Objective Daily natural gas load is influenced by multiple factors, including historical inertia, meteorological conditions, date types, and latent load regimes. However, static calendar variables fail to sufficiently characterize differences in dynamic operating states across different days. To address this gap, this study introduces a short-term natural gas load forecasting method enhanced by historical state features derived from the Gaussian Hidden Markov Model (GHMM). This approach extracts latent, temporally continuous load regimes and state transitions that basic features miss.

Methods A total of 2,557 daily samples spanning 2018–2024 are collected from the French Open Data Réseaux Énergies platform to construct 55 basic features. Seven intraday load-pattern indicators, three meteorological indicators and three calendar indicators are selected as observation variables for the GHMM. GHMMs with 3, 4, 5, and 6 hidden states are fitted on the training set. The RMSE of the coupled GHMM-XGBoost model on the validation set is used to select the number of states. After the optimal 5-state model is determined, the GHMM is refitted on the combined training and validation sets. Thirty-one GHMM-enhanced features are constructed according to the historical state sequence obtained via offline decoding. For forecasting, Persistence serves as the benchmark, while XGBoost and Light Gradient Boosting Machine (LightGBM) learn the daily load increment relative to the previous day. The forecasting performance of XGBoost, GHMM-XGBoost, LightGBM, and GHMM-LightGBM is compared.

Results On the test set, GHMM-XGBoost achieves a Mean Absolute Error (MAE) of 43,976.13 MW·h/d, RMSE of 62,883.62 MW·h/d, Mean Absolute Percentage Error (MAPE) of 4.7408％, and coefficient of determination (R²) of 0.9849. Compared to vanilla XGBoost, MAE, RMSE, and MAPE decrease by 2.85％, 2.55％, and 2.02％, respectively, while R² increases by 0.0008. Similarly, GHMM-LightGBM achieves a MAPE of 4.7194％, outperforming vanilla LightGBM (4.7748％). These results demonstrate that GHMM-enhanced features deliver consistent performance gains across different gradient-boosting tree-based models.

Conclusion With a fixed temporal split, fixed tree-model parameters, offline state decoding, and one-step-ahead daily incremental forecasting, GHMM-enhanced features further characterize the persistence and transition patterns of load regimes on the basis of existing historical load, meteorological, and calendar features. They provide tree-based models with supplementary temporally structured information. The research findings can offer methodological references for dynamic natural-gas load-regime identification, multi-source feature fusion, and temporal-state enhancement for conventional machine learning models.