Zero-shot Microclimate Prediction with Deep Learning

TL;DR

The paper tackles microclimate prediction for unmonitored locations by framing a zero-shot forecasting problem and using transfer learning across weather stations. It introduces a location-aware Transform component within an Informer-based encoder-decoder to map embeddings from data-rich source stations to a target station without local history, enabling forecasts over a horizon of $L_y$ from a past window of length $L_x$. The key contributions include a two-phase training scheme, a differentiable Transform that leverages station locations, and empirical evidence showing improved accuracy over HRRR and a plain Informer on both synthetic data and real AgWeatherNet data. This approach offers practical value for agricultural planning, urban design, and climate resilience by reducing dependence on nearby sensors for new sites.

Abstract

Weather station data is a valuable resource for climate prediction, however, its reliability can be limited in remote locations. To compound the issue, making local predictions often relies on sensor data that may not be accessible for a new, previously unmonitored location. In response to these challenges, we propose a novel zero-shot learning approach designed to forecast various climate measurements at new and unmonitored locations. Our method surpasses conventional weather forecasting techniques in predicting microclimate variables by leveraging knowledge extracted from other geographic locations.

Figures (3)

• Figure 1: Structure of the transform model. $st_1$ and $st_2$ are the train stations on which the encoder-decoder has been trained, and $st_{tar}$ is the target station which we did not have any training data for. We transform the encoding of $st_1$ and $st_2$ at time $t$ using our Transform component and then pass it through the decoder to get the next 24-hour forecast for the target station.
• Figure 2: Left: Informer model with and without the transform component when more and more training stations are added. The test mean squared error of HRRR model and when the Informer model is trained and tested on the same station data are shown. The error bars show the standard deviation of 5 runs of the models. Right: the average MSE values across multiple days for each of the 24 forecast hours.
• Figure 3: Predictions of our best zero-shot model compared with HRRR predictions and ground truth on the last two weeks of January 2023.