Abstract: Accurate forecasting of the evolution of deformation patterns of a concrete dam is critical to ensuring its safe operation and facilitating long-term maintenance. However, conventional prediction models often suffer from insufficient accuracy and poor interpretability in dealing with nonlinear, multi-dimensional, and long-term deformation time series data. In this work, the Convolutional Neural Network (CNN) is adopted to extract local spatial features from deformation data; the Inverted Transformer (iTransformer) with its multi-head attention mechanism are used to capture long-time series dependencies. And, we use Long Short-Term Memory (LSTM) to enhance the memory of time series data, and optimize globally the key model hyperparameters using the Particle Swarm Optimization algorithm (PSO). Then, we develop an interpretable hybrid model integrating PSO, CNN, iTransformer, and LSTM, and conduct global sensitivity analysis to explain the model’s results. We compare and verify multiple gauge points and multiple models in a case study involving 11 years of displacement measurements on the Xiluodu large concrete arch dam, with model performance assessed via the root mean square error (RMSE) metric. The results show our new model achieves more accurate predictions of the dam deformation trend, featured with higher accuracy, stronger generalization ability, and better stability, and good interpretability.

Key words: concrete dam, deformation prediction, convolutional neural network, iTransformer, long short-term memory network, particle swarm optimization, interpretability