Abstract: China has planned and built many pumped-storage hydropower stations which are characterized by rapid and large-amplitude water-level fluctuations. Developing dam deformation factor models and prediction methods that can describe these operating characteristics is crucial for dam safety assessment. Based on traditional models, this study develops an improved deformation factor model that is equipped with an additional factor representing the water-level change rate, to address the limitation of conventional factor models caused by lack of this rate factor. To mitigate redundancy in high-dimensional factors, we work out a new hybrid framework that is used to integrate kernel principal component analysis and a deep autoregressive neural network, thereby enhancing predictive performance through jointly applying factor dimensionality reduction and deep-learning-based modeling. Application to a case study of the dam for a pumped-storage station shows this method achieves an accuracy significantly higher than that of the conventional models by adding the water-level change rate factor. And, it reduces the mean absolute error by roughly 36.27% on average relative to the benchmark models, demonstrating a significant improvement in model prediction, as a new approach to dam deformation analysis and safety evaluation.

Key words: pumped-storage power station, dam deformation, deep autoregressive neural network, kernel principal component analysis