Explainable time-varying directional representations for photovoltaic power generation forecasting
Zhijin Wang, Hanjing Liu, Senzhen Wu, Niansheng Liu, Xiufeng Liu, Yue Hu, Yonggang Fu
Abstract
Accurate photovoltaic (PV) power generation forecasting is crucial for optimizing the integration of solar energy into power grids and advancing towards a cleaner, more sustainable energy future. However, the inherent variability and complexity of PV power generation data pose significant challenges for accurate forecasting. To address these challenges, this paper introduces Collaborative Directional Representation (CoDR), a novel deep learning model that extracts and represents the directional fluctuations of solar irradiance data to improve forecasting accuracy and reliability. CoDR utilizes a series of steps, including data preprocessing, fluctuation extraction, directional representation, linearization, de-extraction, mapping, and data postprocessing, to capture both the temporal and spatial dependencies within the data. CoDR leverages a unique directional representation to capture both temporal and spatial dependencies in the data, enabling superior forecasting accuracy and robustness compared to twenty-two state-of-the-art benchmark methods. We validate CoDR on a real-world dataset of PV power generation in Belgium, demonstrating its effectiveness through extensive experiments, ablation studies, and sensitivity analyses. Importantly, CoDR enhances the transparency of the forecasting process by revealing the causal relationships and directional influences among input variables and PV power output. This explainability feature provides valuable insights into the underlying drivers of PV power generation, promoting trust and informed decision-making in the transition to cleaner energy systems.