Author(s): Huang Minshui; Deng Zhihang; Zhang Jianwei
Linked Author(s):
Keywords: Concrete dam displacement prediction; Variational Mode Decomposition; Long Short-Term Memory neural network; ARIMA model; Hybrid prediction model
Abstract: Research on traditional displacement prediction of concrete dam mostly relies on single models, which often struggle to comprehensively capture the nonlinear characteristics of dam displacement data, resulting in poor prediction accuracy and generalization capability. Additionally, commonly used signal decomposition methods, such as wavelet transform and empirical mode decomposition, suffer from serious mode mixing issues, making it challenging to effectively extract signal features. Therefore, this study aims to address the impact of mode mixing on signal accuracy by employing Variational Mode Decomposition (VMD) for data decomposition and reconstruction. This is combined with Long Short-Term Memory (LSTM) neural network and Autoregressive Integrated Moving Average (ARIMA) models to enhance the accuracy and robustness of dam displacement prediction models. Initially, VMD is utilized for preprocessing the original displacement data to extract high-frequency periodic components, high-frequency random components, and low-frequency trend components. Subsequently, the fused comprehensive high-frequency sequence derived from the VMD decomposition is input into the LSTM for modeling and prediction, while the ARIMA model is employed for predicting the low-frequency trend component. The prediction results from the two models are then combined with weighted averaging, and the weights are iteratively adjusted to obtain the best dam displacement prediction results, ensuring accurate prediction of the dam displacement time series. Finally, through engineering examples, the proposed VMD-LSTM-ARIMA model demonstrates superior prediction accuracy and stability compared to traditional single prediction models, indicating that this hybrid prediction model has excellent fitting and predictive capabilities and holds significant potential for practical engineering applications.
Year: 2024