A Bi-Directional Missing Data Imputation Scheme Based on CNN and Short-Term Window Sliding for Thermal Power Units Data

Under the global clean energy and low-carbon development trend, the thermal power generation industry faces multiple pressures, including technological upgrades, industrial adjustments, and environmental protection. High-quality data recording of thermal power units is crucial for ensuring efficient economic dispatch of electricity. However, data loss is common due to sensor failures and equipment maintenance. Although data missingness has been extensively studied in other fields, there is limited in-depth research in the field of thermal power units, which mainly relies on traditional statistical methods or machine learning techniques. To address this, this paper proposes an innovative method that combines Convolutional Neural Networks, bidirectional imputation, and short-period window sliding techniques (CNN-BIS). This method integrates the powerful modeling capabilities of deep learning with the characteristics of thermal power unit data. Using unit data from a power plant for validation, we found that the results show the performance of the CNN-BIS model is significantly improved compared to baseline models across datasets with different missing rates and time periods. In modeling the flue gas oxygen content before and after data imputation, the prediction performance of various models is significantly enhanced, with the Root Mean Square Error (RMSE) reduced by 0.59% to 16.53%.