Learning global and local features of power load series through transformer and 2D-CNN: An image-based multi-step forecasting approach incorporating phase space reconstruction

Abstract

As modern power systems continue to evolve, accurate power load forecasting remains a critical issue in energy management. The phase space reconstruction (PSR) method can effectively retain the inner chaotic property of power load from a system dynamics perspective and thus is a promising knowledge-based method for power load forecasting. To fully leverage the PSR method’s capability in modeling this high-dimensional, non-stationary characteristics of power load data, and to address the challenges faced by its classical mathematical prediction algorithms in effectively solving contemporary prediction scenarios characterized by massive volumes of data. This study proposes a novel learning-based multi-step forecasting approach that utilizes an image-based modeling perspective for the reconstructed phase trajectories. Firstly, the feature engineering approach that simultaneously utilizes dynamic evolution features and temporal locality features in the trajectory image is proposed. Through mathematical derivation, the equivalent characterization of the PSR method and another time series modeling approach, patch segmentation (PS), is demonstrated for the first time. Building on this prior knowledge, a novel image-based modeling perspective incorporating a global and local feature extraction strategy is introduced to fully leverage these valuable features. Subsequently, within this framework, a novel deep learning model, termed PSR-GALIEN, is designed for end-to-end processing. This model employs a Transformer Encoder and 2D convolutional neural networks (CNNs) to extract global and local patterns from the image, while a multi-layer perceptron (MLP)-based predictor is utilized for efficient correlation modeling. Extensive experiments on five real-world datasets show that PSR-GALIEN consistently outperforms six state-of-the-art deep learning models in short-term load forecasting scenarios with varying characteristics, demonstrating its great robustness. Ablation studies further confirm that the image-based modeling approach captures more relevant features than traditional sequential methods, improving the overall forecasting accuracy. Additionally, the attributions of the forecasting results of PSR-GALIEN can be explained through the proposed visualization-based method, which significantly enhances its interpretability.