Data‐driven robust look‐ahead power dispatch with preidentification of the worst‐case scenario using combined multilayer perceptron

Abstract

AbstractTo address the uncertainty caused by renewable energy generation, two‐stage robust optimization (RO) has been proposed for intraday look‐ahead power dispatch (LAPD) at multiple time scales. However, the intensive computational burden of solving RO problems limits its online application. To address this issue, we developed a novel data‐driven robust LAPD method that exploits information transfer in the worst‐case scenario (WCS) involving uncertain parameters between the two stages of RO. By employing the multilayer perceptron (MLP), the WCSs of uncertain parameters are tentatively preidentified as preliminary inputs of the column‐and‐constraint generation (C&CG) algorithm to reduce the stress of the WCS search during the subsequent iterative procedure. In addition, a combined multilayer perceptron is proposed to address correlations between uncertain parameters in the WCS. The MLP model is trained in distinct time segments to reduce the model training complexity and improve the precision of WCS preidentification. Case studies with various system scales and rolling window sizes indicate that the proposed method is effective at accelerating the C&CG process.