Research on Weather Process Similarity Based on 3D CNN Autoencoder

This article proposes a weather process similarity and retrieval algorithm based on a 3D CNN Autoencoder. It uses 3D CNN to extract features of weather elements in multi-dimensional space. The encoder outputs a one-dimensional vector of size 512. The cosine similarity between two weather processes is then calculated based on the two one-dimensional vectors, representing their similarity. Experimental data demonstrates the effectiveness of this algorithm. By searching historical weather processes based on the forecast date and referring to the evolution of historical weather processes, it provides positive guidance for current weather forecasts.