Real-time monitoring of daily carbon emissions and electricity generation from fossil fuel power plants using geostationary satellite band data and deep learning techniques

The power sector is the primary sources of carbon emissions today. Traditional methods for tracking carbon emissions from fossil fuel power plants typically depend on individual reporting that is susceptible to falsification, expensive emission monitoring systems, or satellite data from sun-synchronous orbits that offer limited temporal resolution. To achieve the target of low cost and high-frequency continuous monitoring, this study introduced a cost-effective, real-time monitoring method using deep learning and self-attention mechanisms to predict daily carbon emissions and power generation from fossil fuel plants, employing high-frequency geostationary satellite data for the first time. Validated in the U.S. with GOES-16 satellite data, the model outperformed other benchmarks, and performed well in regional and national monitoring tasks. It effectively estimated carbon intensity (carbon emission per unit of electricity output) over regional scale. In terms of predictive sensitivity, prediction errors increased dramatically when actual emissions exceeded approximately 30000 Tons per day, and seasonal variations also had significantly impact on mode performance. An analysis of model's extended application showed potential recording omissions in U.S. CEMS data ranging between 10 % and 25 %. For interpretability, the model highlighted data from near-infrared bands that highly correlated with CO2 levels and weather condition. The study marks an incremental advance in real-time emission monitoring to continuously track the low-carbon transition process of power industry.