Patch-MLP-TS: A Channel-Independent, Attention-Free Architecture for Long-Term Time Series Forecasting

Patch-MLP-TS is a channel-independent, attention-free architecture for long-term time series forecasting. The work investigates whether patch-based representations can be combined with MLP-based token and feature mixing without relying on attention. During development, a failure mode was identified in a naive patch-plus-MLP design where the absence of positional information and cross-patch interaction caused temporal order information to be lost. The proposed approach addresses this through positional information and cross-patch mixing and is evaluated on the ETT and Electricity benchmarks against linear and attention-based forecasting models. The study includes multi-seed benchmarking, ablation studies, patch-length analysis, and measured training and inference efficiency.