Abstract: As operating years increase, the crack opening displacement of concrete dams serves as a critical indicator for evaluating structural safety, yet its evolution mechanism exhibits high nonlinearity and time-varying characteristics governed by complex environmental factors, such as fluctuating reservoir water levels, seasonal temperature gradients, and progressive material aging dynamics. Efficiently extracting, decoupling, and utilizing long-term monitoring data under multivariate factor interference has thus become an urgent challenge in the field of dam safety monitoring and structural health diagnosis. Traditional statistical models and conventional machine learning approaches often suffer from significant limitations, including feature redundancy, gradient degradation over long sequence horizons, and an inability to dynamically emphasize critical environmental triggers, ultimately leading to a bottleneck in fitting performance and forecasting accuracy. To overcome these structural limitations and significantly improve both prediction accuracy and robustness, this study proposes an end-to-end concrete dam crack opening prediction model based on optimized feature identification by systematically integrating a dual-channel dot-product attention mechanism, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. The methodology first introduces a dual-channel dot-product attention mechanism operating across spatial and temporal feature dimensions, which dynamically computes similarity scores between input vectors to adaptively assign optimal weights to key environmental variables—such as hydraulic pressure head, internal spatial temperature distributions, and time-dependent aging components—thereby enhancing crucial variable representations while effectively suppressing high-frequency background noise. Subsequently, the CNN subnetwork utilizes multi-scale spatial-temporal convolution kernels to automatically extract high-level structural features and local spatial correlations from heterogeneous multi-point monitoring sequences, while the gated recurrent structure of the LSTM subnetwork captures complex long-term temporal dependencies without suffering from vanishing gradients across multi-year observation spans. By executing an adaptive weighted learning paradigm across multivariate environmental information, the proposed architecture establishes a unified feature optimization, spatial-temporal fusion, and sequence forecasting pipeline. An extensive engineering case study was conducted using real-world long-term prototype monitoring datasets from an operating concrete dam to rigorously evaluate the proposed framework. The empirical results demonstrate that the model accurately captures and identifies the intricate time-frequency nonlinear features, cyclic hysteresis behaviors, and localized abrupt changes caused by sudden operational or thermal events within the monitoring signals. Quantitative evaluations across multiple statistical metrics—including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R²)—confirm that the proposed model consistently outperforms benchmark models, including Multiple Linear Regression (MLR), Support Vector Regression (SVR), Random Forest (RF), and standard unmodified CNN-LSTM baselines, achieving substantial reductions in prediction error and demonstrating exceptional stability over multi-step ahead forecasting horizons. The engineering case study results show that, compared with the combined convolutional neural network and long short-term memory network model, the proposed model improves the coefficient of determination by 7.9% in the fitting stage and 15.0% in the prediction stage. In particular, the coefficient of determination in the prediction stage is 58.6% higher than that of the traditional multiple linear regression model. In conclusion, this research successfully resolves the accuracy bottleneck caused by environmental noise and factor coupling, substantially enhancing the precision, sensitivity, and reliability of crack evolution trend predictions, and offering practical engineering application value and theoretical references for safety early warning, health diagnosis, and intelligent operation and maintenance of concrete dam infrastructure.