Enhanced Transformer-EA Evolutionary Algorithm Fusion Method for High-Precision Completion of Power Load Time Series Data

Abstract Aiming at the problem of missing power time-series data in smart grids caused by equipment failures, communication interruptions, etc., this paper proposes a high-precision completion method for power load time-series data based on the deep integration of enhanced Transformer and evolutionary algorithm (EA). This method optimizes the time-series attention mechanism of the Transformer model through a dual-baseline pre-filling strategy, combined with the periodic characteristics and global statistical information of power loads, and introduces a dynamically decaying EA evolutionary algorithm to finely repair missing data. To further improve the completion accuracy, a multi-dimensional weighted loss function is designed in this paper to emphasize the accuracy optimization during peak hours, and physical compliance constraints are used to ensure that the completion results fall within the reasonable range of power loads. Experimental results show that the proposed method exhibits superior completion performance under different missing mechanisms (MCAR, MAR, MNAR) and missing rates (10%, 20%, 30%). Compared with traditional time-series models and basic Transformer, the enhanced Transformer-EA model has significant improvements in core indicators such as R², MAPE, and RMSE. This method provides an efficient solution for power load data missing completion and has broad application prospects and practical value