Resolution-Aware Retrieval Augmented Zero-Shot Forecasting

TL;DR

This work tackles zero-shot microclimate forecasting by introducing a Resolution-Aware Retrieval-Augmented Forecasting framework that decomposes signals into frequency components and retrieves context accordingly. It combines an attention-based transfer mechanism and frequency-aware retrieval with a probabilistic Gaussian forecasting objective, implemented atop an encoder–decoder backbone (Informer). Key contributions include wavelet-based multi-resolution retrieval, location embeddings (Learned and Satclip), a flexible transfer component (FC, GNN, or Location Attention), and a two-phase training procedure, all validated on ERA5 showing superior performance over HRRR, Chronos, and other baselines across multiple tasks. The results demonstrate data-efficient, scalable zero-shot forecasting with strong long-horizon and cold-start capabilities, enabling foundation-model-like microclimate predictions for across-the-US and potentially other spatiotemporal domains.

Abstract

Zero-shot forecasting aims to predict outcomes for previously unseen conditions without direct historical data, posing a significant challenge for traditional forecasting methods. We introduce a Resolution-Aware Retrieval-Augmented Forecasting model that enhances predictive accuracy by leveraging spatial correlations and temporal frequency characteristics. By decomposing signals into different frequency components, our model employs resolution-aware retrieval, where lower-frequency components rely on broader spatial context, while higher-frequency components focus on local influences. This allows the model to dynamically retrieve relevant data and adapt to new locations with minimal historical context. Applied to microclimate forecasting, our model significantly outperforms traditional forecasting methods, numerical weather prediction models, and modern foundation time series models, achieving 71% lower MSE than HRRR and 34% lower MSE than Chronos on the ERA5 dataset. Our results highlight the effectiveness of retrieval-augmented and resolution-aware strategies, offering a scalable and data-efficient solution for zero-shot forecasting in microclimate modeling and beyond.

Figures (8)

• Figure 1: Overall structure of the model with location attention transfer component and resolution-aware retrieval. The limited available context for the target location ($x_{tar}$) is first decomposed into 3 levels of fast, moderate and slow, for each of these components, our retrieval augmented forecasting model with location attention is used. For the fast components we retrieve the fast components of a smaller set of closer points to the target, while for slower components we retrieve more points.
• Figure 2: Correlation between historical temperature data and distance for a subset of point pairs across the U.S. At low frequencies (left), correlation remains high over long distances, while at high frequencies (right), it declines more rapidly.
• Figure 3: Retrieval augmented forecasting model that uses attention over location embeddings as the transfer module.
• Figure 4: Forecasts of our full model agianst those of HRRR and ground truth. The forecasts are ran at every midnight for 24 hours for both models.
• Figure 5: MSE and MAPE for test data for each hours of forecast for three models of Chronos base, Informer model trained on the training stations and the full model with resolution-aware retrieval (Informer + Ret + Dec).