High accuracy carbon emission factor prediction under large scale data missing based on SAITS and N-BEATS

Against the backdrop of the global pursuit of carbon peaking and carbon neutrality, the precise prediction of carbon emission factors in the power sector not only concerns the efficient operation of the carbon market but also directly affects the accuracy of enterprises’ carbon accounting and the reliability of green power certification. However, large-scale, irregular missing values in grid operation time-series data (due to sensor failures, communication interruptions, etc.) severely reduce the accuracy of traditional prediction methods, becoming a bottleneck for precise carbon quantification in the power sector. To solve this problem, this study proposes a hybrid framework integrating SAITS (Self-Attention based Imputation for Time Series) and enhanced N-BEATS (Neural Basis Expansion Analysis for Time Series). The SAITS module achieves effective reconstruction of incomplete carbon coefficient sequences through the diagonal mask self-attention mechanism combined with observed reconstruction task/missing imputation task joint training. The enhanced N-BEATS module uses multi-scale trend-period decomposition and residual learning techniques to predict future values. This framework not only overcomes the problem of error accumulation in traditional processes but also can adapt to dynamic power grid data, providing a reliable solution for the precise prediction of carbon coefficients under high missing rate conditions, thereby offering strong support for carbon reduction decisions in the power industry.