Learning from History: A Retrieval-Augmented Framework for Spatiotemporal Prediction

2025/10/28 by Jia, Hao, Zhao, Penghao, Wu, Hao +3
#Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · doi:10.48550/arxiv.2510.24049

Abstract

Accurate and long-term spatiotemporal prediction for complex physical systems remains a fundamental challenge in scientific computing. While deep learning models, as powerful parametric approximators, have shown remarkable success, they suffer from a critical limitation: the accumulation of errors during long-term autoregressive rollouts often leads to physically implausible artifacts. This deficiency arises from their purely parametric nature, which struggles to capture the full constraints of a system's intrinsic dynamics. To address this, we introduce a novel Retrieval-Augmented Prediction (RAP) framework, a hybrid paradigm that synergizes the predictive power of deep networks with the grounded truth of historical data. The core philosophy of RAP is to leverage historical evolutionary exemplars as a non-parametric estimate of the system's local dynamics. For any given state, RAP efficiently retrieves the most similar historical analog from a large-scale database. The true future evolution of this analog then serves as a reference target. Critically, this target is not a hard constraint in the loss function but rather a powerful conditional input to a specialized dual-stream architecture. It provides strong dynamic guidance, steering the model's predictions towards physically viable trajectories. In extensive benchmarks across meteorology, turbulence, and fire simulation, RAP not only surpasses state-of-the-art methods but also significantly outperforms a strong analog-only forecasting baseline. More importantly, RAP generates predictions that are more physically realistic by effectively suppressing error divergence in long-term rollouts.

Citations

• Advanced Long-term Earth System Forecasting
• Turb-L1: Achieving Long-term Turbulence Tracing By Tackling Spectral Bias
• BeamVQ: Beam Search with Vector Quantization to Mitigate Data Scarcity in Physical Spatiotemporal Forecasting
• OneForecast: A Universal Framework for Global and Regional Weather Forecasting
• Retrieval-Augmented Diffusion Models for Time Series Forecasting
• Retrieval-Augmented Generation for AI-Generated Content: A Survey
• Earthfarseer: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model
• Learning skillful medium-range global weather forecasting
• MgNO: Efficient Parameterization of Linear Operators via Multigrid
• Accurate medium-range global weather forecasting with 3D neural networks
• PastNet: Introducing Physical Inductive Biases for Spatio-temporal Video Prediction
• Convolutional Neural Operators for robust and accurate learning of PDEs
• Solving High-Dimensional PDEs with Latent Spectral Models
• FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators
• Earthformer: Exploring Space-Time Transformers for Earth System Forecasting
• SimVP: Simpler yet Better Video Prediction
• Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting
• PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning
• The ERA5 global reanalysis
• Deep Residual Learning for Image Recognition
• Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting
• U-Net: Convolutional Networks for Biomedical Image Segmentation
• SimVPv2: Towards Simple yet Powerful Spatiotemporal Predictive Learning

Related

• Predicting the Future by Retrieving the Past
• Learning on predictions: Fusing training and autoregressive inference for long-term spatiotemporal forecasts
• LETS Forecast: Learning Embedology for Time Series Forecasting
• Retrieval Augmented Time Series Forecasting
• TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster