Model-Free Self-Supervised Learning for Dispatching Distributed Energy Resources

Abstract

This paper proposes a model-free self-supervised learning (SSL) method for dispatching distributed energy resources (DERs). The proposed method first establishes a data-driven DER dispatch formulation, which aims to build a decision rule that maps operating conditions (such as user demands and renewable generation) directly to cost-effective dispatch decisions. Unlike model-based methods which need to solve computationally intensive multi-period dispatch problems, this decision rule allows operators to make dispatch decisions immediately. We then translate this data-driven formulation into a training task by employing a neural network to learn the decision rule. This task is “self-supervised”, meaning it is achieved by minimizing a specialized training loss without requiring any labels. As this loss is originally model-based, we design a model-free replication using three calibrated surrogates. Then, a dispatch proxy can be trained to deliver high-quality decisions that exhibit marginal constraint violations. Moreover, this training can be accomplished without requiring exact model information (e.g., DER models, network topology, and line impedance), which is often unavailable in practice but essential for most existing model-based or learning-based methods. Simulations on the IEEE 33-bus and 123-bus test systems demonstrate that our method can quickly deliver safe and cost-effective dispatch decisions for DERs without requiring exact model information.