Physics-informed bidirectional temporal convolutional imputation network (PI-BTCIN) for microgrid missing power data

High-quality operational data are fundamental to ensuring comprehensive situational awareness and the stable operation of microgrids, serving as a cornerstone for the sustainable development of distributed energy systems. However, real-world operations frequently suffer from data loss due to communication failures, significantly compromising system observability. Existing data-driven methods often fail to simultaneously capture long-term temporal dependencies and the inherent physical coupling among multi-source variables, leading to imputations violating system operational dynamics. To address these challenges, this paper proposes a physics-informed bidirectional temporal convolutional imputation network (PI-BTCIN). First, a residual temporal convolutional network (TCN) combined with a time decay mechanism is employed to extract multi-scale temporal features from time series. Second, a bidirectional temporal fusion architecture is constructed to integrate forward and backward contextual information, thereby enhancing reconstruction performance over consecutive missing intervals. Furthermore, physical priors are transformed into regularization loss terms and integrated into training, guiding the model to generate estimates that align with operational characteristics. Experiments on a real-world microgrid dataset show that PI-BTCIN obtains the lowest mean absolute error across all tested missing rates, with an average improvement of ∼17% over the strongest existing imputation baseline and ranks first or second in root mean square error under all tested settings. Ablation results further validate the contributions of the residual TCN, time decay, bidirectional temporal fusion, and physical regularization components.