RhyMix: Lightweight AI Forecasts Long-Term Time Series with Adaptive Rhythms

Real-world time series data often exhibits intricate dynamics, including short-term fluctuations, seasonal cycles, long-term trends, and sudden changes. Many existing forecasting models struggle to capture this full spectrum of temporal patterns efficiently. This research introduces RhyMix (RHYthm MIXture), a novel hybrid neural architecture specifically designed for long-term time series forecasting that addresses these limitations. RhyMix employs a parallel dual-path modeling paradigm with adaptive gating. It features a "Cyclic Path" that incorporates explicit seasonal inductive bias through learnable cyclic embeddings, effectively capturing predictable rhythmic patterns. Complementing this is a "Multi-Scale Temporal Convolutional Network with Channel Attention Path," which uses multi-scale depthwise dilated convolutions to capture temporal dependencies across various receptive fields. A key innovation is the adaptive gating mechanism at multiple levels. A path gate dynamically combines four specialized forecasting heads per sample and channel, while a hybrid gate adaptively balances the contributions of the Cyclic and MSTCN-CA Paths based on input characteristics. This design allows RhyMix to adapt to specific temporal patterns while maintaining linear complexity in sequence length, channels, and prediction horizon. Benchmarking across 12 real-world datasets shows RhyMix achieving state-of-the-art performance on 10, all while remaining lightweight and suitable for real-time deployment on resource-constrained devices.