需求响应收益导向的虚拟电厂多时段等效负荷预测方法OA

In the day-ahead invitation-based demand response,the equivalent load forecasting results of virtual power plants(VPPs)directly affect their bidding decisions and operation revenues.However,traditional forecasting methods aim at maximizing forecasting accuracy but neglect the asymmetric relationship between forecasting deviations and decision-induced revenue loss,which may lead to revenue reduction.To address this,this paper proposes a revenue-oriented multi-period equivalent load forecasting method for VPPs participating in demand response.Firstly,this paper reveals the inconsistency between equivalent load forecasting accuracy improvement and VPP revenue enhancement,and constructs a two-stage decision-making model for VPP demand response.Secondly,a framework of"accuracy-oriented training+decision-oriented fine-tuning"is designed,utilizing a surrogate network to effectively characterize the revenue loss caused by forecasting deviations of the decision-making model.Finally,to tackle the learning difficulty for surrogate networks under multi-period coupling decision-making,a local decision-oriented loss function learning method with multi-period decoupling representation is proposed to approximately characterize the multi-period local decision-oriented loss function by the aggregated learning of single-period local losses.Numerical case results demonstrate that,compared with accuracy-oriented methods,the obtained forecasting results can significantly enhance the demand response revenue of VPPs.