During the long-term service of a building structure, various external and internal factors can lead to the failure of key components. The structural strain response serves as an intuitive indicator of the operational status of these components. Therefore, accurately predicting the structural strain response is crucial for the long-term monitoring of building structures. In this paper, we propose a CNN-LSTM-based method for predicting structural strain responses, enhanced through feature selection, and validate it using measured data from the health monitoring system of the Pearl Tower in Northern Xinjiang. The input model’s features are determined through Pearson correlation analysis, and their importance is ranked and selected using the random forest algorithm. We developed a CNN-LSTM structural strain response prediction model and applied the Bayesian optimization algorithm to fine-tune the model’s hyperparameters, thereby enhancing its prediction accuracy. We then compared the prediction results with those obtained from conventional LSTM and BiLSTM models. The results demonstrate that the correlation coefficients between the predicted and actual values of the CNN-LSTM model at all strain measurement points exceed 0.99. This indicates the model’s superior performance in predicting the strain response of components, as well as its higher prediction accuracy compared to the LSTM and BiLSTM models.