What type of study is this?

This is a Literature Review study.

August 18, 2025Open Access

The Application of Hybrid Deep Learning Method of Improved Transformer Model in Load Forecasting Based on Future Network and Edge Computing

LWLei WangTianjin University HSHaitao SongTianjin Research Institute of Electric Science (China)QYQingguang YuHebei Institute of Physical Education

Key Points

The hybrid deep learning model enhances load forecasting accuracy by integrating user-declared load features with historical data, providing a more comprehensive view.
Model evaluation using electricity load data indicates significant improvements in prediction performance compared to traditional methods, particularly in handling nonlinear relationships.
The optimized feature selection strategy utilizing LightGBM enables better identification of relevant load attributes, enhancing the model's adaptability to complex load patterns.
Integration of future networks with edge computing supports real-time processing, emphasizing its importance in modern power system management and forecasting.

Abstract

Short-term load forecasting is the basis for the safe operation of the power system, and is the guarantee for the power system to make a rapid response to load changes. With the promotion of source-grid-load-storage integration, user load has also become a means of regulation, while existing load forecasting methods do not take into account the subjective initiative of users. In addition, the rapid development of future networks and the widespread application of edge computing technology will provide new opportunities for the intelligent management and efficient operation of power systems. Traditional statistical models rely on linear assumptions and smoothness requirements, making it difficult to portray the nonlinear characteristics of loads and multi-factor coupling relationships; machine learning methods improve forecasting flexibility, but their performance is limited by artificial feature engineering and inadequate modeling of time-series dynamics. This paper combines user-declared load features with passive historical-time-meteorological features, which solves the problem of neglecting user initiative in traditional methods. In the model construction, the feature selection strategy is optimized by integrating light gradient boosting machine (LightGBM) algorithm, which enhances the model’s ability to handle complex load data. In addition, this paper also improves the Transformer model by adopting a model design based on the encoder–decoder architecture, which can more accurately reveal the relationship between the load input features and the actual power demand, thus enhancing the prediction accuracy. First, data analysis is performed on the declared load and the actual load, and it is concluded that the declared load is very close to the actual load and has a normal distribution relationship. Subsequently, the load input features are categorized into temporal, historical, meteorological, and specific load attributes. The LightGBM model is employed to filter these attributes, selecting those with high relevance. Following this, Then, a hybrid deep learning model is constructed, combining the efficient data transmission capabilities of future networks with the real-time processing advantages of edge computing to further improve the model’s predictive performance. The model and algorithm were then evaluated using electricity load data from the Tianjin smart energy service platform. The algorithm’s outcomes demonstrated the proposed method’s precision and effectiveness in considering the declared load, and the integration of future networks and edge computing will provide more powerful technical support for power load forecasting.

KI fragen

Bookmark

View Full Paper