RhyMix: A Lightweight Adaptive Multi-Rhythm Network for Long-Term Time Series Forecasting

Traditional time series forecasting models often struggle to balance the need for capturing long-range dependencies, local temporal variations, and periodic cycles simultaneously. Existing architectures typically rely on a single modeling paradigm—such as transformers, convolutions, or linear models—which often leads to a trade-off between computational efficiency and the ability to model complex, multi-scale dynamics. The authors seek to develop a hybrid, lightweight architecture that can adapt to diverse temporal patterns while maintaining linear computational complexity.

To address these challenges, the authors propose RhyMix (Rhythm Mixture), a dual-path neural network. The architecture consists of two primary branches:

1. Cyclic Path: Uses learnable cyclic embeddings to explicitly capture predictable seasonal and rhythmic patterns.
2. MSTCN-CA Path: A Multi-Scale Temporal Convolutional Network with Channel Attention that uses depthwise dilated convolutions to capture temporal dependencies across varying receptive fields.

A central innovation is the use of adaptive gating mechanisms at multiple levels. These gates use statistical features (mean, standard deviation, slope, etc.) extracted from the input to dynamically weight the contributions of four specialized forecasting heads (Direct, Trend-Seasonal, Local Convolution, and Periodic Fusion) and to balance the two main encoding branches on a per-sample basis.

Across 12 real-world benchmark datasets, RhyMix achieved state-of-the-art performance on 10, outperforming several established baselines including PatchTST, iTransformer, and Time-o1. The model is exceptionally lightweight, containing only ~40,000 parameters, which allows for a small memory footprint (157 KB) and low-latency inference (<5 ms). This efficiency makes the model highly suitable for deployment on resource-constrained edge devices where real-time forecasting is required.