Abstract: This study develops a collaborative multi point ensemble forecasting framework to address the limitations of conventional dam deformation prediction models—relying on isolated single point data and thereby failing to simultaneously account for multi source environmental factors and spatial synergistic effects. First, we use a correlation variation-based CA KMeans clustering algorithm to partition the dam’s monitoring stations into several sub clusters by their temporal deformation characteristics and spatial positions, so as to enhance inter point synergy. And, we construct a high dimensional spatiotemporal feature matrix (HST M) by integrating multi source influencing factors－such as water pressure, temperature, time dependent effects, and spatial coordinates－to characterize the dam’s overall deformation behavior comprehensively. Then, an autoencoder is adopted to perform dimensionality reduction and feature refinement on these high dimensional inputs, automatically extracting critical nonlinear correlations while suppressing redundancy. Finally, ridge regression serves as the predictor, leveraging its L2 regularization to deliver stable, well generalized deformation forecasts even under high dimensional, multicollinear conditions. Case studies demonstrate our new framework not only enhances predictive accuracy and robustness but offers high applicability and computational efficiency.

Key words: dam, deformation prediction, CA-Kmeans, high-dimensional spatiotemporal features, autoencoder, ridge regression