Short-term load forecasting based on temporal importance analysis and feature extraction

• Published in: Electric power systems research Vol. 244; p. 111551
• Main Authors: Yuqi, Ji, An, An, Lu, Zhang, Ping, He, Xiaomei, Liu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2025
• Subjects: Convolution neural network; EIIR importance algorithm; load feature extraction; Short-term load forecasting; Short-term load forecasting; Convolution neural network; load feature extraction; EIIR importance algorithm
• DOI: 10.1016/j.epsr.2025.111551
• ISSN: 0378-7796

•An algorithm for extracting the importance of extreme points based on time series is proposed. Through the marking and calculation of the importance of extreme points, the morphological features of load curves are captured, particularly those of the intervals between extreme points.•This paper focuses on the prediction accuracy of extreme points. Using the extreme importance features extracted by the EIIR algorithm proposed in this paper, machine learning models can pay more attention to one or more extreme regions without relying on attention mechanisms, thereby improving the prediction accuracy of these regions.•In order to guarantee the reliability of model prediction, the Gaussian distribution is used to generate a total of 25 groups of noise with different degrees to interfere with the original Drybulb and THI data. The experimental results show that K-E-CNN-LSTM has good robustness to data noise. Efficient and accurate short-term load forecasting plays a crucial role in ensuring the safe and stable operation of power systems and achieving economic management. This paper proposes an EIIR (Enhanced Importance Index Recognize) importance marking algorithm. This algorithm can extract the importance of each point in the load series, especially extreme points, so that machine learning models can focus on areas of high importance during training. This fills the research gap in the morphological characteristics of time series for peak and valley prediction. First, the K-Medoids algorithm is used to cluster the daily load curve, and then the EIIR algorithm is used to extract the numerical features of the extreme value points of various cluster centers. Then the importance features and historical data are reconstructed into a new feature set and input them into the CNN-LSTM hybrid neural network for prediction. Finally, the ISONE public power load data set is taken as an example for analysis and verification. In order to verify the reliability of the model prediction, the robustness of the model was analyzed and verified by adding input interference. The results show that this method can achieve more accurate short-term load prediction, and the model has good stability and robustness.