LightGTS-Cov: Covariate-Enhanced Time Series Forecasting

Time series foundation models are typically pre-trained on large, multi-source datasets; however, they often ignore exogenous covariates or incorporate them via simple concatenation with the target series, which limits their effectiveness in covariate-rich applications such as electricity price forecasting and renewable energy forecasting. We introduce LightGTS-Cov, a covariate-enhanced extension of LightGTS that preserves its lightweight, period-aware backbone while explicitly incorporating both past and future-known covariates. Built on a $\sim$1M-parameter LightGTS backbone, LightGTS-Cov adds only a $\sim$0.1M-parameter MLP plug-in that integrates time-aligned covariates into the target forecasts by residually refining the outputs of the decoding process. Across covariate-aware benchmarks on electricity price and energy generation datasets, LightGTS-Cov consistently outperforms LightGTS and achieves superior performance over other covariate-aware baselines under both settings, regardless of whether future-known covariates are provided. We further demonstrate its practical value in two real-world energy case applications: long-term photovoltaic power forecasting with future weather forecasts and day-ahead electricity price forecasting with weather and dispatch-plan covariates. Across both applications, LightGTS-Cov achieves strong forecasting accuracy and stable operational performance after deployment, validating its effectiveness in real-world industrial settings.

💡 Research Summary

Time series foundation models (TSFMs) have shown impressive cross‑domain performance when pretrained on massive multi‑source corpora, yet most of them are designed for target‑only inputs and treat exogenous covariates as an after‑thought, often concatenating them with the target series. This practice limits their usefulness in covariate‑rich domains such as electricity price forecasting and renewable energy generation, where both historical and future‑known covariates (e.g., weather forecasts, dispatch plans) carry critical information.
The paper introduces LightGTS‑Cov, a covariate‑enhanced extension of the lightweight (≈1 M parameters) LightGTS model. LightGTS‑Cov adds a modest‑size (≈0.1 M parameters) two‑stage MLP plug‑in that operates after the decoder, preserving the original period‑aware tokenization, parallel non‑autoregressive decoding, and overall architecture of LightGTS.
Methodologically, past covariates are tokenized with the same period‑aware patching as the target series, producing time‑aligned latent tokens Z_past. Future‑known covariates are separately patched (with a configurable patch length P*) and linearly embedded into the same latent dimension, yielding tokens f that are aligned to the forecast horizon. For each decoded token position j, the model concatenates the target token Z_target