Stochastic Battery Operations using Deep Neural Networks

Yize Chen (1) , Md Umar Hashmi (2, 3) , Deepjyoti Deka (4) , Michael Chertkov (4)

1 ECE - Department of Electrical and Computer Engineering
2 DI-ENS - Département d'informatique - ENS Paris
3 DYOGENE - Dynamics of Geometric Networks
4 Theoretical Division and Center for Nonlinear Studies

In this paper, we introduce a scenario-based optimal control framework to account for the forecast uncertainty in battery arbitrage problems. Due to the uncertainty of prices and variations of forecast errors, it is challenging for battery operators to design profitable strategies in electricity markets. Without any explicit assumption or model for electricity price forecasts' uncertainties, we generate future price scenarios via a data-driven, learning-based approach. By aiding the predictive control with such scenarios representing possible realizations of future markets, our proposed real-time controller seeks the optimal charge/discharge levels to maximize profits. Simulation results on a case-study of California-based batteries and prices show that our proposed method can bring higher profits for different battery parameters.

Mots clés: Battery energy storage; Machine learning; Power system economics; Scenario forecasts; Generative model

Domaines: Sciences de l'ingénieur [physics]; Energie électrique; Informatique [cs]; Intelligence artificielle [cs.AI]; Apprentissage [cs.LG]

Yize Chen, Md Umar Hashmi, Deepjyoti Deka, Michael Chertkov. Stochastic Battery Operations using Deep Neural Networks. IEEE ISGT NA 2019 - Innovative Smart Grid Technologies Conference, Feb 2019, Washington, United States. ⟨hal-02070232⟩