Analog forecasting is a widely adopted statistical method in operational meteorological services. Traditional single-layer similarity approaches have such limitations as the lack of three-dimensional spatial information, the unstable performance of single similarity criteria, and the frequent interference from synoptic system pattern and intensity (magnitude). To address these challenges and explore the feasibility of deep learning models in synoptic situation recognition and forecasting, in this study we develop a novel approach using the ECMWF fifth-generation reanalysis (ERA5) dataset. We construct a deep learning architecture that integrates convolutional neural networks (CNN) with Transformer modules, incorporating self-attention mechanisms. Verification shows that this model can effectively capture three-dimensional spatial features of synoptic situation. Then, utilizing the extracted feature vectors, we design a comprehensive similarity framework that combines three complementary metrics: Pearson correlation (empha-sizing pattern shape), Euclidean distance (emphasizing magnitude), and Chebyshev distance (considering both shape and magnitude). This integration forms our proposed method: Synoptic Similarity Net. Finally, the operational application effect of this method is tested and evaluated in detail. The results indicate that this method can achieve the highest average structural similarity index (SSIM) and lowest mean squared error (MSE) relative to the traditional methods, demonstrating significant improvements in both metrics. Case studies across seasons confirm that the historical analogs identified by Synoptic Similarity Net exhibit both greater numerical accuracy and superior spatial pattern consistency compared to the original synoptic fields. These results demonstrate the promising potential of this method for meteorological operational applications.