M2Patch Improves Multivariate Time Series Forecasting with Structured Latent Space

A new CNN-based forecasting architecture, M2Patch, has been developed to enhance multivariate time series prediction by explicitly modeling the structural patterns that unfold across multiple temporal scales. Unlike many existing methods that treat learned representations as transient, M2Patch maps observations into a structured latent space using two complementary differentiable constraints. The architecture employs multi-scale patching to decompose input into overlapping temporal granularities, with depthwise separable convolutions extracting scale-specific features efficiently. These features are then compressed into a compact latent representation. The latent space is organized by an intra-scale smoothness constraint, ensuring temporal continuity, and an inter-scale alignment constraint, which restores cross-granularity interaction. This design ensures consistent representations across scales. Experiments on ten real-world benchmarks show M2Patch achieving superior or competitive results in most settings, demonstrating its effectiveness and linear computational complexity.