Multimodal Multiobjective Neural Architecture Search for Lightweight and Failure-Resilient Time Series Forecasting

Time series forecasting (TSF) plays a pivotal role in decision-making and risk mitigation by predicting future values from historical observations. Despite significant improvements in forecasting accuracy, existing TSF models face two critical challenges: increasing computational complexity hinders deployment on resource-constrained platforms, and vulnerability to random failures compromises prediction stability. To address these issues, we propose M3NAS-TSF, a multimodal multiobjective neural architecture search framework that automates the design of lightweight and failure-resilient TSF models. Specifically, to address computational complexity, we introduce a modular time series super-net that defines a flexible and compact search space, enabling the discovery of architectures with reduced model size. To enhance failure resilience, we develop a multimodal multiobjective neural architecture search algorithm that promotes architectural diversity through an architecture-aware niching strategy and an architectural diversity distance metric, ensuring a broad set of robust candidate models. We also propose a node distribution heatmap and a structural entropy index to assess architectural diversity without ground-truth Pareto sets. Extensive experiments on 40 forecasting cases demonstrate that M3NAS-TSF outperforms six representative baselines, achieving superior forecasting accuracy (the maximum reduction in the RMSE reached 14.50%) with a smaller model size while maintaining greater architectural diversity for failure resilience.