Integrated analysis of instrumentation data for structural health assessment and behavior prediction of arch dams

Abstract

In recent years, machine learning techniques have been available to predict and interpret the structural behavior of dams. Continuous monitoring of dam structure safety is vital in preventing possible damage. This study aims to predict the structural behavior by considering data collected for 13 years from instruments in the dam structure. Various machine learning methods are performed to account for the multi-non-linear relationships between dam displacement and the influential factors, thereby exploring the displacement laws of the dam. Three error metric indicators are employed for prediction, validation, and verification techniques to ensure the performance of models. Validation techniques include historical data validation, prediction validation, and the residual behavior over time. Predicting the structural behavior of the dam using the selected model requires data related to the input variables of the model. For this reason, the long short-term memory (LSTM) model, a robust algorithm for predicting time series variables, was used to predict the input variables. LSTM model provided acceptable predictions of changes in the input variables for these years. Additionally, the Boosted Regression Trees model, selected as the most accurate in the evaluation process, was employed to predict the structural behavior of the dam for periods not yet experienced by the dam, using these input variables. The predicted behavior of the dam demonstrated a strong ability to interpret the health of the dam structure and prevent possible damages. The effectiveness of the LSTM model was confirmed as a promising method in predicting time series input variables for ML models to predict dam displacements in the next years.