Dynamic TMoE: A Drift-Aware Dynamic Mixture of Experts Framework for Non-Stationary Time Series Forecasting

Non-stationary time series forecasting is challenged by evolving distribution shifts that static models struggle to capture. While Mixture-of-Experts (MoE) architectures offer a promising paradigm for decoupling complex drift patterns, existing approaches are limited by fixed expert pools and memoryless routing, hampering their ability to adapt to abrupt regime shifts. To address this, we propose Dynamic TMoE, a framework that unifies architectural evolution with temporal continuity during learning phase. By detecting distribution shifts via Maximum Mean Discrepancy (MMD), we dynamically instantiate heterogeneous experts and prune redundant ones to optimize capacity. Additionally, a temporal memory router leverages recurrent states and an anomaly repository to ensure stable, context-aware expert selection without requiring test-time updates. Experiments on nine benchmarks demonstrate state-of-the-art performance, reducing MSE by 10.4\% and MAE by 7.8\%. Code is available at https://github.com/andone-07/Dynamic-TMoE.

Forecasting how things will change over time, such as tomorrow's electricity demand, next month's flu cases, and next week's traffic, is essential but hard because the real world keeps shifting. A model trained on calm weather may stumble during a storm. One tuned before a financial shock can break afterward. Existing AI forecasters typically rely on a fixed team of "specialists," each focused on certain patterns. The team never grows or shrinks and has no memory of past decisions, so when conditions change abruptly, performance drops. We built Dynamic TMoE, a flexible forecasting framework that automatically detects when data patterns change. When one occurs, it diagnoses what new pattern has appeared and recruits a matching specialist. Specialists that go unused are quietly retired. A memory component records past decisions, so when a familiar disruption returns, the team responds consistently rather than erratically. Across nine real-world datasets spanning energy, healthcare, traffic, finance, and weather, Dynamic TMoE produces more accurate forecasts than leading alternatives. The approach offers a blueprint for AI systems that adapt to a changing world rather than freezing when conditions shift.