A Tailings Dam Long-Term Deformation Prediction Method Based on Empirical Mode Decomposition and LSTM Model Combined with Attention Mechanism

Yang Zhu, Yijun Gao, Zhenhao Wang, Guansen Cao, Renjie Wang, Song Lu, Wei Li, Wen Nie, Zhongrong ZhangWater 2022 · DOI

TL;DR

This paper introduces a novel method for predicting long-term deformation in tailings dams by combining Empirical Mode Decomposition (EMD) with an attention-enhanced Long Short-Term Memory (LSTM) network. This approach significantly improves prediction accuracy by effectively handling the complex, non-linear, and time-dependent nature of dam deformation, offering a more reliable early warning system for potential failures.

What Problem Does This Paper Address?

Tailings dams, crucial for storing industrial waste from mining operations, are inherently susceptible to deformation due to factors like rainfall, groundwater infiltration, and their own gravity. This deformation can lead to instability and, in severe cases, catastrophic failure, posing significant environmental and safety risks. Accurately predicting this long-term deformation is challenging due to the complex interplay of these factors and the non-linear, time-varying nature of the dam's behavior. Existing methods often struggle with capturing long-term dependencies, handling irregular changes, or effectively incorporating the influence of external factors with varying time lags.

Methods

The proposed method integrates several key techniques:

  • Empirical Mode Decomposition (EMD): EMD is used to decompose the complex, non-linear tailings dam deformation time series into a set of intrinsic mode functions (IMFs) and a residual term. This process breaks down the signal into simpler, more manageable components, separating trends from periodic oscillations and reducing the non-linearity and instability of the original data.
  • Lagged Autocorrelation Coefficient: This method is employed to determine the optimal input sequence length (lag order) for the LSTM model. It helps identify how far back in time external factors (like rainfall and phreatic line) influence the dam's deformation, addressing the "lagging effect."
  • Long Short-Term Memory (LSTM) Network: LSTMs are chosen for their ability to capture long-term dependencies in sequential data, overcoming the gradient vanishing/explosion problems common in traditional Recurrent Neural Networks (RNNs).
  • Attention Mechanism: An attention mechanism is integrated into the LSTM architecture. This allows the model to selectively focus on the most relevant input features and time steps, assigning different weights to different parts of the input sequence. This is crucial for overcoming LSTM's limitation of generating fixed-length vectors and for better capturing the influence of important historical data points, especially when dealing with long sequences.

The overall process involves decomposing the dam's deformation data using EMD, preprocessing the decomposed components and external factors (rainfall, phreatic line), determining the optimal lag order, and then feeding these processed sequences into the attention-enhanced LSTM model for prediction.

Datasets

The study utilizes data from the Zijin Mountain Gold-Copper Ore mine in Longyan City, Fujian Province, China. The dataset includes:

  • Tailings dam deformation data: Collected at a sampling frequency of 12 hours.
  • Rainfall data: Obtained from the China Meteorological Data Service Center (CMDC) for Shanghang County.
  • Phreatic line data: Also collected at a sampling frequency of 12 hours.

The data spans from September 3, 2019, to March 29, 2020, totaling 417 moments (approximately 208 days). To simulate a complete year and address missing data issues, the dataset was extended to represent a full year under various weather conditions. The data was split into a training set (first 178 days) and a test set (last 30 days) for evaluating long-term prediction capabilities.

Key Insights and Ideas

  • Decomposition is Key: EMD proves highly effective in breaking down the complex dam deformation signal into components that are easier for the LSTM to learn from, significantly improving prediction accuracy compared to using raw data.
  • Attention Matters for Context: The attention mechanism is vital for allowing the model to dynamically weigh the importance of different input factors (rainfall, phreatic line) and their historical influence, leading to more realistic predictions, especially during periods of rapid change.
  • Lagged Effects are Crucial: Explicitly accounting for the time lag between external factors (like rainfall) and their impact on dam deformation is essential for accurate long-term forecasting.
  • Multi-Directional Prediction: The model's ability to predict deformation in X, Y, and Z directions demonstrates its robustness and potential for comprehensive monitoring.

Key Results

The proposed EMD-Attention-LSTM model demonstrated superior performance compared to several baseline models:

  • Improved Accuracy: Compared to the EMD-LSTM model, the proposed method achieved a 23.66% improvement in RMSE and a 27.90% improvement in MAE on the 30-day prediction horizon.
  • Reduced Error: The model significantly reduced prediction errors compared to EMD-ARIMA and multiple regression models, with RMSE and MAE reduced by 4.616 and 6.574 respectively compared to multiple regression.
  • Realistic Predictions: The model showed better performance in capturing "hysteresis mutation" (sudden changes) in deformation, particularly on specific dates like March 7th and 16th, which is attributed to the attention mechanism's ability to leverage past rainfall and phreatic line data.
  • Generalizability: The model showed high prediction accuracy across X, Y, and Z directions, indicating its applicability for comprehensive tailings dam deformation monitoring.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Fixed Lag Order: The current study uses a fixed lag order for input sequences. Future work could explore dynamically adjusting this lag order based on historical data to potentially improve prediction effects.
  • Domain Specificity: The proposed method has only been validated for tailings dam displacement prediction. Applying and testing it on other types of dam deformation or geotechnical scenarios would be beneficial.
  • Computational Demands: Training LSTM models, especially with attention mechanisms, can be time-consuming and require significant computational resources. Future research could focus on reducing these requirements.
  • Data Integration: Integrating statistical methods with machine learning approaches could further enhance the model's capabilities.
  • Hyperparameter Optimization: Developing methods to reduce the reliance on large datasets for hyperparameter tuning would make the model more accessible.
  • Data Expansion: Collecting more diverse types of tailings dam monitoring data is suggested to further expand the model's application scenarios.

Summary updated

Abstract

Tailings dams are constructed as storage dams for ore waste, serving as industrial waste piles and for drainage. The dam is negatively affected by rainfall, infiltration lines and its own gravity, which can cause its instability to gradually increase, leading to dam deformation. To predict the irregular changes of tailings dam deformation, empirical mode decomposition (EMD) is applied to the deformation data to obtain the trend and periodic components. The attention mechanism is used to assign different weights to the input variables to overcome the limitation that the long short-term memory (LSTM) model can only generate fixed-length vectors. The lagged autocorrelation coefficient is applied to each decomposed subregion to solve the lagging effect of external factors on dam deformation. Finally, the model is used to predict deformation in multiple directions to test the generalization ability. The proposed method can effectively mitigate the problems of gradient disappearance and gradient explosion. The applied results show that, compared with the control model EMD-LSTM, the evaluation indexes RMSE and MAE improve 23.66% and 27.90%, respectively. The method also has a high prediction accuracy in the remaining directions of the tailings dam, which has a wide practical application effect and provides a new idea for tailings dam deformation mechanism research.
BibTeX
@inproceedings{zhu2022tailings,
  author = {Yang Zhu and Yijun Gao and Zhenhao Wang and Guansen Cao and Renjie Wang and Song Lu and Wei Li and Wen Nie and Zhongrong Zhang},
  title = {{A Tailings Dam Long-Term Deformation Prediction Method Based on Empirical Mode Decomposition and LSTM Model Combined with Attention Mechanism}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w14081229},
}
Figure