Implementing an enhanced self-attention-based imputation for time series on missing industrial load data

Electrical load data, a fundamental component of grid information systems, harbors a wealth of valuable insights but is frequently subject to data gaps due to sensing and storage challenges. In response to these gaps in electrical load data, this study introduces an enhanced method for completing missing industrial load data using an improved Self-Attention-Based Imputation for Time Series (SAITS) approach. This enhancement entails the development of a novel Dual-Gate Feedforward Network (DGFFN), which incorporates a dual-gate mechanism and local information amplification to improve the precision of data completion tasks. Through a case study analysis employing industrial load data, findings demonstrate that the SAITS method, augmented with DGFFN, is capable of capturing a broader array of local information, more effectively uncovering missing data, and reducing inaccuracies. This approach is shown to be applicable to the preprocessing of load data within electrical power systems.