A domain knowledge-driven visual analytics system for photovoltaic power time series forecasting

Xinjing Yi, Yurun Yang, Yin Song, Baozhu Zhou, Cheng Li, Shuhan Liu, Dazhen Deng, Di Weng, Yingcai Wu

Abstract: Accurate forecasting of photovoltaic (PV) power generation plays a critical role in optimizing production schedules for PV power stations and facilitating efficient maintenance and repair of PV equipment. Deep time series forecasting models have emerged as the leading approach in the field of time series prediction. However, the intricate architectures and numerous parameters in these models frequently lead to forecasts that conflict with established domain knowledge. Furthermore, domain experts face significant challenges in leveraging their specialized knowledge to enhance the predictive performance of these models. Additionally, when dealing with PV strings exhibiting diverse characteristics, it becomes particularly challenging to identify discrepancies in the model’s predictive accuracy across different strings and pinpointing its performance flaws. To tackle the issues, we propose a domain knowledge-driven forecasting method that combines the deep time series forecasting model with decision tree regression. Empirical evaluations demonstrate that our proposed method yields substantial improvements in forecasting accuracy compared to existing approaches. Furthermore, we have developed and implemented a comprehensive visual analytics system. This system aids experts in forecasting PV string power generation trends by employing multi-perspective model evaluation techniques to assess the rationality of forecasts. It also incorporates domain knowledge through an interactive decision tree construction process, thereby enhancing the model’s predictive capabilities. The efficacy of our proposed system is substantiated through in-depth case studies and rigorous user evaluations.

External IDs:doi:10.1007/s12650-025-01067-4