Numerical simulation is the most common method to predict reservoir production temperatures during geothermal energy extraction. Considering the principle of numerical modeling, the numerical simulation establishment process requires a large amount of good exploration data. In addition, it is heavily influenced by subsurface heterogeneity. Also, despite the superior performance of deep learning models, sparse data is a critical challenge in the training process. Therefore, we propose a one-dimensional-convolutional neural network (1D-CNN) model and use data augmentation techniques to build a large-scale multiscale production temperature data set. The network learns the nonlinear relationship between boundary conditions and production temperature from the data set and reaches the production temperature prediction for a three-well geothermal system. The maximum difference in production temperature is 1.8181 °C and the generalization performance is improved by 59.6%. It is worth noting that the excellent generalization capability indicates that the data-driven concept behind the model is an easily interpretable one. As a new data processing concept, the “data-guided approach” is a key step in establishing a universal approach for application in the geothermal field.