SPECTRA: New AI forecasts energy with 5.7% better accuracy

SPECTRA introduces a principled design that treats trend-periodic components as the baseline trajectory and high-frequency residuals as the source of forecast uncertainty. The architecture adaptively separates deterministic and residual streams, aligns external variables (exogenous context) with both, refines the deterministic backbone via multi-resolution spectral-temporal state-space modeling, and then estimates ordered quantile boundaries from the combined representations. This explicit separation allows the model to capture both predictable structures and uncertainty-bearing fluctuations in a single end-to-end framework.

The authors tested SPECTRA on four energy forecasting tasks — load, price, solar, and wind — and compared it against state-of-the-art baselines including transformer-based and diffusion-based methods. The results are statistically significant: SPECTRA achieved the best continuous ranked probability score (CRPS) in 14 out of 18 settings, with an average CRPS reduction of 5.74% and a 7.27% reduction in upper-tail quantile risk. These gains are particularly valuable for risk-sensitive applications like grid balancing and trading. The paper is currently under review at IEEE Transactions on Power Systems.