The Analysis of Time Series Forecasting Based on MLP Models

Abstract

The field of time series forecasting (TSF) increasingly leverages deep learning architectures. This study examines the latest advancements in Multi-layer Perceptron (MLP)-based models for TSF, focusing on the MLP Mixer, MLP Encoder-Decoder, and Frequency-Domain MLP models. Each model's principles are analyzed to identify commonalities and potential areas for improvement. These innovative approaches enhance the ability to capture features and global dependencies in complex and lengthy time series data through techniques such as dimension transformation, channel independence, and residual connections. This results in more accurate and stable predictions. The MLP Mixer alternates between temporal and feature dimensions; the MLP Encoder-Decoder emphasizes intensive information processing during encoding and decoding; and the Frequency-Domain MLP focuses on processing data from the time domain to the frequency domain. Experimental results reveal that the Frequency-Domain MLP model exhibits superior performance, though the choice of lookback window significantly affects results. Future research will aim to optimize these models further, exploring innovations in mixing, residual structures, and large-scale models. Enhancing the generalization ability and computational efficiency of these models will advance the field of TSF.