FFGN: a Fourier graph network for spatiotemporal forecasting

To address the challenges of capturing global dependencies and the "physics-ignorant" limitations in spatio-temporal graph forecasting, this paper proposes the Flattened Fourier Graph Network (FFGN), a novel architecture operating in the frequency domain. FFGN treats the entire spatio-temporal graph as a composite signal , uses the Fast Fourier Transform (FFT) to project it into the frequency domain , and applies a novel Fourier Graph Convolution (FourierGC) layer to efficiently learn global patterns. Furthermore, the model adopts the Physics-Informed Meta-Learning (PIML) framework and is pre-trained using the Reptile algorithm to learn underlying physical principles, such as the price-demand relationship. Experiments demonstrate that FFGN achieves strong performance on standard forecasting tasks and, unlike "physics-agnostic" baselines , it successfully exhibits a qualitatively correct physical response to unseen price pulses.