高维时序预测的ImageNet时刻！首个高维时序预测基准发布，模型领跑多数据集SOTA

Core Insights - The article discusses the introduction of Time-HD, the first large-scale benchmark specifically designed for high-dimensional time series forecasting, addressing the limitations of existing models in handling high-dimensional data [2][11][42] Group 1: High-Dimensional Time Series Forecasting - The transition to high-dimensional time series data is evident across various fields, including finance and smart city traffic networks, indicating a shift towards complex systems with thousands of variables [6][12] - Current mainstream time series forecasting models are primarily focused on low-dimensional datasets, which limits their efficiency and performance in high-dimensional contexts [7][8] - Time-HD includes 16 datasets with variable counts ranging from 1,161 to 20,000, significantly surpassing traditional benchmarks that typically contain only 7 to 862 channels [12][14] Group 2: Features of Time-HD - Time-HD encompasses diverse sources, including both simulated and real-world datasets, enhancing its applicability for evaluating model generalization in practical scenarios [14] - The benchmark offers datasets of varying scales, with four large-scale (GB-level), eight medium-scale (hundreds of MB), and four small-scale (tens of MB) datasets, facilitating resource-efficient model evaluation [16] - It covers multiple sampling frequencies, reflecting real-world applications, and employs corresponding prediction lengths rather than fixed steps, aligning with actual forecasting needs [17][18] Group 3: U-Cast Model - The U-Cast architecture is introduced to tackle challenges posed by the surge in variables, utilizing a hierarchical latent query network to efficiently extract and compress key information from high-dimensional data [22] - U-Cast demonstrates a 15% reduction in mean squared error (MSE) across multiple datasets compared to existing models, while also achieving faster training speeds and lower memory usage [36][37] - The model incorporates full-rank regularization to mitigate redundancy in high-dimensional time series, promoting the learning of independent and structured feature representations [30][41] Group 4: Impact and Future Directions - The release of Time-HD and the open-source Time-HD-Lib framework, along with the U-Cast method, sets a new benchmark for high-dimensional time series forecasting, providing a robust baseline for future research [42][43] - The advancements in high-dimensional time series forecasting are expected to spur a new wave of innovation, paving the way for more extensive and realistic forecasting applications [44]