Multistep photovoltaic power forecasting based on multi-timescale fluctuation aggregation attention mechanism and contrastive learning
Liang Yuan, Xiangting Wang, Yao Sun, Xubin Liu, Zhao Yang Dong
Abstract
• Introduced an MFA-attention method for PV forecasting, modeling trends and seasonal info separately. • Developed a similar day selection method using contrastive learning and DTW for better predictions. • Proposed a multi-step prediction approach to avoid error accumulation and validated it extensively. The integration of photovoltaic (PV) power into electrical grids introduces significant uncertainty due to the inherent volatility and intermittency of solar energy, underscoring the need for precise short and medium-term PV power forecasting. Despite the superior performance of Transformer-based time series methods, their application to PV power prediction remains suboptimal. In response to this deficiency, this paper proposes a novel attention mechanism that aggregates fluctuations across multiple time scales. This mechanism enhances the segmentation and extraction of nonlinear correlations between PV power outputs and meteorological factors, assigning variable weights to patterns of change across different time scales. Furthermore, a novel approach for selecting similar days is also developed based on contrastive learning, which enables self-supervised identification of similarities among PV power samples and enhances the model’s attention to local dynamic variations. Comparative analysis with eight state-of-the-art benchmark methods shows that the proposed MFA-attention model achieves lower prediction errors and improved effectiveness. © 2017 Elsevier Inc. All rights reserved.