Huawei and ECNU Top International Time-Series Benchmarks with Foundation Model
A joint research team from Huawei and East China Normal University developed an Ascend-native time-series pre-trained foundation model named Qiyao, which topped the TIME-leaderboard and GIFT-Eval benchmarks. This achievement places domestic time-series foundation models into the global front ranks, marking a shift from following to catching up in AI research for crucial industries like energy and meteorology. The model supports zero-shot time-series forecasting across scenarios like power grid load analysis and industrial anomaly detection, while also enabling low-cost secondary fine-tuning for custom vertical industry needs.
## BACKGROUND
Time-series foundation models are a new paradigm in machine learning that break away from traditional single-dataset models by using massive pre-training to handle arbitrary new sequences. Benchmarks like TIME-leaderboard and GIFT-Eval evaluate the generalizability and predictive accuracy of these models across diverse domains.