Power transmission line geological disaster time sequence InSAR prediction method based on improved LSTM model

Abstract

According to the power transmission line geological disaster time sequence InSAR prediction method based on the improved LSTM model, an improved long short-term memory network LSTM model is constructed, and a basic data set is established for LSTM model training by using obtained power transmission line monitoring area time sequence InSAR deformation result data. And a regularization technology, an attention mechanism and an adaptive learning rate adjustment strategy are introduced in a model training process, so that an overfitting risk is reduced, the generalization ability of the model is improved, and a high-precision deformation prediction model for time sequence InSAR monitoring of the power transmission line is obtained. The method is of great significance to early warning and deformation trend analysis of geological disasters around the power transmission line, and can provide scientific basis and decision support for safe operation of power transmission facilities, thereby effectively reducing potential hazards of the geological disasters to a power system.

Description

TECHNICAL FIELD

[0001] The present application relates to power transmission line geological disaster monitoring and prediction, in particular to a power transmission line geological disaster time series InSAR prediction method based on an improved LSTM model. BACKGROUND

[0002] Geological vulnerability, rainfall, weathering and human activities lead to stress imbalance of rock strata, especially in karst mountainous areas, which easily cause collapse and landslide, seriously threatening the safety of power transmission lines and causing incalculable losses.

[0003] InSAR technology has become an important tool for landslide monitoring with its wide coverage and high resolution. In the past two decades, multi-temporal InSAR methods based on permanent scatterers (PS) and distributed scatterers (DS) have been developed to identify risks and track landslide deformation. Permanent scatterer interferometry (PSI) technology selects stable reference points to remove noise and ensure monitoring accuracy. For example, IPTA, StaMPS and quasi-permanent scatterer technology have promoted the application of InSAR. The small baseline subset (SBAS) method uses short temporal and spatial baseline interferograms to alleviate decorrelation problems. Improved methods such as multi-scale InSAR time series, intermittent SBAS (ISBAS) and MintPy further improve monitoring efficiency. SqueeSAR uses maximum likelihood estimation and redundant interference network strategy to deal with decorrelation.

[0004] However, mountainous areas with dense vegetation face challenges such as temporal and spatial decorrelation, complex terrain and atmospheric interference, which seriously weaken the accuracy of InSAR. In particular, along the power transmission line, karst landslides occur frequently and are covered with vegetation, resulting in a lack of PS targets and limited deformation data acquisition. At the same time, the DS method is complex to calculate, making it difficult to establish a connection between deformation and damage mechanisms, affecting the understanding and prediction of landslide dynamics.

[0005] To address these challenges, researchers have explored new methods to improve phase estimation accuracy in vegetation areas, such as using eigenvalue decomposition (EVD) of the covariance matrix to optimize interferogram phase estimation and increase the number and quality of coherent targets. Regularization strategies such as M-estimation, Hadamard-spectral regularization and shrinkage techniques also show potential, especially in long-term coherence scenarios.

[0006] However, it should be noted that these strategies are still limited in fast decorrelation environments. Low-coherence interferometry performs well in PSI but not in DSI. Fast decorrelation caused by vegetation and terrain weakens the reliability of phase measurements and the accuracy of deformation time series, posing new challenges for landslide monitoring. SUMMARY

[0007] In view of the shortcomings of the prior art, the present application provides a power transmission line geological disaster time series InSAR prediction method based on an improved LSTM model, which introduces regularization techniques, attention mechanisms and adaptive learning rate adjustment strategies, effectively reduces the risk of overfitting and improves the generalization ability of the model, and can be effectively used for early InSAR time series monitoring and early warning analysis of power transmission line geological disasters.

[0008] To achieve the above object, the application adopts the following technical solutions:

[0009] The power transmission line geological disaster time series InSAR prediction method based on the improved LSTM model comprises the following steps:

[0010] Step 1: The original time series InSAR monitoring data is preprocessed by using time series analysis and data standardization methods;

[0011] Step 2: An LSTM model is established for the processed time series, and an attention mechanism is introduced into the LSTM model to establish an Attention-LSTM model;

[0012] Step 3: The Attention-LSTM model is subjected to adaptive regularization and adaptive learning rate adjustment strategies, and the cosine annealing algorithm is combined with the MSE loss to control the update of the learning rate during the training process, thereby improving the training efficiency and model performance;

[0013] Step 4: The input data is smoothed by using a sliding window method, and the prediction results are weighted and fused according to the time correlation to obtain the prediction results.

[0014] Step 1 comprises:

[0015] Step 1.1: The original time series InSAR monitoring data is denoised;

[0016] Step 1.2: The missing values are completed based on the neighborhood mean;

[0017] Step 1.3: The data is normalized by using a robust standardization method.

[0018] In step 2, for a long time series, a hierarchical LSTM structure is used to gradually extract deep features in the time series through a multi-layer LSTM network to improve the prediction accuracy; the output of each layer is used as the input of the next layer, deep features in the time series are gradually extracted through a multi-layer LSTM network, the bottom layer LSTM captures short-term time dependence, and the high-layer LSTM captures long-term time dependence; the long time series is a time series longer than 1 year.

[0019] The attention mechanism is introduced into the LSTM model, specifically:

[0020] Step 2.1: The LSTM layer outputs a hidden state h t at each time step t, which contains sequence information from the beginning to the current time step;

[0021] Step 2.2: Calculate an attention weight α t for each time step t, which represents the importance of the time step to the current prediction task; t

[0022] e t = h t · s t-1 (1)

[0023]

[0024] where h t is the hidden state of the current time step, s t-1 is the context vector of the previous time step, T is the sequence length, and α t is the normalized weight of the current time step;

[0025] Step 2.3: Use the attention weight to weight and sum the hidden state of the LSTM to generate the context vector c t :

[0026]

[0027] The context vector c t contains the weighted information of all time steps in the sequence, which can highlight the features of key time steps;

[0028] Step 2.4: Combine the context vector c t with the hidden state h t of the current time step to generate the final prediction output:

[0029] y t = OutputLayer([h t ; c t ]) (4)

[0030] where [h t ; c t ] represents concatenating h t and c t .

[0031] In step 3, the learning rate formula of cosine annealing is:

[0032] ​

[0033] where η t is the learning rate, η min is the minimum learning rate, η max is the maximum learning rate, T cur is the number of iterations within the current period, T max is the total number of iterations for a period.

[0034] The learning rate is dynamically adjusted according to the current MSE loss combined with the cosine annealing algorithm:

[0035]

[0036] where η t is the adjusted learning rate, α is a hyperparameter for controlling the degree of influence of the MSE loss on the learning rate, and MSE is the current mean square error loss.

[0037] Step 4 specifically includes:

[0038] Step 4.1: sliding partitioning of time series data X = [x1, x2, …, x T ]:

[0039] X window = [x t , x t+1 , …, x t+w-1 ] (9)

[0040] where t is the starting time step of the window, and w is the window size.

[0041] Step 4.2: smoothing processing of data within each window:

[0042]

[0043] where ω i is the weight coefficient, which gradually decreases with increasing distance.

[0044] Step 4.3: weighted fusion of multiple prediction results according to time correlation in the prediction stage,

[0045]

[0046] where y is the fused prediction result, y i is the i-th prediction result, ω i is the time correlation weight of the i-th prediction result, and satisfies

[0047] The positive and beneficial effects of the present application are:

[0048] This paper proposes a time-series InSAR prediction method for transmission line geological hazards based on an improved LSTM model. This method is superior to other technologies in time-series InSAR prediction of transmission lines. At the same time, it enhances the model's ability to capture complex deformation features while improving prediction accuracy, which helps to accurately predict the deformation trend of geological hazards on transmission lines and avoid damage to transmission lines caused by geological hazards. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the present invention;

[0050] Figure 2 This is the structure of the Attention-LSTM model;

[0051] Figure 3 Training curve diagrams for the LSTM model and the improved LSTM model;

[0052] Figure 4 The prediction result curves of the time series prediction algorithm are as follows: the first row is the traditional statistical model Sarima algorithm, the second row is the LSTM model, and the third row is the improved Attention-LSTM model;

[0053] Figure 5 is the Pearson correlation coefficient and mean square error between the prediction results of the time series prediction algorithm and the original data. DETAILED DESCRIPTION

[0054] In order to facilitate ordinary technicians in this field to understand the purpose, technical solutions and advantages of the present invention, the present invention is further described in detail below in conjunction with specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0055] The time series InSAR prediction method for transmission line geological hazards based on the improved LSTM model uses time series analysis and data enhancement methods to preprocess the original time series InSAR monitoring data to improve the data quality and the input effect of the model.

[0056] like Figure 1 As shown in the figure, the time series InSAR prediction method for geological hazards of transmission lines based on the improved LSTM model includes the following steps:

[0057] Step 1: The original time series InSAR monitoring data are preprocessed using the time series analysis and data standardization method, specifically including:

[0058] Step 1.1: De-noise the original time series InSAR monitoring data to eliminate the influence of random noise;

[0059] Step 1.2: Fill in missing values based on neighborhood mean to ensure data integrity;

[0060] Step 1.3: Adopt robust standardization method to normalize data to eliminate the interference of outliers, and provide high-quality, standardized input data for model training.

[0061] Step 2: Establish an LSTM model for the processed time series and introduce an attention mechanism into the LSTM model to establish an Attention-LSTM model;

[0062] Step 3: Adopt adaptive regularization and adaptive learning rate adjustment strategies for the Attention-LSTM model, use cosine annealing algorithm combined with MSE loss to control the update of learning rate during training, and improve training efficiency and model performance;

[0063] Step 4: Smooth the input data using the sliding window method and weight the prediction results according to the time correlation to obtain the prediction results.

[0064] In step 2, the LSTM layer in the LSTM model is the core part of the model. LSTM controls the flow of information through its internal gating mechanism (including input gate, forget gate and output gate), so it can remember or forget information in the sequence. The attention mechanism calculates the weight through the Softmax function, converts the attention score to a probability distribution, and makes the model focus on important parts. After calculating the attention weight, weighted summation is performed, the hidden state of the input sequence is multiplied by the corresponding attention weight and summed to obtain a weighted context vector that is passed to a fully connected layer to generate the final output.

[0065] Preferably, for long time series (i.e. time series greater than 1 year), a hierarchical LSTM structure is adopted to gradually extract deep features in the time series through multiple LSTM networks to improve prediction accuracy; the hierarchical LSTM structure, the output of each layer is used as the input of the next layer, and deep features in the time series are gradually extracted through multiple LSTM networks. The bottom LSTM captures short-term temporal dependencies, and the high-level LSTM captures long-term temporal dependencies. Through the hierarchical structure, the model can capture both short-term and long-term temporal dependencies, avoid information loss in long time span, make the data better reflect the sequence rule, and improve the prediction performance.

[0066] Preferably, the attention mechanism is introduced into the LSTM model, specifically:

[0067] Step 2.1: The LSTM layer outputs a hidden state h t at each time step t, which contains sequence information from the beginning to the current time step;

[0068] Step 2.2: Calculate an attention weight α for each time step t t , which represents the importance of this time step to the current prediction task; t

[0069] e t = h t ·s t-1 (1)

[0070]

[0071] where h t is the hidden state of the current time step, s t-1 is the context vector of the previous time step (initialized as a zero vector), T is the sequence length, and α t represents the normalized weight of the current time step.

[0072] Step 2.3: Weighted sum the hidden states of the LSTM using the attention weight to generate the context vector c t :

[0073]

[0074] The context vector c t contains the weighted information of all time steps in the sequence, which can highlight the features of key time steps;

[0075] Step 2.4: Combine the context vector c t with the hidden state h t of the current time step to generate the final prediction output:

[0076] y t = OutputLayer([h t ; c t ]) (4)

[0077] where [h t ; c t ] represents concatenating h t and c t .

[0078] Step 3 adopts an adaptive learning rate adjustment strategy, which uses a cosine annealing algorithm combined with the MSE loss to control the update of the learning rate during training, improving training efficiency and model performance:

[0079] The learning rate formula of cosine annealing is:

[0080]

[0081] where η​t is the learning rate, η min is the minimum learning rate, η max is the maximum learning rate, T cur is the number of iterations within the current period, T max is the total number of iterations for one period.

[0082] According to the current MSE loss combined with the cosine annealing algorithm to dynamically adjust the learning rate:

[0083]

[0084] where: η t is the adjusted learning rate, α is a hyperparameter for controlling the influence of MSE loss on the learning rate, and MSE is the current mean square error loss.

[0085] The training loss curve comparison chart is shown in Figure 3 , using the adaptive learning rate adjustment strategy, dynamically adjusting the learning rate according to the loss change in the training process, accelerating the model convergence and avoiding the overfitting problem. The L1 and L2 regularization methods are introduced in the loss function, and the adaptive parameter is used to balance the regularization weight size, smooth the model output and preserve the high-dimensional data features, further improving the generalization ability of the model.

[0086] L1 and L2 regularization are introduced in the loss function:

[0087]

[0088] where y is the true value, is the predicted value, n is the sample size, ω is the weight coefficient, λ1 and λ2 are the L1 and L2 regularization coefficients, respectively, indicating their allocation proportion.

[0089] Step 4 specifically includes:

[0090] Step 4.1: sliding partition is performed on the time series data X = [x1, x2, …, xn]: T

[0091] X window = [x t , x t+1 , …, x t+w-1 ] (9)

[0092] where t is the starting time step of the window, and w is the window size;

[0093] Step 4.2: smooth the data in each window:

[0094]

[0095] where ω​i are weight coefficients, which gradually decrease with the increase of distance.

[0096] Step 4.3: The prediction stage weights and fuses multiple prediction results according to time correlation, further improving prediction performance.

[0097]

[0098] wherein, is the fused prediction result, y i is the i-th prediction result, ω i is the time correlation weight of the i-th prediction result, satisfying

[0099] The sliding window method can effectively smooth noise and improve data quality. The weighted fusion method can combine the advantages of multiple prediction results, improve the stability of prediction, and dynamically adjust the weight according to the time correlation, making the prediction result more reliable.

[0100] Figure 4 For comparative analysis experiments, the image shows the comparison of original data (ori_data) and data processed by SARIMA, LSTM and Attention-LSTM. The model preprocesses the original deformation data through time series analysis and data enhancement methods, effectively reducing the interference of noise and outliers, improving the quality of data and the input effect of the model. For long time series, a hierarchical LSTM structure is used to gradually extract deep features in the time series, enhancing the model's ability to capture long-term dependencies and improving prediction accuracy. After introducing the attention mechanism, the model can dynamically adjust the weight of each time step in the input sequence, enhancing the ability to capture key deformation features, making the prediction results of Attention-LSTM more accurate and close to the true value. In addition, in order to suppress the influence of environmental noise and random interference on deformation prediction, the sliding window method is used to smooth the input data, and the prediction results are weighted and fused according to the time correlation, significantly improving the stability and reliability of the prediction results.

[0101] Figure 5 For the comparison of Pearson correlation coefficient and mean square error (MSE) of the prediction results and the true data, the Pearson correlation coefficient calculation equation is:

[0102]

[0103] wherein, y i is the i-th true value, is the i-th predicted value, is the mean of the true value, is the mean of the predicted value.

[0104] The calculation formula of mean square error (MSE) is as follows:

[0105]

[0106] Wherein, y i is the i-th real value, is the i-th predicted value, and n is the sample quantity.

[0107] The Attention-LSTM performs best in both Pearson correlation coefficient (0.97) and MSE (0.05), the linear correlation between the predicted value and the real value is the strongest, and the prediction error is the smallest. Through the joint action of the improved points, the Attention-LSTM performs outstandingly in the time series task.

[0108] Therefore, the beneficial effects of the present application are:

[0109] 1. The hierarchical LSTM structure is adopted, the deep features in the time series are gradually extracted through the multi-layer LSTM network, and the prediction accuracy is improved.

[0110] 2. The attention mechanism is introduced into the LSTM model, the weights of each time step in the input sequence are dynamically adjusted, and the capture ability of the model to key deformation features is enhanced.

[0111] 3. The adaptive learning rate adjustment strategy is adopted, the learning rate is dynamically adjusted according to the loss change in the training process, so as to accelerate the convergence of the model and avoid overfitting.

[0112] 4. The L1 and L2 regularization are introduced into the loss function, the L1 and L2 regularization weight sizes are balanced through adaptive parameters, the model output is smoothed, and the high-dimensional data features are saved.

[0113] 5. In order to effectively suppress the influence of environmental noise and random interference on deformation prediction, the sliding window method is used to smooth the input data, and the prediction results are weighted and fused according to the time correlation, so as to improve the stability and reliability of the prediction.

[0114] It should be noted that the above examples are only used to illustrate the technical solutions of the present application, but not limit the present application. Other modifications or equivalent replacements to the technical solutions of the present application made by those skilled in the art should be covered in the scope of the claims of the present application, as long as they do not deviate from the spirit and scope of the present application.

Claims

1. A time-series InSAR prediction method for geological hazards on transmission lines based on an improved LSTM model is characterized by: The steps include: Step 1: Preprocess the original time series InSAR monitoring data using time series analysis and data standardization methods; Step 2: Build an LSTM model for the processed time series and introduce the attention mechanism into the LSTM model to build an Attention-LSTM model. Step 3: Adaptive regularization and adaptive learning rate adjustment strategies are applied to the Attention-LSTM model. The cosine annealing algorithm is combined with the MSE loss to control the learning rate update during training, improving training efficiency and model performance. Step 4: The input data is smoothed using the sliding window method, and the prediction results are weighted and fused according to time correlation to obtain the prediction results.

2. The method for predicting geological disasters in a time series of transmission lines using an improved LSTM model according to claim 1, wherein: Step 1 includes: Step 1.1: De-noise the original time series InSAR monitoring data; Step 1.2: Fill in missing values ​​based on the neighborhood mean; Step 1.3: Use robust normalization method to normalize the data.

3. The time series InSAR prediction method for geological hazards of power transmission lines based on the improved LSTM model according to claim 1 is characterized in that: In step 2, for the case of long time series, a hierarchical LSTM structure is adopted, and the deep features in the time series are gradually extracted through a multi-layer LSTM network to improve the accuracy of the prediction; in the hierarchical LSTM structure, the output of each layer is used as the input of the next layer, and the deep features in the time series are gradually extracted through a multi-layer LSTM network. The bottom layer LSTM captures short-term time dependencies, and the high-level LSTM captures long-term time dependencies; the long time series is a time series with a time series greater than 1 year.

4. The method for predicting geological hazards of power transmission lines using time series InSAR based on the improved LSTM model according to claim 1, characterized in that: The introduction of the attention mechanism into the LSTM model is as follows: Step 2.1: The LSTM layer outputs a hidden state h at each time step t t , the hidden state contains the sequence information from the beginning to the current time step; Step 2.2: For each time step t, calculate an attention weight α t , attention weight α t Indicates the importance of this time step to the current prediction task; e t =h t ·s t-1 (1) Among them, h t is the hidden state of the current time step, s t-1 is the context vector of the previous time step, T is the sequence length, α t represents the normalized weight of the current time step; Step 2.3: Use the attention weights to perform weighted summation on the hidden states of the LSTM to generate the context vector c t : context vector c t Contains weighted information of all time steps in the sequence, which can highlight the characteristics of key time steps; Step 2.4: Transform the context vector c t and the hidden state h at the current time step t Combined, the final prediction output is generated: y t =OutputLayer([h t ;c t ]) (4) where [h t ;c t ] means h t and c t Stitched together.

5. The method for predicting geological hazards of power transmission lines using time series InSAR based on the improved LSTM model according to claim 1, characterized in that: In step 3, the learning rate formula for cosine annealing is: where η t is the learning rate, η min is the minimum learning rate, η max is the maximum learning rate, T cur is the number of iterations in the current cycle, T max is the total number of iterations in a cycle; Dynamically adjust the learning rate based on the current MSE loss combined with the cosine annealing algorithm: Where: η t is the adjusted learning rate, α is a hyperparameter used to control the degree of influence of MSE loss on the learning rate, and MSE is the current mean squared error loss.

6. The method for predicting geological hazards of power transmission lines using time series InSAR based on the improved LSTM model according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: For the time series data X = [x1, x2, ..., x T ] to perform sliding split: X window =[x t ,x t+1 ,…,x t+w-1 ] (9) Where t is the starting time step of the window and w is the window size; Step 4.2: Smooth the data in each window: Among them, ω i is the weight coefficient, which decreases gradually with increasing distance. Step 4.3: In the prediction stage, multiple prediction results are weighted and fused according to time correlation. in, is the prediction result after fusion, y i is the i-th prediction result, ω i is the time correlation weight of the i-th prediction result, satisfying

Images