A Multi-Task Regression Model for Seismic Key Parameters Prediction Based on Transformer and ResNet Parallel Architecture Considering Fusion Mechanisms

Since the development of smart cities and sustainability rely on a stable social environment, effective emergency methods have been seen as important stabilizing means, especially for earthquakes. Seismic parameter prediction is one of the relevant core research contents, including traditional and deep learning methods. However, traditional methods have many limitations and cannot effectively adapt to the complex seismic environment or understand the nonlinear effects deeply. In this paper, the author proposed a multitask regression model based on the parallel architecture of Transformer and ResNet considering the fusion mechanism, which aims to predict key seismic parameters. It could efficiently extract the features of three-phase seismic waves by considering both local and long-term dependence. After that, the experimental results showed that the new model outperformed the other four models— fusionless version, Transformer only, ResNet only, and CNN-BiLSTM —with a Loss of 0.106, MAE of 0.346, RMSE of 0.445 and R² of 0.708. Furthermore, the author found that the configuration with a learning rate of 0.0003, a batch size of 64, and a maximum number of iterations of 60 achieved the most incredible balance between model effectiveness and training stability. Compared to others, it has the best prediction accuracy, error control, and goodness-of-fit under the same architecture and data conditions. The stability and effectiveness of the optimal model were further validated by observing the training loss, validation loss, MAE, RMSE, R², and dynamic weights throughout the training phase. Additionally, the effectiveness of the cosine annealing strategy was evident from the learning rate curve. At the same time, the steady variation in computational complexity, inferred from the epoch time, demonstrated the model’s suitability for practical training and resource management. Finally, the author concluded the study with a summary and proposed directions for future research.