A Machine Learning Approach to Boost the Vehicle-2-Grid Scheduling

TL;DR

The paper tackles leveraging fleets of electric vehicles as a distributed BESS to provide capacity-related grid services in wholesale markets, focusing on the charging/discharging scheduling problem under operational constraints. It proposes a two-stage framework that first generates scheduling policies via approximate dynamic programming using historical data, then trains kernel-based classifiers (SVC/QSVC) to predict aggregate EV actions from the current state. Empirical results show that with large fleets, the approach attains objective values comparable to CPLEX and approximate DP but with substantially reduced run times, enabling frequent reoptimization as system conditions change. This work demonstrates a scalable, data-driven path to harness EVs for grid services in DAM/ICM environments, with practical implications for real-time energy arbitrage and grid flexibility.

Abstract

Electric Vehicles (EVs) are emerging as battery energy storage systems (BESSs) of increasing importance for different power grid services. However, the unique characteristics of EVs makes them more difficult to operate than dedicated BESSs. In this work, we apply a data-driven learning approach to leverage EVs as a BESS to provide capacity-related services to the grid. The approach uses machine learning to predict how to charge and discharge EVs while satisfying their operational constraints. As a paradigm application, we use flexible energy commercialization in the wholesale markets, but the approach can be applied to a broader range of capacity-related grid services. We evaluate the proposed approach numerically and show that when the number of EVs is large, we can obtain comparable objective values to CPLEX and approximate dynamic programming, but with shorter run times. These reduced run times are important because they allow us to (re)optimize frequently to adapt to the time-varying system conditions.

Figures (5)

• Figure 1: Schematic illustration of the problem of leveraging EVs a BESS in the DAM. The blue dashed line indicates the energy demand and the red dashed line is the total energy bought from the markets (DAM + ICM). When the total energy bought is larger than the demand (blue areas), it is possible to store the energy (given that there is enough storage capacity). The red areas indicate an energy shortage that must be covered with storage energy. If there is insufficient energy in storage, the ES must buy more energy from the ICM.
• Figure 2: CPLEX optimality gap in euros as a function of the run time for 200 EVs and different levels of complexity. The optimality MIP gap is difference between the best integer solution and best (continuous) lower bound. The results are the average of 50 days from EPEXSPOT epexspot selected uniformly at random (2015-2019). The shaded area indicates the standard deviation. The solver is run for 10 minutes and the optimality gap is collected every second.
• Figure 3: Schematic illustration of the proposed approach.
• Figure 4: Value of the objective function of the best solution obtained with different methods (average over 100 scenarios), when the number of EVs varies. Shaded areas indicate 0.2 standard deviation. Lines are drawn for the proposed SVC/QSVC-based algorithm, for the approximate DP used in data generation, and for CPLEX with a runtime limited to 10 minutes.
• Figure 5: Run times of the methods in Fig. \ref{['fig:obj_fn']}, excluding CPLEX which is time-bound to 10 minutes. Run time is measured as wall-clock time, and is restricted to inference since training is performed offline. Shaded areas indicate the standard deviation, which is inappreciable for SVC and QSVC.