Abstract: Existing studies primarily employ neural network models to predict single outcomes such as ultimate failure load, which cannot fully reflect the service performance of strengthened beams. Furthermore, refined finite element modeling is highly sensitive to material constitutive and interface parameters, while theoretical models suffer from excessive simplification and limited generalizability. To address these challenges, this paper proposes a hybrid model combining a convolutional neural network (CNN) and a long short-term memory (LSTM) network to predict the full-range load–deflection curves of fiber-reinforced polymer (FRP) near-surface mounted (NSM) reinforced concrete beams. The model integrates multi-source input features (both numerical and categorical data) and optimizes feature fusion using domain knowledge from civil engineering, thereby improving prediction accuracy and computational efficiency. The model is trained and tested on 134 sets of experimental data, and evaluated using R², RMSE, MAE, and RPD. Results show that the CNN-LSTM model achieves R² values of 0.97 and 0.92 on the training and testing sets, with RMSE of 4.22 and 8.75, MAE of 2.89 and 5.21, and RPD values all exceeding 2. The overall prediction error is controlled within 30%, and the model accurately captures the three-stage characteristics of the load–deflection curve. Furthermore, Pearson correlation analysis and feature sensitivity analysis reveal the influence of 15 key design parameters on the ultimate load-bearing capacity, leading to proposed material–configuration compatibility optimization strategies. This study provides a high-precision and generalizable data-driven method for predicting the flexural performance and guiding the engineering design of FRP NSM-strengthened beams.