An Enhanced Multivariate Long-Term Time Series Robust Forecasting Model with Distinguishable Variable Identifier

Recently, in the field of multivariate long-term series prediction, Linear models have benefited from the training strategy of Channel Independence, outperforming most previous Transformer-based models and questioning the ability of the attention mechanism to capture time dependence. To regain the Transformer's dominance, we have further improved the Patch-based Transformer model, but some articles point out that the model designed based on Channel Independence has the defect of spatial indistinguishability. In order to solve this problem, our Variable Identifier, embedding corresponding to a specific variable, is specially designed for the Transformer model to give it the ability to distinguish among different variables. The Variable Identifier we have learned can also roughly describe the similarity between variables. In addition, to enhance the model's resistance to data noise interference and prevent overfitting, we incorporate a novel loss function that combines the advantages of both Mean Squared Error and Mean Absolute Error loss functions. In order to further improve the model's performance but not add extra computation, we try to design a multi-scale CNN structure inspired by computer vision. Finally, our model outperforms recently proposed linear models and previous Transformer-based models, and our extensive experiments on widely used open-source datasets show a maximum 4.4% relative improvement over state-of-the-art methods.