Photovoltaic power data imputation based on spatio-temporal fusing and improved TimesNet

Acquiring high-quality photovoltaic (PV) power data is a crucial prerequisite for the reliable and accurate dispatch and operation of future power grids. However, due to communication interruptions, equipment failures, and sensor malfunctions, station data frequently suffer from missing values, thereby severely compromising power forecasting, dispatch optimization, and system stability. Most existing methods focus on data imputation based on information from individual stations, although single-station imputation methods demonstrate acceptable performance in random missing scenarios, they are fundamentally constrained by the 'information island' effect. Due to the lack of spatial reference information, these models may have difficulty distinguishing abrupt power drops caused by weather variations from actual data absence when consecutive data losses occur because of equipment failures. Consequently, such approaches tend to yield deterministic smooth curves, severely underestimating the instantaneous volatility of PV power output and failing to meet the requirements for high-precision grid dispatch. To address these challenges, this paper proposes the multi-station adaptive graph attention network method. The model employs an adaptive graph learning module to adaptively fuse stations based on geographic proximity and statistical correlation. A parallel dual-branch encoder utilizes graph attention to capture spatial dependencies, while simultaneously leveraging TimesNet to characterize temporal cycles and capture temporal relationships. Furthermore, a mask-aware period folding mechanism is introduced within the temporal network to embed missing-position indicators into period detection and multi-scale convolutions, thereby enhancing imputation accuracy across various missing data rates. Extensive experiments show that the proposed multi-station adaptive graph TimesNet (MSAG-TimesNet) achieves the lowest mean absolute error (MAE) and root mean square error (RMSE) across all tested missing rates. Even under the severe 90% missingness condition, it maintains an MAE of 1.774 ± 0.161, demonstrating its ability to reconstruct PV power data under substantial data loss.