Robust imputation of missing photovoltaic power data using a weather- and context-aware hybrid transformer framework

Abstract

Accurate imputation of missing photovoltaic (PV) power data is critical for ensuring the reliability of downstream energy management systems. This paper proposes a novel imputation framework that leverages both external knowledge and internal data patterns to enhance imputation performance in complex scenarios with high missing data rates. A weather-prompt and context-knowledge fusion mechanism is designed to incorporate meteorological features alongside coarse imputation results. These semantic prompts provide valuable environmental and temporal context, improving the model's ability to better understand missing data regions. The core architecture features a hybrid design that integrates Transformer modules with diagonal masked self-attention (DMSA) to capture different levels of temporal dependencies. These modules work synergistically with coarse-to-fine imputation layers and context-aware refinement blocks, enabling progressive data reconstruction and robust generalization across varying conditions. Comprehensive robustness evaluations demonstrate the model's ability to maintain high imputation accuracy even under extremely high missing rates. Compared with the strongest baseline, our model reduces MAE and RMSE by up to 50.5 % and 55.0 % on the DKASC dataset and 47.4 % and 52.9 % on the Hebei dataset under 90 % missing conditions. Furthermore, the model remains effective in scenarios where meteorological inputs are unavailable, when missing rates differ between the training and testing phases, or when the input PV data is collected at different time resolutions. These findings highlight the strong adaptability and practical applicability of the proposed imputation framework for real-world PV data scenarios.