Privacy-preserving short-term carbon intensity forecasting in power systems

Power systems significantly contribute to global carbon emissions. It is the energy demand that drives power generation then resulting in carbon emissions; thus, it is crucial to link carbon emissions to electricity consumption through carbon intensity measures. Accurate short-term carbon intensity forecasting can help electricity consumers perform low-carbon demand responses to minimize overall carbon emissions and operational costs. However, consumption-based carbon forecasting is dependent on variational generation data and the power flow information down to every node in the networks, which is hard to obtain. In addition, privacy concerns arise from the sharing of sensitive information. This paper solves the problem by proposing a D ifferentially P rivate two-level F eature E ngineering-assisted C arbon I ntensity F orecasting (DP-FECIF) framework. The proposed method aggregates the energy production under local differential private generation data, then produces day-ahead forecasts for each energy source using machine learning models. The forecasting accuracy is enhanced using an automated feature generation and selection process. Case studies are conducted on two open carbon intensity datasets from Denmark and PJM ISO. The proposed model outperforms benchmark models with an average RMSE improvement of 9.79%, and the feature engineering method restores RMSE statistics with an average of 12.32% across different test cases.