Agentic AI for Weather Forecasting: A Reproducible Fourier-Domain Neural Operator Pipeline for Short-Horizon, Station-Level Multivariate Weather Predict

Abstract

Traditionally, weather forecasting has been achieved by numerical weather prediction (NWP) systems built from physics-based models that numerically integrate the equations of motion for air on a supercomputer, which are accurate but expensive to run, can be in error due to chaotic dynamics in the atmosphere, and require a significant amount of computing power to make forecasts. In the last ten years, two parallel developments have changed this landscape: firstly, data-driven deep-learning weather models, like FourCastNet, Pangu-Weather, GraphCast, FengWu, FuXi, GenCast and Aurora, that learn atmospheric behaviour directly from historical reanalysis data, and can reach or even surpass NWP accuracy with much less computational cost, and secondly, agentic AI frameworks such as ReAct, AutoGPT, MetaGPT, AutoGen, and Reflexion, that allow for autonomous planning, tool-use, self-verification, and multi-agent orchestration, in the software-engineering and general-purpose task domains. The analysis of 35 papers within both fields reveals no current approach that integrates autonomous, self-verifying, uncertainty-aware reasoning with a short-term forecasting model that is driven by data. Agentic applications found in the weather field are largely post-hoc interpretation, reporting and orchestration layers around externally provided forecasts. Conceptually, this is inspired by FourCastNet, but in this paper we propose and evaluate a lightweight Adaptive Fourier Neural Operator (AFNO) type forecasting model, which is derived from a global gridded atmospheric dataset to a single-station, tabular multivariate time-series setting, and provides an empirical basis for closing this gap. Using 24 hours of six continuous surface weather variables (temperature, dew point temperature, relative humidity, wind speed, visibility and pressure) as input, two stacked Fourier-domain temporal-mixing blocks with a 64-dimensional hidden representation, residual connections and layer normalisation directly forecast the same six variables one hour ahead. The model is trained and evaluated on 8,784 hourly single station observations with a chronological split of 70/15/15 for the training, validation and test sets, respectively, yielding an overall test-set MAE of 1.9138, MSE of 12.5279, RMSE of 3.5395 and R² of 0.8854; the best accuracy is achieved for pressure (R² = 0.9921) and the lowest for visibility (R² = 0.7151). These findings, along with a qualitative comparison to other models of weather forecasting and nowcasting that use global data, demonstrate the effectiveness of Fourier-domain temporal mixing as a lightweight inductive bias for station-level, short-horizon weather forecasting and provide an empirical benchmark for future investigations of integrating this forecasting capability with an autonomous, self-verifying, uncertainty-aware agentic reasoning layer.