Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LightGTS-Cov: Covariate-Enhanced Time Series Forecasting

Yong Shang, Zhipeng Yao, Ning Jin, Xiangfei Qiu, Hui Zhang, Bin Yang

TL;DR

This work addresses covariate-rich time-series forecasting in industry by extending the lightweight LightGTS backbone with a decoder-side residual fusion module that incorporates both past covariates and horizon-aligned future covariates. The core idea is to fuse exogenous information through a two-stage post-decoder MLP with time-aligned token representations, enabling covariate-conditioned refinements without altering the backbone’s decoding process. Empirically, LightGTS-Cov improves over the LightGTS baseline and remains competitive with larger covariate-aware models across public benchmarks (EPF and Energy) and two real-world deployments in PV power forecasting and day-ahead electricity pricing. The approach demonstrates strong practical value, offering a deployable, parameter-efficient solution that enhances forecasting stability and accuracy in covariate-rich industrial settings.

Abstract

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.

LightGTS-Cov: Covariate-Enhanced Time Series Forecasting

TL;DR

This work addresses covariate-rich time-series forecasting in industry by extending the lightweight LightGTS backbone with a decoder-side residual fusion module that incorporates both past covariates and horizon-aligned future covariates. The core idea is to fuse exogenous information through a two-stage post-decoder MLP with time-aligned token representations, enabling covariate-conditioned refinements without altering the backbone’s decoding process. Empirically, LightGTS-Cov improves over the LightGTS baseline and remains competitive with larger covariate-aware models across public benchmarks (EPF and Energy) and two real-world deployments in PV power forecasting and day-ahead electricity pricing. The approach demonstrates strong practical value, offering a deployable, parameter-efficient solution that enhances forecasting stability and accuracy in covariate-rich industrial settings.

Abstract

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 1M-parameter LightGTS backbone, LightGTS-Cov adds only a 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.
Paper Structure (31 sections, 12 equations, 5 figures, 6 tables)

This paper contains 31 sections, 12 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time Series Foundation Models
  4. Covariate-aware Deep Forecasting Models
  5. Adaptation Methods for TSFMs
  6. Methodology
  7. Problem Settings
  8. Structure Overview
  9. Target Backbone Representations
  10. Covariate Pathways
  11. Post-Decoder Fusion
  12. Training Strategy
  13. Experiments
  14. Experimental Setup
  15. Results on Public Benchmarks
  16. ...and 16 more sections

Figures (5)

  • Figure 1: Architecture of LightGTS-Cov. It takes as input past targets, past covariates, and future-known covariates, and consists of (i) a pretrained LightGTS backbone, (ii) a two-stage post-decoder MLP fusion module for past and future covariates, and (iii) a shared output head that projects tokens to the final forecast.
  • Figure 2: Long-term PV power forecasting results in Industrial Deployment I: MSE (left) and MAE (right).
  • Figure 3: Day-ahead electricity price forecasting results in Industrial Deployment II: MSE (left) and MAE (right).
  • Figure 4: PV power forecasting comparison over two inverter channels (top: #17; bottom: #23) using 10-day history to predict the next 15 days at 5-minute resolution. The dashed line marks the forecast start. LightGTS-Cov better aligns peak timing and magnitude. Curves show ground truth (blue) and predictions from LightGTS-Cov (red), Chronos-2 (orange), and Sundial (green).
  • Figure 5: Day-ahead electricity price forecasting under two representative windows (top: weekday; bottom: weekend). The dashed line marks the forecast start and the shaded region indicates the forecast horizon. LightGTS-Cov better tracks trend changes. Curves show ground truth (blue) and predictions from LightGTS-Cov (red), Chronos-2 (orange), and Sundial (green).