Natural Gas Consumption Forecasting Model Based on Feature Optimization and Incremental Long Short-Term Memory
Natural gas, as a vital component of the global energy structure, is widely utilized as an important strategic resource and essential commodity in various fields, including military applications, urban power generation and heating, and manufacturing. Therefore, accurately assessing energy consumptio...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-05-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/25/10/3079 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Natural gas, as a vital component of the global energy structure, is widely utilized as an important strategic resource and essential commodity in various fields, including military applications, urban power generation and heating, and manufacturing. Therefore, accurately assessing energy consumption to ensure a reliable supply for both military and civilian use has become crucial. Traditional methods have attempted to leverage long-range features guided by prior knowledge (such as seasonal data, weather, and holiday data). However, they often fail to analyze the reasonable correlations among these features. This paper proposes a natural gas consumption forecasting model based on feature optimization and incremental LSTM. The proposed method enhances the robustness and generalization capability of the model at the data level by combining Gaussian Mixture Models to handle missing and anomalous data through modeling and sampling. Subsequently, a weakly supervised cascade network for feature selection is designed to enable the model to adaptively select features based on prior knowledge. Finally, an incremental learning-based regression difference loss is introduced to promote the model’s understanding of the coupled relationships within the data distribution. The proposed method demonstrates exceptional performance in daily urban gas load forecasting for Wuhan over the period from 2011 to 2024. Specifically, it achieves notably low average prediction errors of 0.0556 and 0.0392 on the top 10 heating and non-heating days, respectively. These results highlight the model’s strong generalization capability and its potential for reliable deployment across diverse gas consumption forecasting tasks within real-world deep learning applications. |
|---|---|
| ISSN: | 1424-8220 |