Microseismic monitoring technology is one of the most critical technologies used in hydraulic fracturing. The positioning accuracy and efficiency of microseismic sources significantly influence the performance of this technology. This study proposes a microseismic source localization method based on convolutional neural networks that transforms the inversion problem of solving source positions into a mapping problem of constructing the probability distribution of source positions from microseismic data. First, when constructing the dataset, factors affecting the accuracy of source positioning are considered, and velocity model errors and noise interference are introduced for data augmentation. Second, a convolutional neural network model, termed MEL-Net, is developed, which is based on the classic U-Net network architecture and integrates an attention mechanism and a spatial hole multiscale pooling module to improve feature extraction. Furthermore, no feature concatenation operation is performed in shallow-level encoding and decoding to reduce the interference of irrelevant information with positioning tasks. Finally, the applicability of the method is verified using a simple layered velocity model and a complex Marmousi model. The results show that MEL-Net can achieve accurate source location predictions. In the measurement process, it is more robust than U-Net. Compared with the traditional reverse-time positioning algorithm, it is insensitive to factors such as speed model errors and noise interference. It significantly improves the positioning speed while providing accurate microseismic source location predictions.