Prediction optimization fusion learning-based approach for day-ahead carbon aware scheduling in distribution network

As an important measure for the greenhouse emission mitigation of the power system, low carbon load management methods are adopted by system operators, which require information on carbon emission intensity. The accurate prediction of the nodal carbon intensity (NCI) is challenging due to numerous uncertain resources. Existing prediction methods mitigate the inevitable prediction error, and overlook the impact on downstream decision-making of load scheduling. Hence, this paper proposes a prediction optimization fusion learning-based (POFL) approach for shiftable load scheduling to compensate for the accuracy loss in decision-making. The sequence neural network based prediction model is integrated with a two-stage dispatch model, which generates low carbon shiftable load scheduling plans based on predicted NCI. Moreover, the surrogate dynamic regret is derived to measure the gap between the actual decision under predicted NCI and the oracle decision under true NCI in the load scheduling problem with time variant parameters. Simulation results illustrate that the proposed approach performs better in carbon-aware decision-making with comparable prediction accuracy compared with the state-of-the-art methods.