Model-Free Aggregation for Virtual Power Plants Using Input Convex Neural Networks

Abstract

The virtual power plant (VPP) has been advocated as a promising way to aggregate massive distributed energy resources (DERs) in a distribution system (DS) for their participation in transmission-level operations. This requires identifying the feasible set of VPP power transfers fulfilling DS operational constraints. The identification task is completed using the existing methods by relying on constraint parameters (e.g., line impedance). However, the constraint parameters in a DS may be inaccurate and even missing in practice. Consequently, this paper develops a model-free aggregation method for VPPs. The proposed method first develops an input convex neural network (ICNN)-based surrogate for the feasible set of VPP power transfers. Our ICNN-based surrogate can be determined in a model-free manner by historical data. Furthermore, it is proven by leveraging the convexity and epigraph relaxation of an ICNN that our ICNN-based surrogate can be reformulated as a linear programming model without binary variables. This allows efficiently embedding our ICNN-based surrogate in transmission-level operations so that numerous VPPs can be efficiently coordinated at the transmission level. The proposed method is verified by numerical experiments in the IEEE 33-bus and IEEE 136-bus test systems.