MLP-Transformers: Multi-Layer Perceptron coordinated Transformers for Time Series Forecasting

This paper proposes MLP-Transformers, which combines multi-layer perceptron (MLP) and Transformer models to automatically divide time series data into batches and call the corresponding Transformer model for prediction. This method trains MLP and Transformer simultaneously during the training phase. First, a corresponding number of Transformer models and MLP models are created according to the number of categories of time series data, and then the training set is input into the MLP model for classification. Next, the most appropriate Transformer model is selected for training based on each batch of data to enhance the generalization ability of the model. During the prediction phase, the test set is fed into the trained MLP-Transformer model. After MLP classification, the corresponding Transformer model is used for prediction. Performances of MLP-Transformers are thoroughly evaluated using IEEE PHM 2012 dataset provided by FEMTO-ST Institute and jet engine lubrication oil dataset provided by airlines respectively. Experimental result shows that MLP-Transformers outperforms existing methods for dataset, demonstrating the effectiveness and the wide range applicability of the proposed method in solving time series problems.