Deep Reinforcement Learning based Solution Approach for Unit Commitment under Demand and Wind Power Uncertainty

Abstract

In this work, we present a deep reinforcement learning-based solution strategy for the day ahead scheduling of power generation resources in energy systems with intermittent renewable energy resources. Fluctuations in wind power generation and the stochastic nature of electrical demand are considered in the unit commitment of generation resources. Typically, solution techniques for the stochastic multistage unit commitment problem are either computationally expensive or yield overly conservative solutions. In the proposed solution technique, temporal and spatial correlational structures of uncertainties present in the system are captured with a neural network function approximator. A causal policy is obtained which relies only on previously observed wind power and demand forecast errors. We conduct computational experiments on the IEEE 39-bus test case to demonstrate the effectiveness of the proposed solution strategy and improvement over existing unit commitment solution techniques.