Abstract: Concrete dams serve as critical hydraulic infrastructure, whose safe operation is directly linked to the national economy and people's livelihood. Deformation prediction plays a vital role in dam safety monitoring and management, providing essential guidance for early warning systems and maintenance decisions. However, traditional global modeling strategies for concrete dam deformation prediction face significant challenges in adapting to the time-varying characteristics inherent in the complex deformation process, where environmental conditions, material properties, and operational states continuously evolve. These conventional approaches typically establish a single static model based on historical data, which becomes increasingly inadequate as the dam's behavior changes due to aging, seasonal variations, and fluctuating operational conditions. To address these limitations, this study proposes an innovative adaptive prediction method for concrete dam deformation based on improved just-in-time learning (IJITL) and weighted extreme learning machine (WELM). The proposed methodology fundamentally differs from traditional global modeling approaches by implementing a dynamic, locally adaptive framework that continuously evolves with changing deformation patterns. The methodology involves several key innovations. First, a comprehensive feature correlation analysis framework is established using both the Pearson correlation coefficient and mutual information to quantify linear and nonlinear relationships between input features and deformation responses. This dual-metric approach ensures that both simple linear dependencies and complex nonlinear interactions are properly captured and weighted. The Pearson correlation coefficient effectively measures linear relationships, while mutual information, derived from information theory, captures nonlinear dependencies that traditional correlation measures might overlook. These correlation measures are then normalized and combined to form a comprehensive correlation index, which serves as the basis for feature importance weighting. Second, the traditional just-in-time learning framework is significantly enhanced by integrating feature importance weights into the sample similarity calculation. This improvement addresses a key limitation of conventional JITL methods, which treat all input features equally when selecting relevant historical samples. By incorporating feature importance derived from correlation analysis, the improved similarity measure focuses on the most influential features for deformation prediction, thereby selecting more representative and relevant historical samples for local model construction. Third, a weighted extreme learning machine is proposed as the local prediction model within the JITL framework, replacing the conventional partial least squares regression commonly used in traditional JITL applications. While PLS captures only linear relationships, WELM models complex nonlinear mappings between input factors and dam deformation. Furthermore, WELM incorporates sample weights based on similarity measures, ensuring that samples more similar to the target prediction point contribute more significantly to local model training. The proposed method was comprehensively validated using real monitoring data from the Jinping I arch dam, the highest concrete arch dam in the world. The validation dataset spans from July 31, 2016, to December 31, 2018, encompassing diverse operational conditions and seasonal variations. Experimental results demonstrate exceptional performance in long-term prediction scenarios, where traditional global models typically degrade due to their inability to adapt to evolving deformation patterns. Comparative analysis against four representative global modeling methods—multilayer perceptron (MLP), support vector regression (SVR), extreme gradient boosting (XGBoost), and long short-term memory neural networks (LSTM)—reveals substantial improvements. The proposed IJITL-WELM method achieved root mean square error reductions of 94.4%, 94.0%, 91.3%, and 81.4% compared to MLP, SVR, XGBoost, and LSTM, respectively. It maintains prediction accuracy even over extended time horizons, effectively addressing the performance degradation issue that affects global models through dynamic database updates and adaptive local modeling capabilities. Ablation experiments systematically evaluate the individual contributions of the two key improvements. Results confirm that feature importance weighting significantly enhances sample selection quality, while the weighted extreme learning machine substantially improves nonlinear mapping capabilities compared to traditional linear models. The universality analysis across multiple monitoring points further validates the method’s generalizability and robustness.