Hourly Natural Gas Load Forecasting Based on STL-XGBoost-NBEATSx

Authors:

Shao Bilin, Ren Meng, Tian Ning

Affiliation:

School of Management, Xi’an University of Architecture and Technology, Xi’an, Shaanxi 710311

Author Biography:

Shao Bilin (1965-), male, professor, master’s supervisor, doctoral supervisor, research interests include big data, artificial intelligence, data information and management, energy sustainable development.

E-mail:

[email protected]

Corresponding Author:

Ren Meng (1999-), female, master’s student, research interests include energy forecasting, machine learning, data information and management.

E-mail:

[email protected]

Classification Number:

TP 996; TP 183

Document Identifier:

A

Article Number:

1673-5005(2024)03-0170-10

DOI:

10.3969/j.issn.1673-5005.2024.03.019

The hourly natural gas load forecasting is influenced by external feature factors and forecasting methods. To improve its forecasting accuracy and solve the issues of poor interpretability and long training time in other deep learning models or ensemble models, this paper introduces a new feature factor called “hourly impact degree” and proposes a forecasting method based on the Extreme Gradient Boosting Trees (XGBoost) model and the interpretable neural network model NBEATSx. The XGBoost model serves as a feature selector to filter the feature set data, and the filtered dataset is then input into NBEATSx for training, improving the training speed and forecasting accuracy of NBEATSx. The load data and feature data are decomposed into trend components, seasonal components, and residual components using the Seasonal and Trend decomposition using Loess (STL) algorithm, which are then input into XGBoost for prediction to reduce the noise impact in the original data. The optimized NBEATSx and XGBoost models are combined using the variance reciprocal method to yield the prediction results of the STL-XGBoost-NBEATSx ensemble model. The results indicate that the new feature, “hourly impact degree,” is an important influencing factor for hourly load forecasting. The STL-XGBoost-NBEATSx model shows improved training speed, good interpretability, and higher prediction accuracy, with the average absolute percentage error, mean square error, and mean absolute error reduced by 54.20%, 63.97%, and 49.72% respectively compared to other single models, and reduced by 24.85%, 34.39%, and 23.41% compared to other ensemble models. The model’s coefficient of determination is 0.935, indicating a good fit to the observed data.

Natural gas load forecasting; hourly impact factors; extreme gradient boosting trees; interpretability; NBEATSx; ensemble model