Three Operational Problems of the Platform for managing the Distributed Energy Resources

Abstract

Distributed energy resource(DER) is an energy resource that generates and manages small-scale and geographically dispersed electricity (e.g. solar photovoltaic(PV), energy storage systems(ESSs), and demand response(DR)). It is expected that the market share of the DERs will increase gradually. Then, the influence of the DERs on the market grows and the DERs will be able to replace traditional generators. However, since the output of the DERs is relatively unstable, they have difficulty directly participating in the market. To help the participation of the DERs, a virtual power plant(VPP), a platform that collects small-scale DERs and acts as a single plant, emerges.

Furthermore, due to the volatility and uncertainty of the DERs' output, the VPP faces several operation problems. Among the several operation problems, we focus on three operation problems: (1) DER bidding decision under the DER forecast uncertainty, (2) multi-ESSs cooperative operation in real-time, and (3) renewable energy certificate (REC) management under the volatile market price. The objective of this dissertation is to provide practical algorithms for the VPP to maximize its own and its customer group's total profits under DER markets and a REC market.

At first, since the DER market requires the responsibility to provide the amounts of the DERs as much as the VPP bids, the decision on the bidding amount affects the VPP and the customers' profits. In detail, to prevent the DER from damaging the stability of the market, the market imposes a penalty for the shortage or provides an incentive for adequate supplies. In the first study, we formulate a bidding decision problem under an incentive-based market structure as a Markov decision process model. To quantify and describe the forecast uncertainty of the DERs, we propose frameworks to generate scenario trees or lattices. Especially, we devise the framework to generate the scenario lattices for the clustering-based forecast model. Furthermore, we incorporate heuristic techniques into the algorithm to reduce the computational burden. As a result, the computation time decreases by about 14% in the pilot test.

After submitting the bidding amounts to a day-ahead market, the VPP supplies the DERs as much as the requirement on the day of operation. By operating the ESSs, the VPP can fulfill the requirement amounts more easily. In the second study, we deal with the ESS real-time operation in a DR market with a non-fulfillment penalty. The VPP participates in the DR market by grouping the customers together. Moreover, since the VPP aims to maximize the group's total profits, the VPP controls the ESSs cooperatively. However, the computation time increases proportionally to the number of customers. Hence, we provide a multi-agent reinforcement learning-based algorithm. To improve the performance of the algorithm, we propose an action adjustment process and a problem-specific contribution structure. The algorithm can derive the ESS charging/discharging operation schedules in a short time. In addition, the algorithm makes a decision considering the forecast uncertainty of the electricity demand.

As well as the electricity, a REC, a by-product of renewable energy generation, can be sold. The certificate plays a role as a subsidy to promote renewable energy generation. However, there exist several challenges: a minimum capacity requirement for market participation and volatile market prices. Then, the VPP acts as a broker making REC sales decisions on behalf of individual renewable energy generators. In the last study, we formulate a VPP's REC management problem using a MDP. From the formulation, we analyze the structural properties of the optimal policy. Although we obtain the structural properties of the optimal policy, the VPP can not utilize the policy in practice. The reason is that the mathematical model-based approach requires the long-term prediction of the volatile price. Then, we develop a deep reinforcement learning (DRL)-based algorithm that incorporates graph neural networks (GNN) and the structural properties of the optimal policy. We numerically validate the proposed algorithm has improved convergence and solution quality compared to several baseline solution methods.

We numerically validate the performances of all proposed algorithms under real data-based simulation environments and check whether the computation times of the algorithms are appropriate to the actual operations. This dissertation deals with several operation problems for the VPP. We expect that this dissertation will serve as a key reference for VPP companies interested in operating the DERs and market designers of the newly introduced DER market.