Multi-Scale Patch Embedding with Transformer and LSTM for Adversarial Learning

Time series prediction has important application value in multiple fields, but due to the complexity and uncertainty of time series data, accurate prediction still faces many challenges. This article proposes a new collaborative adversarial learning framework that combines multi-scale patch embedding methods, transformers, and Long Short Term Memory (LSTM) networks to fully leverage the complementary advantages of these two models. The transformer model can capture feature patterns of time series at different time scales through multi-scale patch embedding, while LSTM can effectively simulate the long-term dependencies of sequences. We have designed a collaborative mechanism that enables two models to learn and share knowledge with each other, while introducing adversarial training strategies to improve the model's generalization ability and prediction stability. Experiments on ETTh1 and other public time series data sets show that the proposed collaborative confrontation learning framework significantly improves the prediction accuracy compared with a single model and performs well in processing complex time series data. Especially in long-term prediction tasks, Our method surpasses some existing methods This study provides a new and effective paradigm for time series prediction, which has important practical significance for multiple fields such as industrial production, energy management, and financial analysis.