stationarity test on the obtained displacement component sequence, and dividing data sets and performing normalization processing respectively;

[0044] S4: establishing and training an ARIMA-LSTM model according to the stationarity test result;

[0045] S5: saving the optimal model and optimal parameters to files respectively;

[0046] S6: calling the saved optimal model file and optimal parameter file to predict the component sequence;

[0047] S7: performing reverse normalization on the component sequence prediction results, and adding the components with equal weights to obtain a landslide cumulative displacement prediction sequence.

[0048] The specific implementation steps can be divided into three parts, and the implementation flowchart is as shown in Figure 1 .

[0049] Firstly, data set acquisition and processing (including steps S1 and S2 described above).

[0050] (1) Data set acquisition.

[0051] The data set uses real-time displacement monitoring data and historical monitoring data generated by a receiver device, and the data acquisition process is completed by a Python program. Firstly, historical displacement data is obtained from a database, and the time length and time step of the data are configured through a configuration file. This process is realized through a multi-process mode.

[0052] (2) Data "decomposition-denoising-reconstruction".

[0053] Considering that there is noise in the original displacement monitoring data, in order to improve the robustness and prediction accuracy of the model, the training data is processed by "decomposition-denoising-reconstruction". The application proposes a "decomposition-denoising-reconstruction" method based on singular spectrum analysis and sample entropy reconstruction, and the steps are as follows:

[0054] a) Decomposition: Singular Spectral Analysis (SSA) is used to decompose the cumulative landslide displacement into periodic, trend, and noise sequences. This includes four steps: embedding, singular value decomposition, grouping, and diagonal averaging. The matrices obtained from grouping are then transformed into new sequences Z of length N using the diagonal formula, which are the component sequences. Let L be the window length and K be the embedding dimension. * =min(L,K), K * The diagonal formula for max(L,K) is as follows; please refer to the singular spectrum analysis results. Figure 2 .

[0055]

[0056] b) Denoising: The series of components obtained from step a) contains effective displacement components and noise. Calculate the contribution rate of each component, and then calculate the cumulative contribution rate. The sequence corresponding to the node where the cumulative contribution rate curve tends to stabilize is noise and can be removed. The contribution rate β is calculated using the following formula, where λ... i For D=XX T The eigenvalues, where X is the original displacement time series, and d = L * .

[0057]

[0058] c) Reconstruction: Set the pattern dimension m to 2 and the similarity tolerance threshold r to 0.2. Calculate the sample entropy of each component. Then, sort the sample entropy values ​​of the displacement component sequence in descending order and classify the sample entropy values ​​according to the change step size of 0.1. The component sequences are reconstructed into three categories: high frequency sequence, medium frequency sequence, and low frequency sequence. At this time, the model prediction result is optimal. The sample entropy calculation formula is as follows.

[0059]

[0060] d) Classification: The reconstructed sequences obtained from step c) have different stationary properties. These components are distinguished by performing a stationarity test. Then, the ARIMA model is used to predict stationary sequences, and the LSTM model is used to predict non-stationary sequences.

[0061] Based on the above steps, the original displacement data is extracted and preprocessed, and the original sequence is finally classified into stationary and non-stationary sequences. Due to their different characteristics, different models are used for prediction.

[0062] The second step is model training and prediction (steps S3 to S6).

[0063] The dataset is divided and normalized for all stationary and non-stationary sequences, with 70% used as the training set and 30% as the test set. The training set is used for model training. For the prediction model structure and process, please refer to [link / reference]. Figure 3.

[0064] (1) Train the ARIMA model using stationary sequences. The principle formula of the ARIMA model is as follows:

[0065] ARIMA (p,d,q) =AR (p) +DIFF (d) +MA (q)

[0066] The parameters p, d, and q are found using a grid search algorithm. The number of autoregressive terms p ranges from [0, 6], the difference order d ranges from [0, 5], and the number of moving average terms q ranges from [0, 5]. (p) For a p-order autoregressive model, DIFF (d) For a d-order difference model, MA (q) This is a q-order moving average model. Its training process is controlled by a Python program, and the steps are as follows:

[0067] Step 1: Initialize the grid search algorithm parameters, enumerate the possible parameters, and have the Python program calculate all possible parameter combinations. Then, input all the parameters into the model for model training. This process is run in a multi-threaded manner, and the model input data is the first 80% of the training set data.

[0068] Step 2: Construct the ARIMA model with the coefficient of determination R. 2 As the fitness function of the model, the last 20% of the training set is used as the validation set input. The fitness of the model is calculated for each training iteration. After reaching a set number of iterations, the R-value of the model for each parameter combination is calculated. 2 For comparison, R 2 The parameter combination closest to 1 is the optimal parameter combination;

[0069] Step 3: Save the combination to the parameter file, and the corresponding model to the model file. This yields the optimal parameters and optimal model for the ARIMA model.

[0070] (2) Train the LSTM model using non-stationary sequences. Set the hyperparameters of the LSTM model: the number of neurons and the learning rate are obtained by genetic algorithm (GA), including the following steps:

[0071] Step 1: Genetic algorithm parameters are set as follows: number of generations: 30, population size: 20, crossover rate: 0.6, mutation rate: 0.1, and model input data is the first 80% of the training set.

[0072] Step 2: Construct the LSTM model and determine the coefficient of determination R. 2As the fitness function, the last 20% of the data of the training set is input as the validation set, the fitness of each trained model is calculated, and the R 2 of each parameter combination of the model is compared after reaching the set number of iterations. 2 The parameter combination closest to 1 is the optimal parameter combination.

[0073] Step 3, save the combination to the parameter file, and save the corresponding model to the model file. Thus, the optimal parameters and the optimal model of the LSTM model are obtained.

[0074] (3) The above processes (1) and (2) save the optimal model and the optimal parameters of the ARIMA model and the LSTM model to the file, and the prediction program calls the model file and the parameter file of ARIMA and LSTM to preprocess and predict the test set data, respectively. The prediction results are shown in Figure 4 、 Figure 5 .

[0075] The above model training and prediction are performed respectively, the model is trained every 10 minutes, the model file and the parameter file are updated, and the model prediction is performed every second by calling the latest model file and parameter file, and the process is independent, which ensures the accuracy of the prediction and improves the prediction efficiency.

[0076] Step 3, predict the data fusion processing (step S7).

[0077] The prediction results of the ARIMA and LSTM prediction models are obtained respectively, and the prediction results are added equally to obtain the final cumulative displacement prediction value, wherein the R 2 of the original cumulative displacement and the predicted cumulative displacement is 0.985, the RMSE is 6.727 mm, and the MAPE is 0.221%, which proves that the present application has high prediction accuracy and low prediction error. The cumulative displacement prediction result is shown in Figure 6 .

[0078] The above only discloses one preferred embodiment of the present application, and of course cannot limit the scope of the present application. Those skilled in the art can understand that the above-mentioned embodiment can be implemented in whole or in part, and equivalent changes made according to the claims of the present application still fall within the scope of the present application.

Claims

1. A landslide displacement prediction method based on a singular spectrum analysis and ARIMA-LSTM combined model, characterized in that, The method comprises the following steps: Step 1: obtaining landslide cumulative displacement history monitoring data from a database; Step 2: decomposing, denoising and reconstructing the landslide cumulative displacement history monitoring data based on singular spectrum analysis and sample entropy to obtain displacement component sequences; The process of decomposing, denoising and reconstructing comprises the following steps: The original landslide cumulative displacement is decomposed into periodic sequences, trend sequences and noise sequences by SSA, the contribution rate of each component sequence is calculated, noise components are removed, and effective displacement sequences are retained; Sample entropy reconstruction is performed on each effective displacement sequence, entropy values of sequences with similar entropy values are reconstructed according to the size of sample entropy values, and high-frequency sequences, medium-frequency sequences and low-frequency sequences are formed; Stationarity test is performed on the component sequences to distinguish stationary sequences and non-stationary sequences, training set and test set are divided according to a ratio of 7:3, and data normalization is performed; In the process of sample entropy reconstruction, the pattern dimension m is 2, and the similar tolerance threshold r is 0.2; Step 3: performing stationarity test on the obtained displacement component sequences and dividing and normalizing the data sets respectively; Step 4: establishing and training an ARIMA-LSTM model according to the stationarity test results; In the process of establishing and training the ARIMA-LSTM model, the stationary sequences are trained by constructing an ARIMA model, and the non-stationary sequences are trained by constructing an LSTM model; the parameters of the ARIMA model are optimized by a grid search algorithm, the optimal parameter combination and the optimal model obtained by training are saved to a file; the parameters of the LSTM model are optimized by a genetic algorithm, and the optimal model and the optimal parameters are obtained; Step 5: saving the optimal model and the optimal parameters to files respectively; Step 6: calling the saved optimal model file and optimal parameter file to predict the component sequences; Step 7: reversing the normalization of the component sequence prediction results, adding each component with equal weight to obtain a landslide cumulative displacement prediction sequence.

2. The landslide displacement prediction method based on the combination of singular spectrum analysis and ARIMA-LSTM model according to claim 1, wherein In the process of searching the parameters of ARIMA model by grid search algorithm, the parameter interval of ARIMA model is set as follows: the value interval of autoregressive term p is [0, 6], the value interval of difference order d is [0, 5], the value interval of moving average term q is [0, 5], the step length is set as 1, and the coefficient is the fitness function.

3. The landslide displacement prediction method based on the combination of singular spectrum analysis and ARIMA-LSTM model according to claim 2, wherein In the process of using genetic algorithm to optimize the parameters of LSTM model, the evolution number is 30, the population size is 20, the crossover rate is 0.6, the mutation rate is 0.1, and the determination coefficient is the fitness function.

Images