Most of the previous studies on dam seepage pressure prediction did not distinguish the differences in the contributions of influencing factors to different characteristic components of seepage pressure

thereby reducing the interpretability of their models. Most of the previous prediction models in literature adopted a single algorithm and thus suffered such problem as a difficulty in exploring accurately the patterns of seepage pressure sequences with highly nonlinear

non-stationary mixed characteristics. This paper develops an interpretable hybrid prediction model for seepage pressure based on seasonal and trend decomposition using loess (STL) and an integrated learning strategy. First

the model adopts STL to decompose the original seepage pressure time series into seasonal

trend and remainder components

so as to avoid the deficiency of pattern confusion encountered in the previous predictions. Then

it identifies the variability characteristics of different components using an ensemble learning model that consists of a kernel extreme learning machine (KELM) optimized by the multi-strategy-improved sparrow search algorithm (MSISSA) and a convolutional neural network combined with a gated recurrent unit (CNN-GRU). In addition

the One-Factor-At-A-Time (OFAT) method is used to analyze the contributions of influencing factors to the different characteristic components of seepage pressure so as to change the weights of input factors and successfully increase model interpretability. Case study shows that while ensuring model interpretability

this new hybrid model improves prediction accuracy by an average of 48.44% compared to a single algorithm-based model and an average of 11.42% compared to other ensemble prediction models. This verifies the model and provides a new modelling approach for dam seepage safety monitoring.