DCU: Unidimensional Data Prediction with Graphical Model

In view of the limited number of available models for one-dimensional data prediction tasks at present and in the past, and the underlying problems of RNN-like networks in most general models, we propose a new data prediction method. This method utilizes a two-dimensional model with a larger sampling area and more extensive branches to carry out the prediction task of one-dimensional time series data from the image level. We integrate one-dimensional data according to the time nodes and transform it into two-dimensional multi-channel images. We enhance the traditional encoding-decoding network structure by replacing part of the convolutional layers with Tokenized MLP. This modification ensures more accurate weight calculation of the model on a specific coordinate axis. Additionally, we improve the training mode of the traditional network architecture, and add a custom penalty mechanism on top of the loss function. This mechanism directly impacts the weight of the model nodes beyond the learning process, enhancing the controllability of the network training process. Through comparative experiments, our method demonstrates an $8.4 \%$ improvement in MAE, a $7.8 \%$ improvement in 12 fitting, and a $51.3 \%$ improvement in test time compared with the traditional methods.