REMIND: Processed Spatiotemporal Forecasting Datasets for Traffic, Energy, and Hydrology
Abstract
This dataset collection provides processed data used in REMIND experiments for spatiotemporal graph forecasting across traffic, energy, and hydrology domains. When to Trust Memory: Retrieval-Guided Probabilistic Spatiotemporal Forecasting under Distribution Shift. (NeurIPS 2026) The collection contains four dataset variants: ENERGY: Four time-series segments with 103 nodes, derived from the Energy-Stream benchmark described in Expand and Compress. CAMELS: A hydrological subset containing 12,699 daily observations across 7 nodes. PEMS: Traffic-flow observations from 2015–2017 across 581 nodes. PEMS-7years: Traffic-flow observations from seven annual periods spanning 2011–2017, using a consistent set of 581 nodes. Each annual file contains a subset of the year rather than full-year coverage. Each dataset includes time-series arrays, an adjacency matrix, precomputed training/validation/test splits, and normalization parameters. The seven-year PEMS variant additionally includes source-column indices. These files support both reuse of the provided splits and regeneration of model inputs from the included time-series arrays. The provided splits use 12-step input and prediction windows, a sequential 60%/20%/20% split, and normalization fitted on training histories. Retrieval caches, model checkpoints, and prediction outputs are not included. This collection contains processed experimental subsets, not complete or unmodified releases of the source datasets. Dataset-specific documentation describes array formats, processing conventions, and known limitations. Source references Wei Chen and Yuxuan Liang. Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting. ICLR, 2025. Xu Chen, Junshan Wang, and Kunqing Xie. TrafficStream: A Streaming Traffic Flow Forecasting Framework Based on Graph Neural Networks and Continual Learning. IJCAI, pp. 3620–3626, 2021. https://doi.org/10.24963/ijcai.2021/498 Nans Addor, Andrew J. Newman, Naoki Mizukami, and Martyn P. Clark. The CAMELS data set: catchment attributes and meteorology for large-sample studies. Hydrology and Earth System Sciences, 21(10):5293–5313, 2017. https://doi.org/10.5194/hess-21-5293-2017
