.png){kind=icon}
.png){kind=icon}
Return
Short-term electrical load forecasting based on pattern label vector generation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.enbuild.2025.115383.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Short-term load forecasting (STLF) is critical for achieving grid-load interactions and energy-efficient operations. Nevertheless, natural and social uncertainties increase the difficulty of electrical load forecasting. This study proposes a novel method based on pattern label vector generation (PLVG), which includes labeling and generation steps, to guide the electrical load all-day forecasting. During the labeling process, PLVG considers historical electrical load, weather, and social behavior pattern information to form a single-dimensional combinatorial label vector. During the generation process, the Word2Vec model is utilized to map each label to a word vector, thereby generating the corresponding dataset. Subsequently, a Convolutional Neural Network (CNN) combined with an attention-based Bidirectional Long Short-Term Memory (BiLSTM) model is employed to generate electrical load pattern labels covering the full forecasting cycle based on the word vector dataset. These labels are then matched with the cluster center vectors to obtain feature vectors. The feature vectors extracted by the PLVG model serve as inputs to the forecasting model, effectively guiding electrical load forecasting. Moreover, the features include explainable patterns of real electrical load changes that are closely linked to societal patterns. Experimental results on the GEFCom2014 public dataset and an office building dataset in Shanghai, China, show that the PLVG model outperforms mainstream electrical load forecasting methods, such as Transformer, Seq2Seq, etc., which verifies its excellent performance and generalizability. The proposed approach significantly improves the accuracy of electrical load forecasting and provides precise predictive support for resource allocation and scheduling in power systems, highlighting its substantial practical application value.
Keywords:
All-day forecasting
Bidirectional long short-term memory
Pattern label vector generation
Short-term load forecasting
Word2vec model